Kawésqar SECURE LINUX · AGENT FABRIC · PATAGONIALINUX SÉCURISÉ · AGENT FABRIC · PATAGONIESICHERES LINUX · AGENT FABRIC · PATAGONIENSIKKER LINUX · AGENT FABRIC · PATAGONIALINUX SEGURO · AGENT FABRIC · PATAGONIALINUX SEGURO · AGENT FABRIC · PATAGÔNIA
EN
Agent Fabric · lab notebook

More than 70 labs. The ones that failed too.

This is the lab notebook of the Agent Fabric: every experiment we can publish, with the plane it ran on, the verdict, the figure and the limit. Failures and refutations sit next to the passes, because a lab that only shows what worked is showing you a pitch, not evidence.

0+
experiments in the lab, and counting
0
shown here, each with its verdict, plane and limit
0
tested: PASS, HELD or mitigated by policy
0
published as FAIL or REFUTED, not buried

patent pending  Patent applications filed with INAPI (Chile); in process.

Verdict board

Every lab, by plane.

One chip per lab. Green is a tested pass, amber is partial, unclear, open or a measured residual, orange is a recorded failure or refutation. The plane matters as much as the colour: our results run nested (inside a VM, which validates neither isolation nor performance) or on bare metal, the only place those claims can close. Metal means one laptop workstation, n=1, unless a card says otherwise.

Nested VM · 16 labs

12 tested · 2 partial or open · 2 failed

Nested cloud · 1 lab

0 tested · 1 partial or open · 0 failed

Lab72 also carries a TLA+ model checked against traces, counted here under bare metal. In the cloud we show only what orients; it never closes a verdict.

What failed

What failed, said plainly.

A lab that only publishes its passes is selling. These are the places where the data said no, or said less than we hoped, and what we changed after.

nested VM  FAIL

lab14 · The rump guest printed nothing

FAIL on nested Xen. The guest reached long mode and ran rump_init, yet two console faults kept the marker from ever appearing. The failure dictated the order of fixes that lab16 closed.

nested VM  FAIL

lab15 · The observer did not fail closed

F9 FAIL: 38,577 mutations produced 4,520 crashes, 85 false countersignatures (only with a compromised stream) and 56 hangs. lab29 later reproduced the failure on metal, hardened the observer and reached 0 crashes in 3,000 variants.

bare metal  FAIL

lab25 · The disposable cell leaked on disk

RAM held (0 hits) but a reused volume let cell B read cell A's pattern: 6,882,854 hits. Zeroing the whole volume on create brought it to 0, a fix verified in the lab before it reached the product.

bare metal  REFUTED

lab64 · Gate plus order cost more than 1 ms

We pre-registered under 1 ms. The measured median is 2.735 ms (P=1, n=200). The hypothesis is refuted and the number is the one we publish.

bare metal  REFUTED

lab70 · The faster Gate still missed its target

H1 asked for under 1.5 ms of wall time. The full arm adds +3.210 ms (95% CI 3.074 to 3.378), and +3.216 ms again with the performance governor. Refuted twice.

bare metal  FAIL

KZ-F1 phase 0 · C1

FAIL against the pre-registered criterion: the register path returns an I/O error under Xen 4.20, before and after load. An alternative interface answered and followed the load, but it is not validated (n=1, no external meter) and two registered checks of it missed their criteria. It affects only energy measurement, and no energy figure is cited.

bare metal  REFUTED

KZ-F1 phase 2 · H7b

We pre-registered at most 1,800 lines of code for the production path. The count is 7,840, and above 2,000 counts as refuted. The hypothesis is refuted and the number is published as measured, without redefining the count.

bare metal  PARTIAL

lab20a · An observation we must not oversell

One DMA denial was seen once and could not be reproduced; under strict mode the positive control was itself denied. It is an observation, not a result, and we do not cite it as “strict denies unmapped transfers”.

logic  PARTIAL

C5 · The close was not won

A broker that forges the storage confirmation crossed the gate in the first logic wave (23/23 behaviors reproduced). Later work added a separate store; the card for kaf-store shows what that closes and what it does not.

bare metal  TIE

lab27 · “Rump exposes less” was false

The hypothesis was falsified: rump keeps 2 live event channels against 0 for minimal Linux, and they tie on grants and xenstore nodes. We concluded rump did not justify deep investment on that ground alone.

bare metal + nested cloud  PARTIAL

kz-az1 · The 2x claim only half holds

On 100 ms jobs the 2x edge holds against untuned Linux (2.01x, borderline) and is falsified against tuned Linux (0.92x to 1.03x). Before the trim, Unikraft was ready sooner. We lead with the part that did not hold.

Timeline

Six stages, one thread.

The notebook follows the order we worked in: sandbox first, then Xen nested in a VM, then the physical machine, then attacks on our own design, then the fast cells, and last the cloud. Bar height is the number of labs shown for each stage; the reference code (11 more cards) sits apart at the end.

timeline
2SANDBOX14NESTED XEN27BARE METAL16RED TEAM15KZ CELLS2CLOUDSANDBOX2NESTED XEN14BARE METAL27RED TEAM16KZ CELLS15CLOUD2
timeline   Labs shown per stage. A nested or cloud result never counts as an isolation or performance verdict: only bare metal can close those.
Stage 1 · 16 labs

Nested Xen, the first answers.

Inside a VM (Xen under QEMU/KVM). These runs show the architecture working as logic on real domains. They are never an isolation or performance verdict, and each card says so.

Each card: lab, plane and verdict, then what was found, then the limit that comes with it. Figures are the ones the lab recorded; where a later campaign superseded them, the card says so.

lab03 · nested VM  PARTIAL

NetBSD rump kernels: closing the unknowns

The real IP stack uses 52 hypercalls (base 46) and 28 of 46 are invoked at startup; in the official order the real definition wins in 3 of 3 duplicates; the sysctl warning is benign.

limit  One nested guest; partially reproducible build (186/190 identical); no conclusion on performance or isolation.

lab04 · nested VM  PASS

Build non-determinism, guest disk, skill tests

The difference was only embedded source paths: two MKREPRO builds come out identical, 190/190. It corrects the lab03 hypothesis; skill tests 12/12 after rerunning one.

limit  n=2 builds per condition; the exact cause is strong inference, not fact; one case failed on the first pass.

lab05 · nested VM  PASS

Nested Xen and a minimal PVH cell

The cell prints its marker and exits on its own; 5 cycles and 1 destroy during boot leave 0 domains and 0 residual xenstore keys.

limit  All nested and without IOMMU: not valid for performance or isolation; not a benchmark.

lab06 · nested VM  PARTIAL

Skill validation and nested Xen cell (A to D)

A: 10/12 PASS; B: 7 PASS and 1 FAIL (one case invented a format); C: 5/5, 30 cycles with 0 residue and 0 failures; D: 5/5. The rump hypercall count is 46 on both sides.

limit  Nested, no IOMMU; three cases cannot be reproduced because their prompt and criterion are not on disk.

lab07 · nested VM  PASS

S4.0 xenstore probe between two cells

Without an ACL the cross read returns EACCES (30/30 in the control); the owning cell grants access and the other reads the nonce and gets 3 events. dom0 is not required.

limit  Nested, N=1 per arm, a single xenstored variant; no quotas or concurrency.

lab07b · nested VM  PASS

S4.2 gntalloc+gntdev+evtchn ring between two cells

Bidirectional ring with identical hashes at both ends, no dom0 privilege. The first run FAILED because of our own faulty protocol and was fixed.

limit  Nested, n=1 PASS; N≥30 pending; it was not shown that a third domain cannot map the grant.

lab08 · nested VM  PASS

Placeholder domains for control, broker and observer

30 cycles (90 domains): 0 XenStore residue and 90/90 exit code 0; a test Cell was denied in 15/15 operations on the three role branches.

limit  Nested, one run, placeholder domains with no real logic; not an isolation verdict.

lab09 · nested VM  PASS

Role logic inside an initramfs

The three suites pass inside a 256 MB PVH domain (17/17, 31/31, 9/9) with ~167-172 MB free and 0 out-of-memory signals.

limit  Nested, n=1; functional viability and memory headroom only.

lab10 · nested VM  PASS

Broker in one domain and a Cell over the ring

5/5: a 9000 B artifact with identical hash both ways; foreign or forged CID, replay and forced destination denied; 0 processes of the pair in dom0.

limit  Nested, n=1; no conclusion on performance or resistance on Xen bare metal.

lab11 · nested VM  PASS

Observer that countersigns and issues a destruction receipt

4/4: countersignature only while the Cell was alive during a real STORE; destruction receipt only after destroying it; 17/17 attempts to read, alter or delete the stream denied.

limit  Nested, n=1; does not cover a hostile Cell with kernel capabilities.

lab12 · nested VM  PASS

Chain A, receipt, B across real domains

All 4 conditions hold: B does not exist before the order, without a countersignature the Gate denies, a Cell trying to sign the order is denied (18 times) and the hash is identical end to end.

limit  PASS bounded to architecture: nested, n=1, first-generation symmetric signature; the first attempt failed on packaging.

lab13 · nested VM  PASS

B1.8 Falsifications only the substrate can refute

All 3 falsifications PASS: CIDs from the previous epoch give STALE_EPOCH after recreating the broker; a third domain cannot map the grant (128 refs denied); a kill between STAGED and COMMITTED gives NOT_COMMITTED. Finding: epoch and nonces must be durable outside the domain.

limit  Nested, n=1, HMAC v0 signature; the kill instant is forced; dom0 remains in the trust base.

lab14 · nested VM  FAIL

T15b First boot of rump as a PVH guest under Xen

Instructive FAIL: the guest entered via PVH, reached long mode and ran rump_init, but did not print the marker. Two console faults (unprogrammed hypercalls and a buffer removed by the compiler) dictated the fixes that lab16 closed.

limit  Nested, n=1; it stayed unobserved whether rump_init returned 0.

lab15 · nested VM  FAIL

S9.2 Observer falsifications 8 and 9

F8 PASS: the observer survives a cell crash and the Gate does not use its self-report. F9 FAIL: with 38,577 mutations there were 4,520 crashes, 85 false countersignatures (only with a compromised stream) and 56 hangs. A crash signs nothing: it is an availability failure.

limit  Nested, n=1, HMAC v0; the lie is a fixed script; dom0 remains in the trust base.

lab16 · nested VM  PASS

T15c Closing the rump axis: the rump guest prints KZ-RUMP-OK under Xen

With the two console fixes, the rump guest prints the marker, rump_init returns 0 and it shuts down by hypercall with code 0. It closes the bit lab14 could not read.

limit  Nested, n=1: init and console only; no performance or hard isolation.

lab17 · nested VM  PASS

T16 Adversarial guest: A_effective contained in A_delegated

289 attempts to map foreign grants and 289 to bind event channels gave 0 successes, 0 foreign pages read, 0 self-grants. What Xen contains is separated from what Kawésqar contains (replay, STALE_EPOCH, NOT_COMMITTED).

limit  Nested, n=1; the guarantee assumes a correct Xen; grant and channel denials are Xen's, not Kawésqar's.

Stage 2 · 12 labs

The first campaign on metal.

On the physical control domain of one laptop workstation, n=1. Here the first honest failures appear: a covert channel we could not measure and a disk that leaked.

lab18 · bare metal  PASS

Physical baseline and gate for the lab plane

Inventory captured and preflight 11/11. The CPUs in this Xeon are not affected by L1TF, MDS or TAA, so the side-channel lab was re-scoped.

limit  ZBook, n=1, a single day; this is inventory and preparation, not an isolation or performance result.

lab19 · bare metal  PASS

Re-attesting the architecture core on metal

7/7 core verdicts PASS on the physical control domain. Minimal cell cycle: p50 0.95 s, p99 1.09 s, 200/200 without failure, real residue 0.

limit  ZBook, n=1, control domain capped at 4 vCPU: correctness, not performance; the chain on real domains is still HMAC.

lab20a · bare metal  PARTIAL

VT-d enforcement on bare metal with a named DMA device

VT-d configuration was attested and a denial was seen once, but it is not reproducible. Under strict mode the instrument ended UNCLEAR: the positive control was denied.

limit  ZBook, n=1. It must not be cited as “strict denies unmapped transfers”: it is an observation, not a tested result.

lab21 · bare metal  PASS

Baseline latency of a cell on bare metal

200/200 per variant, 0 residue: create with xl 184 ms (p50) and kernel plus init boot 607 ms (p50, p99 790 ms). A cell with work adds a bounded overhead.

limit  First-generation figures (Linux kernel and xl), long before today's cells: do not compare them.

lab22 · bare metal  PASS

Scale and load

0 Gate failures in 256 runs under load; 2,775 cycles completed at 100%; creation steady around 237 ms from N=1 to N=414; 0 residual cells.

limit  This is the Gate logic under scheduling pressure, not real cell transitions at scale; the resource ceiling was never stressed.

lab23 · bare metal  UNCLEAR

Contention covert channel

The positive control did not cross the threshold: best case 9.64 bps with 95% CI [-33.06 ; 52.31], indistinguishable from noise. Neither presence nor absence of a channel is claimed.

limit  Instrument not validated; recorded as an unmeasured channel, never as negligible.

lab23b · bare metal  MITIGATED

Covert channel: closed by policy

Risk closed by a placement policy (do not co-schedule mutually distrusting cells on sibling threads), not by measurement: the strongest SMT channel does not exist by construction.

limit  It is not “we showed there is no channel”: shared-cache and cross-core memory channels remain as residual risk.

lab24 · bare metal  OPEN

XSAs that really apply

Three advisories patched, the speculative ones inapplicable on this silicon or mitigated; one applicable advisory remains unpatched (a liveness DoS in the xenstore daemon). None is an escalation or cross-domain leak.

limit  State tied to a package version; it must be rerun when the build changes.

lab25 · bare metal  FAIL

Disposable cell: RAM and disk residue

RAM PASS: 0 hits with forced overlap of ≥548 MiB. Disk FAIL: cell B sees A's pattern (6,882,854 hits); zeroing the whole volume on create brings it to 0 hits.

limit  The mitigation was verified in the lab but was not yet in the product; device DMA residue out of scope.

lab26 · bare metal  PARTIAL

Closing the bare metal campaign

Architecture: bounded PASS, conditional close. No amplification of authority or evidence on real Xen (7/7) and under load; disposable cell falsified on disk and PASS on RAM.

limit  No verified physical isolation is claimed, nor unconditional disposable cell, nor the full thesis; covert channel UNCLEAR.

lab27 · bare metal  TIE

Boundary shootout: KZ-Rump against minimal Linux

The hypothesis “rump exposes less” was falsified: rump keeps 2 live event channels (minimal Linux 0); they tie on grants (0) and on 38 xenstore nodes. Finding: an adapter abort gave reason code 0, same as a clean shutdown.

limit  Observed surface, not reachable; both runtimes are near-stubs; one machine, one run.

lab28 · logic + bare metal  PASS

Delegation tree n>1 and non-amplification by induction

Real Gate: 1,958 edges decided, 0 amplifications; real cells at the exact envelope edge (768 of 768 MiB); a grandchild asking for more than its parent is denied with AUTHORITY_NOT_SUBSET; 5 mutants detected.

limit  The parentage tree is modeled (Xen does not know it); depth above 2, fan-in and fuzz are logic. The lab found real Gate defects, since fixed.

Stage 3 · 15 labs

Pushing the limits.

Still on metal: observer fuzzing, the signed chain inside real cells, saturation to 412 live cells, strict IOMMU and observed hypercalls.

lab29 · bare metal  PASS

Falsification 9: observer surface

The first run FALSIFIED the property: 47 crashes and 208 violations through holes reachable via JSON. After hardening the observer: 0 crashes and 0 violations across 3,000 variants, 3 mutants detected, 0 residue.

limit  PASS bounded to the wire (decoded JSON stream); in-process violations remain, declared as a GAP. n=1.

lab30 · bare metal  PASS

Falsification 10: transactional STORE under kill

At no stage was a partial COMMITTED deliverable left (7 real kill -9); replay after restart always gave STALE_EPOCH; 3 mutants detected; real cells destroyed.

limit  Journal persistence is modeled by the harness: it tests the contract's atomicity, not the production persistent CAS (declared GAP).

lab31 · bare metal  PASS

Adversarial Gate under load

0 acceptances out of 6,870 concurrent adversarial attempts; 1,194 of 1,194 legitimate ones cross under the same load; the naive race over-reserves 6 of 6 (the instrument sees the race); 0 residue.

limit  The ledger lock is in-process; production persistent CAS is out. A finite corpus, not a general proof of absence.

lab32 · bare metal  PASS

Stability soak of the create-run-destroy cycle

400 of 400 iterations: zero drift (Kendall tau 0.02 to 0.08), residue 0, Xen free memory 55,883 to 55,883 MiB, dom0 stable.

limit  Bounded window (not months), one minimal cell, one machine, n=1.

lab33 · bare metal  PASS

Saturation at maximum scale and throughput

Peak of 412 simultaneously live cells (52,736 of 53,835 MiB), 0 amplification and 0 residue across 14 levels; at most 14.78 cells/s creation and 8.98 cells/s full cycle.

limit  n=1. At the peak, destruction reaches p99 32 s because the toolstack serializes; at scale only a sample carries a console marker.

lab34 · logic + bare metal  PASS

Resource fan-out (Falsification #4)

In logic: 4,681 nodes and 7,608 edges with 0 amplifications; on metal, with 128 MiB cells, every step completes, 0 amplifications and per-cell overhead of 1.5 to 2.5 MiB.

limit  The first two physical attempts were UNCLEAR (cause found in lab40); the close is the third. n=1.

lab35 · bare metal  PASS

Minimal observed boundary of the cell under load

Full classification of 18 interfaces and 49 hypercalls; load did not break the boundary, 0 residue in all 4 scenarios. Xen, not the runtime, bounds isolation.

limit  Characterization, not an isolation verdict. Trace observation stayed UNKNOWN on all 49 rows for lack of a decoder (lab44 closed it).

lab38 · bare metal  PASS

End-to-end Ed25519 cryptographic chain inside cells

The chain of cell A, countersignature, Gate, cell B and destruction receipt runs on metal; 4 mutants (dom0 forgery, altered receipt, B without evidence, replay) and 7 negative vectors rejected without creating B.

limit  Verifying the signature before storing is modeled by the harness, not by the reference Gate. n=1.

lab39 · bare metal  PASS

Smoke and invariant regression under strict IOMMU

All 4 groups pass: 60/60 iterations, 0 accepted out of 3,444 adversarial attempts with 600/600 legitimate admitted, and zeroing verified; create rises about 12% (some 23 ms) under strict.

limit  Bounded regression, not exhaustive re-testing; the clean A/B in the same boot is pending. n=1.

lab40 · bare metal  PASS

Boot memory floor of a PVH cell

Measured floor of 96 MiB: with 64 MiB the cell does not boot, with 96 and 128 MiB it does. Closes the root cause of the lab34 failure and falsifies the harness race hypothesis.

limit  Coarse floor (n=3 per point), valid for this lab kernel and initrd; the exact value between 65 and 96 MiB was not refined.

lab41 · bare metal  PASS

Audit of dom0 grants over cell pages

A diskless cell, idle or under load, has exactly 1 active grant mapped by dom0 (the console ring), 0 data pages; with a disk 67 refs appear (positive control).

limit  The first run was UNCLEAR because of a blind positive control and was fixed. That they enter dom0's IOMMU is inference from the source. n=1.

lab42 · bare metal  PASS

Entropy boundary inside the cell

14 of 14 distinct keys; in no boot was a key generated while the random generator was not ready, even with RDRAND off (median 1.47 s versus 0.14 s).

limit  It shows availability and non-determinism, not entropy quality. The no-RDRAND case rests on reading the source. n=1.

lab43 · bare metal  PASS

Cryptographic chain in a cell with n greater than 1

The chain and the fan-out cross end to end and every mutant is rejected with its exact reason; the fan-out fills the exact envelope (2 by 128 MiB). Finding: the pure gate does not consume the prev link of destruction receipts.

limit  Signature verification and the prev link are modeled in the harness (production gap declared). n=1.

lab44 · bare metal  PASS

Hypercalls observed through a real trace

In all 4 scenarios what is observed is a subset of the declared set of 10 hypercalls; the rump cell leaves a minimal footprint (5 distinct numbers) and Linux adds 4 more. Negative control: 0 calls with the cell paused.

limit  The hypercall number is observed, not the sub-op or the arguments. n=1.

lab45 · bare metal  PASS

Soak under strict IOMMU and structure leaks

400 of 400 cycles without leaks: active grants flat at 2, maptrack flat at 4 and Xen free memory constant (55,883 MiB before and after).

limit  Free memory is read in MiB (detectable minimum near 80 KiB per cycle); active grants are a partial proxy. n=1.

Stage 4 · 16 labs

Attack, then verify outside dom0.

The decision moves out of the control domain, then we attack our own design, measure what dom0 can still read and compare against Linux.

lab46 · logic  PASS

Denial conformance and live agent against the real channel

12 of 12 checks and the independent oracle PASS, each denial with its exact reason; 0 escalations crossed with the 8 language models tested, including the 3 that did attack through the real channel.

limit  Logic level, not metal nor an agent inside the cell. A later campaign corrected the lab's own overclaims (mutants that mutated data, not code).

lab47 · bare metal  PASS

Verifier outside dom0 (chain A to B)

The positive control crosses and all 11 mutants are denied with their exact reason; creations = admissions = verifier PASS (1 = 1 = 1); dom0 forgeries and replays die by signature. Re-attested on metal: 3 of 3 runs PASS.

limit  n=1, one host. dom0 still creates domains and can read the RAM of a live domain (hardware residual). The first physical run was UNCLEAR because of an oracle bug, fixed.

lab48 · bare metal  PASS

Hardened gate (gate_verified) against the harness

With the harness switched off, the old gate ACCEPTS and creates the child in the 4 main mutants; the new one DENIES them with its reason. The properties move from modeled in the harness to enforced by the gate.

limit  n=1. dom0 still generated the private keys in this run (lab47 attacks that); a lying tap is not covered.

lab49-c5 · logic + bare metal  PASS

Closing C5: compromised broker

The old observer accepts both forgeries; kaf-store denies both and the byte witness denies the first. On metal, a compromised broker that signed without backing the bytes leaves the medium empty and does not cross.

limit  The witness lets through a false COMMITTED over bytes honestly in staging. This does not settle the later red-team marker, which records the C5 close as not won. n=1.

lab50-dma · bare metal  PASS

DMA confinement of an assigned device

dom0 reads the USB stick before assignment, loses the device on assignment, and the domU reads the same number of sectors with 0 VT-d faults in strict mode. It measures confinement, not adversarial isolation.

limit  Partial: a malicious driver attempting DMA into another domain's memory was not tested. It ran in HVM because PVH does not support passthrough. Do not cite as closed DMA isolation.

lab51-netless · bare metal  PASS

lab51-netless · cell with a kernel without a network stack

A PVH cell with a kernel built without a network stack boots and runs: /proc/net, /sys/class/net and iplink ABSENT and the load closes with sum=500500, with the console ready in under 1 s of kernel boot.

limit  A fact about how a kernel was built, not an isolation or performance verdict; the evidence is the console and the configuration. n=1.

lab52-smt · bare metal  PASS

lab52-smt · placement against the SMT channel

A neighbor on the sibling thread slows the victim 4.98 to 5.94 times; on another core it drops to 1.07 to 1.13 times, near baseline. Placement suppresses the SMT channel (3 runs).

limit  It measures throughput interference, not a bit rate or key extraction. The ~1.08x residual is the cross-core channel (L3). n=1.

lab53-llc · bare metal  PARTIAL

lab53-llc · cross-core channel through the L3

The cross-core channel exists and is small: 1.03 to 1.14 times interference (average ~1.08x), an order of magnitude below the SMT channel. A residual measured and bounded, not closed.

limit  Without Intel CAT/RDT the L3 cannot be partitioned on this CPU. Small, noisy signal (3 runs); it measures throughput. n=1.

lab54-meminspect · bare metal  CONFIRMED

lab54-meminspect · the control domain reads a live cell

Positive control: dom0 mapped 10,076 frames of the live cell and found the secret without the cell's cooperation (2 runs). The signature authenticates, but gives no secrecy against dom0.

limit  It measures the exposure, it does not close it. Only hardware-encrypted memory (SEV-SNP/TDX) closes it; this CPU lacks it. n=1.

lab54b · bare metal  CONFIRMED

lab54b-meminspect-incell · secret born inside the cell

3 of 3 runs with 3 distinct secrets: 1 candidate per run, hash verified, 977 contiguous frames. The secrecy residual against dom0 is now measured.

limit  It does not close the residual: that needs hardware-encrypted memory (TDX/SEV-SNP), which this machine lacks, and that closing test has not been run. The same residual exists in every cloud, AWS and GCP included, unless confidential computing is on: whoever controls the host, the provider or an attacker who gets root on it, can read the memory of what runs on it. It is a property of the platform, not of Kawésqar. We did try confidential VMs in both clouds (TDX on GCP, SEV-SNP on AWS): memory is encrypted, but the cell cannot run inside (no KVM) and a confidential VM per job takes 50 to 58 s to come up. n=1.

lab55 · bare metal  PASS

lab55-composition · the five properties together

5 of 5 runs PASS: the hardened gate decides outside dom0, with no private key in dom0; forged or repeated orders die by signature and mutants are denied. Closes the “Composition” item.

limit  n=1, 5 repetitions (reproducibility, not statistics). dom0 still creates domains and can inspect memory (lab54 residual).

lab56 · bare metal  PARTIAL

lab56-substrate · full NetBSD or trimmed Linux

Instantiation: NetBSD PVH ~1.65 s, Linux PVH 2.06 s, NetBSD PV 3.19 s. Both substrates sustain 8 concurrent cells with near-flat scaling from N=4 to N=8.

limit  Early-stage figures, earlier than today's cells; the lab's text disagrees with its own table, so only the table is cited. n=1.

lab57-gate · bare metal  PASS

Real rump cell under the Gate

The honest run PASSES and 5 mutations give their named DENY before running; a chain of 3 real cells, 8/8 cells gated in parallel, 20/20 in stability. An “echo” cell without rump passed with a bytes-only witness and is denied once the kernel is attested.

limit  Bounded PASS, NOT containment shown: the console is a degraded medium and dom0 can read the cell's RAM (hardware residual). Small n.

lab58-overhead · bare metal  PASS

Xen and rump cell overhead against Linux

Rump cell cycle without work ~94 ms versus ~500 ms for the Linux cell (5.3x); boot 0.9 ms versus 237 ms; Xen charges +8% compute in dom0 and +11% in a Linux PVH cell versus bare metal; the rump cell sits at +4% of native.

limit  n=50 per cell, one machine. An anomaly in the rump sha phase (+15 to +20%) stays UNCLEAR; no I/O measured. Create figures predate the direct launcher.

lab59 · logic + bare metal  PARTIAL

Cell-escape red team (campaigns A to F)

On metal a compromised cell read 0 pages of a neighbor, opened 0 unauthorized channels and gained no authority; the launcher's prefix cleanup did kill cells of another job (closed in logic); native fuzz 85M executions without crashes; 15 of 16 ledger items closed in logic.

limit  What is closed is the logic plane and has not been re-measured on metal. Still open: the verifier that does not re-hash the bytes, the cross-core cache channel and dom0 reading cell RAM.

redteam-escape · logic  PARTIAL

redteam-escape · Wave 1 of logic attacks (C5, E5, E6, E8, CGAPS)

C5: the close was not won (a broker that forges the storage confirmation crosses; 23/23 behaviors reproduced); E6: prefix cleanup kills cells of another job; E8: the memory check admits 11 extra cells; E5: DENY. The failures were repaired later in lab59.

limit  Logic plane in a single process, no metal; this lab records the state before the fix.

Stage 5 · 15 labs

Cells, a signed chain and the clock.

Direct launchers, the cost of the Gate, sealed images, the minimal builder in three phases and the energy per cell. Three hypotheses were refuted and one check failed here, and all are kept.

lab61 · logic  PASS

Secure Boot negatives and adversarial map

5 of 5 cases as expected: the valid control verifies, and unsigned, foreign-key and mismatched-certificate artifacts are rejected; the artifact without SBAT is detected as such.

limit  Offline signature verification: loader rejection at real boot is described as expected, not measured; 17 adversarial dimensions remain pending.

lab62 · bare metal  HELD

kz-k2 on KVM against Firecracker

With the cell before the trim, ready from zero in 4.8 ms (kz-k2 on KVM) against 8.2 ms for Firecracker with the same cell; against Firecracker with a Linux guest the pre-registered hypothesis holds at 2.14 to 2.34x; at P=8 on the minimal cell, 251 versus 147 cells/s.

limit  n=1. With the same cell the edge is 1.04 to 1.10x on 100 ms jobs: it lives in the cell, not the VMM. On 1000 ms jobs it is lost (~1x). Against Firecracker with an untuned Linux, 3.5x measured on the same machine.

lab63 · logic  PASS

lab63 · Gate to launcher chain with Ed25519

107 negative tests pass (altered receipt, missing or foreign countersignature, broken lineage, badly signed order, replay, replay after restart) and the closure of the 5 images gives 84 of 84 checks.

limit  Tested in logic and static build, without Xen or a VM; the end-to-end Xen evidence with Ed25519 is in later labs.

lab64 · bare metal  REFUTED

lab64 · Gate and order cost over the cell cycle

The pre-registered “under 1 ms” hypothesis is refuted: Gate plus order add a median of 2.735 ms (P=1, n=200); the A/B delta in total time is +2.574 ms, and +10.815 ms at P=8.

limit  n=1, ondemand governor. At P=8 the Gate queues behind a single lock, so this is not “the Gate is slower under load”. A recorded failure, published as a trait.

lab65 · bare metal  PASS

lab65 · verifier outside dom0 over real rump cells

On metal 16 of 16 scenarios as predicted; the extension with artifact substitution gives 19 of 19 in three runs, 0 orphans; in logic 62 fidelity tests and 10 mutation tests pass.

limit  n=1. A dom0 that lies coherently gets PASS, declared a deliberate NO-CLAIM. Gate state lives in the verifier's RAM. dom0 relay channel, no mTLS. A bounded verdict, not “containment shown”.

lab66 · bare metal  PARTIAL

lab66 · kz-c2 with a sealed kernel (closing the lab64 TOCTOU)

Verifying the hash on every cycle costs +7.16 ms on a cycle of about 15 ms and is discarded; the second variant costs -0.27 ms in total_ms (15.345 to 15.079), within noise, with 9 negative tests OK.

limit  The TOCTOU attack was shown only with a stand-in for the loader, not real Xen. n=1, no SHA-NI. The full chain with Gate was measured in lab70 (+3.21 ms).

lab68 · bare metal  PASS

lab68-e2e-real · real evidence from A to cell B

10 of 10 chains of A, Gate, order, direct launcher and B on metal (2 batches of 5), oracle PASS in each; in logic 128 launcher checks and 54 preflight checks without failures.

limit  n=1. dom0 remains the trust base for the tap, and B's FETCH is a signed echo, not a real FETCH to the broker. No mTLS on that channel. The first smoke on metal was UNCLEAR and was archived.

lab70 · bare metal  REFUTED

lab70-gate-fast-fd · cost of the fast Gate over launching without a Gate

Hypothesis H1 (under 1.5 ms of wall time) is refuted: the full arm adds +3.210 ms median (95% CI 3.074 to 3.378) over the cell without Gate; repeated with the performance governor: +3.216 ms.

limit  n=1, Xen, P=1. The descriptor arm includes a durable nonce fsync that the others do not pay.

lab71 · bare metal  PASS

lab71-gate-evidence · single activation and release order

1000 of 1000 rounds with exactly one activation (the baseline without the predicate crosses 300 of 300); unsigned, foreign or repeated orders give 0 creations; 10 of 10 cycles with the release confirmed before the verdict.

limit  n=1, N up to 32, no power cut mid-write. In one run an arm gave an environment FAIL and a probe was fixed after three instrument defects: reported as they are.

lab72 · bare metal  PASS

lab72-b-evidence · evidence for Application B

After destroy, residue enumeration gave 0 in 100 of 100 cells, with destruction verified in at most 0.093 ms; persistent ledger with kill -9 and restart without violations; creating the Xen domain through the direct launcher 41 ms versus 115 ms with xl. It includes a TLA+ model with trace conformance.

limit  n=1. fs-verity and dm-verity were BLOCKED on Xen and were only tested with a KVM cell on another machine. The 41 ms is Xen domain creation, not cell latency.

kz-f1 · f0 · bare metal  PARTIAL

KZ-F1 phase 0: energy, CPUID policy and zeroed pages

C1, energy reading from dom0: FAIL against the pre-registered criterion; the MSR register path returns an I/O error under Xen 4.20. An alternative interface answered, but it is not validated (n=1, no external meter; two registered checks of it also missed their criteria) and it affects only energy measurement. C2, cell without CPUID policy: REVIEW (80 of 80 pairs with identical behavior, the dump differs). C3, zeroed pages after reusing memory: PASS (210 domains of 256 MiB, at least 51,636 MiB reused, 0 non-zero bytes).

limit  ZBook, n=1, one run per check. C1 is not a citable energy figure.

kz-f1 · f1 · bare metal  PASS

KZ-F1 phase 1: signed cycle, build time and xenstore

Full cycle with a signed order: smoke PASS (8 of 8 valid orders; 4 of 4 invalid orders rejected without creating a domain), repeated with the final front: PASS. H2, build time against kz-c2: INSIDE (paired difference of -2.1 to -3.9 ms; criterion ≤ +1.0 ms), repeated with the final front: INSIDE. H3, xenstore nodes and writes per cell: PASS twice (0 nodes and 0 writes across 200 cells). H7a: PASS. Sweep smoke (the worker dies with the cell alive): PASS, 8 of 8.

limit  ZBook, n=1: the smoke, H2 and H3 ran twice on the same machine; H7a and the sweep once. Build time is not end to end: end to end this phase was slower than kz-c2 (+7.9 to +16.2 ms).

kz-f1 · f2 · bare metal  PARTIAL

KZ-F1 phase 2: resident builder, recovery and scale

Cycle of 15.1 to 18.1 ms against 22.7 to 26.9 ms for the reference arm, Gate included. H1 INSIDE: paired improvement of -8.672 to -7.243 ms with 200 pairs per variant (criterion ≤ -4.0 ms). H6 PASS: 13 states × 100 and 200 clean endings, 0 violations. Smoke 8 of 8 with 3 negatives rejected; sweep 21 of 21; real failures 20 of 20 across the four cells. Not met: H1b INCONCLUSIVE (1.022 to 1.059 ms against a threshold of ≤ 1.0); H4 BETWEEN THRESHOLDS (entry 3.443 to 3.647 ms and 224 to 280 cells/s; criterion ≤ 3 ms and ≥ 243/s); H5 BETWEEN THRESHOLDS (1,800 cells per variant, no wall, lowest decile 74.1/s against a criterion of ≥ 78/s); H7b REFUTED (7,840 lines of code against a criterion of ≤ 1,800).

limit  ZBook, n=1, one run per test, unsealed. The cycle definitions differ from the KVM ones and from the rival's (Xen against KVM, with and without the signing Gate) and are not chained. “Between thresholds” and “inconclusive” are not passes.

energy · bare metal  PARTIAL

Energy per cell with RAPL counters, native kernel without hypervisor

Same task of about 100 ms, processor and memory energy per cell, with idle subtracted: native work 0.74 J, our cell 0.89 J, Firecracker with our same cell 0.98 J, Firecracker with Linux 2.04 J. Overhead over the native work: 0.15 J against 1.30 J. That is less than half the processor energy of a Firecracker microVM with Linux, for the same work. Against Firecracker with the same cell it is close to a tie (0.89 against 0.98 J), and with a new process per cell, 0.99 J, it ties. Full life cycle of a cell (boots, passes its test, shuts down), with no work: minimal 24 mJ, with network 28, with files 32, with network and files 38. 5 repetitions of 400 cells per arm, 0 failures. Coherent with kz-f1 f0: there the energy reading was taken from dom0 under Xen and failed; here the kernel is native, without a hypervisor (KVM).

limit  ZBook (one laptop), native kernel, synthetic load (a hash in a loop). This is chip energy (RAPL: processor and memory), not the wall plug or the electricity bill; it does not cover network, disk, fans or power supply. The saving against Firecracker with Linux comes from replacing Linux with the cell, not from the cell against Firecracker. No money figures. Still to repeat on a server and to validate against an external meter. Evidence recorded with hash and timestamp.

kz-c1 · bare metal  PASS

kz-c1, cell launcher with direct libxl

A cycle of about 38 ms (create 26 + life 2 + destroy 9) against about 94 ms plus harness with xl; 64/64 cells ok in each configuration and a ceiling of 50 to 60 cells/s per node from P=4, because creation serializes in the toolstack.

limit  Measured on the ZBook, before later generations: not the current cycle nor a latency headline. No confidence intervals.

Stage 6 · 2 labs

The cloud, kept in its place.

Nested runs orient and nothing else: the performance and isolation verdicts are the ones measured on metal.

kz-gc1 · cloud (nested)  PARTIAL

kz-gc1: the KVM cell in a cloud VM (nested KVM)

Cell ready from zero in 6.9 ms against 107 ms for untuned Firecracker + Linux. On ~100 ms jobs: 2.11x at the worst P, against untuned Linux the edge shows; with the same cell on Firecracker it drops to 1.04x (it orients). A burst of 20,000 cells without failures and 13,053 live cells at 1.86 MB each.

limit  Exploratory · in the cloud (nested) · it orients, it is not a verdict; the performance and isolation verdicts are the ZBook's. One round per point; the 13,053 was cut by time, not a RAM ceiling.

kz-az1 · bare metal + nested cloud  PARTIAL

kz-az1: duel on metal and runs on nested AWS

On metal, the 2x hypothesis on ~100 ms jobs holds only against untuned Linux (2.01x, borderline) and is falsified against tuned Linux (0.92x to 1.03x, technical tie). With the best measured configuration after the trim, kz-k2 ready 1.51 versus 2.32 ms for Unikraft and 1.03 versus 1.49 MiB per cell; before the trim Unikraft arrived first (2.25 versus 4.83 ms).

limit  One machine, one round per point. At P=15 Unikraft stays ahead on rate in 3 of 4 cells; with defaults Unikraft leads on ready time. Nested AWS orients, it is not a verdict. Snapshot-restore stays out of the rankings. The trim mechanism is not published.

Stage 7 · 11 labs

The reference code.

The components themselves, as standalone scripts that print their checks and exit 0 only if every vector matched its expected verdict. Mostly logic plane.

impl · gate · logic  PASS

Gate (admission gate for the Job)

42/42 checks without failures (rc=0); reservation without failures and indexed ledger 8/8. Across the 31 extended adversarial vectors (gate, broker, chain): 19 DENY, 9 declared GAPs, 3 LIMIT by design.

limit  Logic plane. The 9 GAPs are published without hiding them; C5 was the only high-severity one. The Gate's metal run is lab71.

impl · broker · logic  PASS

Broker (mediated I/O with CID)

71/71 checks (68 vectors, purity, derivation and signatures), 0 failures, rc=0. Covers the peer key pin: another certificate from the same CA with another key is denied.

limit  Logic plane. Declared GAPs remain (B2, B5, B6 by design, and B7, state rollback, which depends on dom0 durability). Not an isolation verdict.

impl · chain · logic  PASS

Chain (receipt and lineage chain)

37/37 scenarios without failures, rc=0. Signature coverage now attests the full set of fields (moving any gives BAD_SIGNATURE) and lineage requires a real destruction receipt.

limit  Logic plane. Starting findings that are published: lineage was vacuous because of a public constant and signature coverage was partial; both were fixed in the real code.

impl · observer · logic  PASS

Observer (countersignature and destruction receipt)

14/14 checks without failures and the destruction-receipt chain with rc=0. Removing or replacing the countersignature (6+3 variants) is denied with MISSING_COSIGNATURE.

limit  Logic plane. The observer countersigned without an observed commit fact (C5); it was addressed in the real code with a separate store. Independence in production is not claimed here.

impl · ring · logic  PASS

Ring (framing of the request and response channel)

62/62 vectors without failures, rc=0. The receiver validates only a snapshot of the page, and large artifacts travel in chunks.

limit  Logic plane (loopback on one machine, no real Xen channel). Python's fence is best-effort. Not an isolation or performance verdict.

impl · transport · logic  PARTIAL

Transport (mTLS for the control plane)

In logic: mtls 22/22, serve 21/21, integration 13/13 (real handshake, 0 SKIP), tier2 6/6, ca 18/18, identity 8/8. On metal, once: 12/12 suites with a real TLS 1.3 handshake. Another certificate from the same CA with another key is denied.

limit  Open: running it between distinct OS principals and between separate Xen domains. On metal both ends had the same UID; the current inter-domain channel is a relay without mTLS.

impl · crypto · logic  PASS

kzsig: Ed25519 asymmetric signing

15/15 checks, 0 failures: with public keys only the verifier checks but cannot forge; 300 random signatures, a reused signature and a signature made with the public key all fail; a dom0 signing with its own key is rejected.

limit  Pure logic. Authentically distributing the three public keys remains open; dom0 can still drop or delay; the private keys are not protected by this layer.

impl · kaf-store · logic  PASS

kaf-store: per-process isolated storage (closing C5)

22/22 checks (rc=0): asking for a signature over bytes that were never stored, storing garbage and declaring another hash, or writing straight to the medium, is denied; no private key reachable from the broker process.

limit  It closes only the compromised-broker-process case. Against dom0 with root, process isolation is not enough: that is a hardware residual. Not tested on metal in this directory.

impl · alicebob · logic  PARTIAL

alicebob: notary and key registry for console agents

Two adversarial reviews found and fixed about twenty defects (rewritable or truncatable chain, per-Job owner, injection in the audit query, among others); the Go tests for chain, signature, server and store pass.

limit  Finding NOT fixed in code: if daemon and agent run under the same UID there is no agent/human boundary. Linear log, not Merkle; no performance figure.

impl · rump-xen · bare metal  PASS

KZ-Rump cell on Xen PVH (and the same image on KVM)

Honest sequence: T15a link PASS; T15b the console FAILED; T15c KZ-RUMP-OK; T16a fs-only BLOCKED (UNCLEAR); T16b fs-only PASS 3/3; T16c net-only PASS, 8-thread contention without aborts and a deliberate deadlock that dies observably at 30 s. A single ELF boots on Xen and on KVM.

limit  n=1, one run per test. A first audit found an insufficient watchdog and it was fixed. Current boot figures are those of the trim, not of these reports.

mtls-metal · bare metal  PARTIAL

mtls-metal: control-plane mTLS on the physical dom0

12/12 suites green on metal with a real TLS 1.3 handshake (a certificate from the same CA with another key is denied). The harness between principals with distinct UIDs never ran.

limit  Both ends with the same UID and pure-Python TLS: not separation between OS principals nor between Xen domains. n=1.

Memory

The memory, in every cloud.

Whoever controls the host can read the memory of what runs on it: the cloud provider, or an attacker who gets root on the host. Every cloud has this (AWS, GCP, any other) unless you use confidential computing. It is not always said in these words, but that is how it works. We say it plainly so nobody takes it for a Kawésqar problem.

memory
Cloud withoutconfidential computingAWS, GCP, any cloudyour VM / your workloadmemory in the clearsecrethypervisor / hostrootreadsThe provider, or an attacker withroot on the host, reads the memory.Own server,Xen dom0measured on metal, n=1the cellmemory in the clearsecretdom0rootreadsRoot in dom0 reads the cell memory.Measured: 3 of 3, a secret borninside the cell.ConfidentialcomputingTDX · SEV-SNPyour VM / your workloadmemory encrypted▒▓░ x9#f 7a▒ ░▓hypervisor / hostrootblockedThe host sees only ciphertext.The read is blocked.Same security model in all three;the only thing that changes is memory encryption.Cloud withoutconfidential computingAWS, GCP, any cloudyour VM / your workloadmemory in the clearsecrethypervisor / hostrootreadsThe provider, or an attacker withroot on the host, reads the memory.Own server,Xen dom0measured on metal, n=1the cellmemory in the clearsecretdom0rootreadsRoot in dom0 reads the cell memory.Measured: 3 of 3, a secret borninside the cell.ConfidentialcomputingTDX · SEV-SNPyour VM / your workloadmemory encrypted▒▓░ x9#f 7a▒ ░▓hypervisor / hostrootblockedThe host sees only ciphertext.The read is blocked.Same security model in all three;the only thing that changes is memory encryption.
memory   Who can read what. Orange arrow: a read by the host. Only the third stack, with hardware encryption, stops it.
platform

It belongs to the platform, not to Kawésqar

Any workload on any host without confidential computing is readable from the host. It is a property of the platform, the same for every product that runs on it.

we tried it

We tested it in confidential cloud

TDX on GCP and SEV-SNP on AWS. Memory stays encrypted, but the cell cannot run inside (there is no KVM) and one confidential VM per job takes 50 to 58 s to come up.

what we ask

So the closure is validated on physical hardware

The cell closure is validated on physical hardware (TDX or SEV-SNP), which is what we are asking for.

How we use it

Critical stays home, the rest goes fast.

We do not go back against the move to the cloud. We use Kawésqar speed in the cloud for less critical workloads, or where the risk can be tolerated, and we keep at home only what deserves it: sovereign, on-prem or local. The most critical, such as a bank's keys, stays on your own hardware with full memory encryption (TDX / SEV-SNP).

ladder
Risk-tolerantworkload→ CloudKawésqar speedSensitiveworkload→ On-premsovereign, at homeA bank's keys→ On-prem with full memory encryptionTDX / SEV-SNPvalidated on physical hardwareless criticalmore criticalRisk-tolerantworkload→ CloudKawésqar speedSensitiveworkload→ On-premsovereign, at homeA bank's keys→ On-prem with full memory encryptionTDX / SEV-SNPvalidated on physical hardwareless criticalmore critical
ladder   Where each kind of workload goes. Only what deserves it stays at home.
same contract

Same contract, same Gate

On both sides the same contract, the same Gate and the same security model apply. Only where the workload runs changes.

measured

What is measured today

Closing the cell with full memory encryption is validated on physical hardware (TDX or SEV-SNP), which is what we ask Intel and AMD for. Today the residual is measured and declared (lab54b, 3 of 3), not closed.

not yet

What does not exist yet

Federation between on-prem and cloud is a design direction, still without code.

Keep reading

Evidence, then the ask.

The evidence page puts the headline figures in context and names the gaps still open. If you want to reproduce something here, tell us plainly.

the rule

Not a certificate, not a claim of regulatory compliance: a dated, bounded record of what ran.

Agent Fabric · cuaderno de laboratorio

Más de 70 labs. Los que fallaron también.

Este es el cuaderno de laboratorio del Agent Fabric: cada experimento que podemos publicar, con el plano en que corrió, el veredicto, la cifra y el límite. Los fallos y las refutaciones están junto a los aciertos, porque un laboratorio que solo muestra lo que funcionó te muestra un pitch, no evidencia.

0+
experimentos en el laboratorio, y contando
0
mostrados aquí, cada uno con su veredicto, plano y límite
0
probados: PASS, HELD o mitigados por política
0
publicados como FAIL o REFUTED, no enterrados

patent pending  Solicitudes de patente presentadas ante INAPI (Chile); en trámite.

Tablero de veredictos

Cada lab, por plano.

Una ficha por lab. Verde es un acierto probado, ámbar es parcial, poco claro, abierto o un residual medido, naranja es un fallo o una refutación registrada. El plano importa tanto como el color: nuestros resultados corren anidados (dentro de una VM, lo que no valida aislamiento ni rendimiento) o en bare metal, el único lugar donde esas afirmaciones pueden cerrarse. Metal es una estación portátil, n=1, salvo que la tarjeta diga otra cosa.

VM anidado · 16 labs

12 probados · 2 parciales o abiertos · 2 fallidos

Nube anidada · 1 lab

0 probados · 1 parciales o abiertos · 0 fallidos

El lab72 incluye además un modelo TLA+ contrastado con trazas, contado aquí en bare metal. En la nube mostramos solo lo que orienta; nunca cierra un veredicto.

Lo que falló

Lo que falló, dicho claro.

Un laboratorio que solo publica sus aciertos está vendiendo. Estos son los lugares donde los datos dijeron que no, o dijeron menos de lo que esperábamos, y qué cambiamos después.

VM anidado  FAIL

lab14 · El guest rump no imprimió nada

FAIL en Xen anidado. El guest llegó a modo largo y ejecutó rump_init, pero dos fallos de consola impidieron que apareciera la marca. El fallo dictó el orden de arreglos que cerró el lab16.

VM anidado  FAIL

lab15 · El observador no falló cerrado

F9 FAIL: 38.577 mutaciones produjeron 4.520 crashes, 85 falsas contrafirmas (solo con el stream comprometido) y 56 colgados. El lab29 reprodujo después el fallo en metal, endureció el observador y llegó a 0 crashes en 3.000 variantes.

bare metal  FAIL

lab25 · La celda desechable filtró en disco

La RAM resistió (0 hits), pero un volumen reutilizado dejó que la celda B leyera el patrón de A: 6.882.854 hits. Cerar el volumen completo al crear lo dejó en 0, un arreglo verificado en el lab antes de llegar al producto.

bare metal  REFUTED

lab64 · Gate más orden costaron más de 1 ms

Pre-registramos menos de 1 ms. La mediana medida es 2,735 ms (P=1, n=200). La hipótesis se refuta y esa cifra es la que publicamos.

bare metal  REFUTED

lab70 · El Gate más rápido tampoco llegó a la meta

H1 pedía menos de 1,5 ms de pared. El brazo completo añade +3,210 ms (IC95 3,074 a 3,378), y +3,216 ms otra vez con el governor performance. Refutada dos veces.

bare metal  FAIL

KZ-F1 fase 0 · C1

FAIL según el criterio pre-registrado: la vía de registros devuelve error de E/S bajo Xen 4.20, antes y después de la carga. Una interfaz alternativa respondió y siguió la carga, pero no está validada (n=1, sin medidor externo) y dos comprobaciones registradas sobre ella incumplieron su criterio. Solo afecta a la medición de energía, y no se cita ninguna cifra de energía.

bare metal  REFUTED

KZ-F1 fase 2 · H7b

Pre-registramos como máximo 1.800 líneas de código para el camino de producción. El conteo es 7.840, y por encima de 2.000 se da por refutada. La hipótesis queda refutada y se publica la cifra tal como se midió, sin redefinir el conteo.

bare metal  PARTIAL

lab20a · Una observación que no debemos inflar

Se vio una denegación DMA una vez y no se pudo reproducir; en modo strict el propio control positivo fue denegado. Es una observación, no un resultado, y no la citamos como «strict deniega transferencias no mapeadas».

lógica  PARTIAL

C5 · El cierre no se ganó

Un broker que forja la confirmación de almacenamiento cruzó la compuerta en la primera ola lógica (23/23 comportamientos reproducidos). Un trabajo posterior añadió un almacén separado; la tarjeta de kaf-store muestra qué cierra y qué no.

bare metal  TIE

lab27 · «Rump expone menos» era falso

La hipótesis quedó falsada: rump mantiene 2 event channels vivos contra 0 de Linux mínimo, y empatan en grants y nodos de xenstore. Concluimos que rump no justificaba por sí solo una inversión profunda.

bare metal + nube anidada  PARTIAL

kz-az1 · El 2x se sostiene solo a medias

En trabajos de 100 ms la ventaja de 2x se sostiene contra Linux sin ajustar (2,01x, al borde) y se falsa contra Linux ajustado (0,92x a 1,03x). Antes del recorte, Unikraft llegaba antes a lista. Empezamos por lo que no se sostuvo.

Línea de tiempo

Seis etapas, un hilo.

El cuaderno sigue el orden en que trabajamos: primero el sandbox, luego Xen anidado en una VM, luego la máquina física, luego los ataques a nuestro propio diseño, después las celdas rápidas y al final la nube. La altura de cada barra es la cantidad de labs mostrados en esa etapa; el código de referencia (11 tarjetas más) queda aparte, al final.

línea de tiempo
2SANDBOX14XEN ANIDADO27BARE METAL16RED TEAM15CELDAS KZ2NUBESANDBOX2XEN ANIDADO14BARE METAL27RED TEAM16CELDAS KZ15NUBE2
línea de tiempo   Labs mostrados por etapa. Un resultado anidado o de nube nunca cuenta como veredicto de aislación ni de rendimiento: solo el metal puede cerrarlos.
Etapa 1 · 16 labs

Xen anidado, las primeras respuestas.

Dentro de una VM (Xen bajo QEMU/KVM). Estas corridas muestran la arquitectura funcionando como lógica sobre dominios reales. Nunca son un veredicto de aislamiento ni de rendimiento, y cada tarjeta lo dice.

Cada tarjeta: lab, plano y veredicto, luego lo que se encontró y luego el límite que lo acompaña. Las cifras son las que registró el lab; cuando una campaña posterior las reemplazó, la tarjeta lo dice.

lab03 · VM anidado  PARTIAL

Rump kernels NetBSD: cierre de incógnitas

La pila IP real usa 52 hypercalls (base 46) y se invocan 28 de 46 al iniciar; en el orden oficial gana la definición real en 3 de 3 duplicados; el aviso sysctl es benigno.

límite  Un solo guest anidado; build reproducible parcial (186/190 idénticas); sin conclusión de rendimiento ni aislamiento.

lab04 · VM anidado  PASS

No determinismo del build, disco del guest, tests del skill

La diferencia eran solo rutas de fuente embebidas: dos builds MKREPRO salen idénticos 190/190. Corrige la hipótesis del lab03; tests del skill 12/12 tras reejecutar uno.

límite  n=2 builds por condición; la causa exacta es inferencia fuerte, no hecho; un caso falló en la primera pasada.

lab05 · VM anidado  PASS

Xen anidado y celda PVH mínima

La celda imprime su marca y termina sola; 5 ciclos y 1 destroy durante el arranque dejan 0 dominios y 0 claves residuales en xenstore.

límite  Todo anidado y sin IOMMU: no vale para rendimiento ni aislamiento; no es benchmark.

lab06 · VM anidado  PARTIAL

Validación de skills y celda Xen anidada (A a D)

A: 10/12 PASS; B: 7 PASS y 1 FAIL (un caso inventó un formato); C: 5/5, 30 ciclos con 0 residuos y 0 fallos; D: 5/5. El conteo de hypercalls rump es 46 en ambos lados.

límite  Anidado, sin IOMMU; tres casos no se pueden reproducir porque su prompt y criterio no están en disco.

lab07 · VM anidado  PASS

S4.0 sonda xenstore entre dos celdas

Sin ACL la lectura cruzada da EACCES (30/30 en el control); la celda dueña concede el acceso y la otra lee el nonce y recibe 3 eventos. dom0 no es requisito.

límite  Anidado, N=1 por brazo, una sola variante de xenstored; sin cuotas ni concurrencia.

lab07b · VM anidado  PASS

S4.2 anillo gntalloc+gntdev+evtchn entre dos celdas

Ring bidireccional con hashes idénticos en ambos extremos, sin privilegio de dom0. La primera corrida FALLÓ por un protocolo propio defectuoso y se corrigió.

límite  Anidado, n=1 PASS; N≥30 pendiente; no se probó que un tercer dominio no pueda mapear el grant.

lab08 · VM anidado  PASS

Dominios placeholder de control, broker y observer

30 ciclos (90 dominios): 0 residuos en XenStore y 90/90 con código 0; una Cell de prueba recibió acceso denegado en 15/15 operaciones sobre las ramas de los tres roles.

límite  Anidado, una corrida, dominios placeholder sin lógica real; no es veredicto de aislamiento.

lab09 · VM anidado  PASS

Lógica de los roles dentro de un initramfs

Las tres suites pasan dentro de un dominio PVH de 256 MB (17/17, 31/31, 9/9) con ~167-172 MB libres y 0 señales de falta de memoria.

límite  Anidado, n=1; solo viabilidad funcional y holgura de memoria.

lab10 · VM anidado  PASS

Broker en un dominio y Cell por el ring

5/5: artefacto de 9000 B con hash idéntico en ambos sentidos; CID ajeno o forjado, repetición y destino forzado negados; 0 procesos del par en dom0.

límite  Anidado, n=1; sin conclusión de rendimiento ni de resistencia frente a Xen en metal.

lab11 · VM anidado  PASS

Observer que contrafirma y emite recibo de destrucción

4/4: contrafirma solo con la Cell viva durante un STORE real; recibo de destrucción solo después de destruirla; 17/17 intentos de leer, alterar o borrar el flujo negados.

límite  Anidado, n=1; no cubre una Cell hostil con capacidades de kernel.

lab12 · VM anidado  PASS

Cadena A, recibo, B sobre dominios reales

Las 4 condiciones se cumplen: B no existe antes de la orden, sin contrafirma el Gate niega, una Cell que intenta firmar la orden recibe acceso denegado (18 veces) y el hash es idéntico de principio a fin.

límite  PASS acotado a arquitectura: anidado, n=1, firma simétrica de la primera versión; el primer intento falló por empaquetado.

lab13 · VM anidado  PASS

B1.8 Falsaciones que solo el sustrato puede refutar

Las 3 falsaciones dan PASS: CIDs del epoch anterior dan STALE_EPOCH tras recrear el broker; un tercer dominio no mapea el grant (128 refs denegadas); un kill entre STAGED y COMMITTED da NOT_COMMITTED. Hallazgo: epoch y nonces deben ser durables fuera del dominio.

límite  Anidado, n=1, firma HMAC v0; el instante del kill está forzado; dom0 sigue en la base de confianza.

lab14 · VM anidado  FAIL

T15b Primer arranque de rump como guest PVH bajo Xen

FAIL instructivo: el guest entró por PVH, pasó a modo largo y ejecutó rump_init, pero no imprimió la marca. Dos fallos de consola (hypercalls sin programar y un buffer eliminado por el compilador) dictaron los arreglos que cerró el lab16.

límite  Anidado, n=1; quedó sin observar si rump_init devolvió 0.

lab15 · VM anidado  FAIL

S9.2 Falsaciones 8 y 9 del observador

F8 PASS: el observador sobrevive a la caída de una celda y el Gate no usa su autodeclaración. F9 FAIL: con 38.577 mutaciones hubo 4.520 crashes, 85 falsas contrafirmas (solo con el stream comprometido) y 56 colgados. Un crash no firma: es fallo de disponibilidad.

límite  Anidado, n=1, HMAC v0; el engaño es un guion fijo; dom0 sigue en la base de confianza.

lab16 · VM anidado  PASS

T15c Cierre del eje rump: el guest rump imprime KZ-RUMP-OK bajo Xen

Con los dos arreglos de consola, el guest rump imprime la marca, rump_init devuelve 0 y apaga por hypercall con código 0. Cierra el bit que el lab14 no pudo leer.

límite  Anidado, n=1: solo init y consola; sin rendimiento ni aislamiento duro.

lab17 · VM anidado  PASS

T16 Guest adversarial: A_effective contenido en A_delegated

289 intentos de mapear grants ajenos y 289 de atar canales de evento dieron 0 éxitos, 0 páginas ajenas leídas, 0 auto-concesión. Se separa lo que contiene Xen de lo que contiene Kawésqar (replay, STALE_EPOCH, NOT_COMMITTED).

límite  Anidado, n=1; la garantía supone un Xen correcto; las denegaciones de grants y canales son de Xen, no de Kawésqar.

Etapa 2 · 12 labs

La primera campaña en metal.

En el dominio de control físico de una estación portátil, n=1. Aquí aparecen los primeros fallos honestos: un canal encubierto que no pudimos medir y un disco que filtraba.

lab18 · bare metal  PASS

Baseline físico y gate del plano de laboratorio

Inventario capturado y preflight 11/11. Los CPU de este Xeon no están afectados por L1TF, MDS ni TAA, así que se re-acotó el lab de canal lateral.

límite  ZBook, n=1, un solo día; es inventario y preparación, no un resultado de aislación ni de rendimiento.

lab19 · bare metal  PASS

Re-atestiguación en metal del núcleo de la arquitectura

7/7 veredictos del núcleo PASS en el dominio de control físico. Ciclo de una celda mínima: p50 0,95 s, p99 1,09 s, 200/200 sin fallo, residuo real 0.

límite  ZBook, n=1, dominio de control capado a 4 vCPU: corrección, no rendimiento; la cadena sobre dominios reales sigue en HMAC.

lab20a · bare metal  PARTIAL

Aplicación de VT-d en bare metal con un dispositivo DMA nombrado

Se atestiguó la configuración de VT-d y se vio una denegación una vez, pero no es reproducible. Bajo modo strict el instrumento quedó UNCLEAR: el control positivo fue denegado.

límite  ZBook, n=1. No debe citarse como «strict deniega transferencias no mapeadas»: es una observación, no un resultado probado.

lab21 · bare metal  PASS

Latencia base de una celda en bare metal

200/200 por variante, 0 residuos: crear con xl 184 ms (p50) y arrancar kernel e init 607 ms (p50, p99 790 ms). Una celda con trabajo agrega un overhead acotado.

límite  Cifras de la generación inicial (kernel Linux y xl), muy anteriores a las celdas actuales: no compararlas con ellas.

lab22 · bare metal  PASS

Escala y carga

0 fallos del Gate en 256 corridas bajo carga; 2.775 ciclos completados al 100 %; creación estable en torno a 237 ms de N=1 a N=414; 0 celdas residuales.

límite  Es la lógica del Gate bajo presión de scheduling, no transiciones reales de celda a escala; el techo de recursos nunca se estresó.

lab23 · bare metal  UNCLEAR

Canal encubierto por contención

El control positivo no cruzó el umbral: mejor caso 9,64 bps con IC 95 % [-33,06 ; 52,31], indistinguible del ruido. No se afirma ni presencia ni ausencia de canal.

límite  Instrumento no validado; se registra como canal no medido, nunca como despreciable.

lab23b · bare metal  MITIGATED

Canal encubierto: cierre por política

Riesgo cerrado por política de ubicación (no co-agendar celdas que no confían entre sí en hilos hermanos), no por medición: el canal SMT más fuerte no existe por construcción.

límite  No es «probamos que no hay canal»: los canales por caché compartida y memoria entre núcleos quedan como riesgo residual.

lab24 · bare metal  OPEN

XSA que aplican de verdad

Tres avisos parcheados, las especulativas inaplicables por silicio o mitigadas; queda una aplicable sin parche (un DoS de liveness en el daemon de xenstore). Ninguna es escalada ni fuga entre dominios.

límite  Estado atado a una versión de paquete; hay que re-correr al cambiar de build.

lab25 · bare metal  FAIL

Celda desechable: residuo de RAM y disco

RAM PASS: 0 hits con solape forzado de ≥548 MiB. Disco FAIL: la celda B ve el patrón de A (6.882.854 hits); cerar el volumen completo al crear lo deja en 0 hits.

límite  La mitigación se verificó en el lab pero no estaba aún en el producto; residuo por DMA fuera de alcance.

lab26 · bare metal  PARTIAL

Cierre de la campaña bare metal

Arquitectura: PASS acotado, cierre condicional. Sin amplificación de autoridad ni de evidencia en Xen real (7/7) y bajo carga; celda desechable falsada en disco y PASS en RAM.

límite  No se afirma aislamiento físico verificado, ni celda desechable incondicional, ni tesis completa; canal encubierto UNCLEAR.

lab27 · bare metal  TIE

Shootout de frontera: KZ-Rump contra Linux mínimo

La hipótesis «rump expone menos» quedó falsada: rump mantiene 2 event channels vivos (Linux mínimo 0); empatan en grants (0) y en 38 nodos de xenstore. Hallazgo: un aborto del adaptador daba reason code 0, igual que un apagado limpio.

límite  Superficie observada, no alcanzable; ambos runtimes son casi stubs; una máquina, una corrida.

lab28 · lógica + bare metal  PASS

Árbol de delegación n>1 y no-amplificación por inducción

Gate real: 1.958 aristas decididas, 0 amplificaciones; celdas reales en el borde exacto del envelope (768 de 768 MiB); un nieto que pide más que su padre es negado con AUTHORITY_NOT_SUBSET; 5 mutantes detectados.

límite  El árbol de parentesco es modelado (Xen no lo conoce); profundidad mayor que 2, fan-in y fuzz son lógica. El lab halló defectos reales del Gate, ya corregidos.

Etapa 3 · 15 labs

Llevando el metal al límite.

Todavía en metal: fuzzing del observador, la cadena firmada dentro de celdas reales, saturación hasta 412 celdas vivas, IOMMU estricto e hypercalls observadas.

lab29 · bare metal  PASS

Falsación 9: superficie del observador

La primera corrida FALSÓ la propiedad: 47 crashes y 208 violaciones por huecos alcanzables vía JSON. Tras endurecer el observador: 0 crashes y 0 violaciones en 3.000 variantes, 3 mutantes detectados, 0 residuos.

límite  PASS acotado al cable (stream JSON decodificado); quedan violaciones de familia inproc declaradas como GAP. n=1.

lab30 · bare metal  PASS

Falsación 10: STORE transaccional bajo kill

En ninguna etapa quedó un COMMITTED parcial entregable (7 kill -9 reales); el replay tras restart dio STALE_EPOCH siempre; 3 mutantes detectados; celdas reales destruidas.

límite  La persistencia del journal es modelada por el harness: prueba la atomicidad del contrato, no el CAS persistente de producción (GAP declarado).

lab31 · bare metal  PASS

Gate adversarial bajo carga

0 aceptaciones sobre 6.870 intentos adversariales concurrentes; 1.194 de 1.194 legítimas cruzan bajo la misma carga; la carrera ingenua sobre-reserva 6 de 6 (el instrumento ve la carrera); 0 residuos.

límite  El lock del ledger es en proceso; el CAS persistente de producción queda fuera. Corpus finito, no prueba de ausencia general.

lab32 · bare metal  PASS

Soak de estabilidad del ciclo crear-correr-destruir

400 de 400 iteraciones: deriva cero (tau de Kendall 0,02 a 0,08), residuo 0, memoria libre de Xen 55.883 a 55.883 MiB, dom0 estable.

límite  Ventana acotada (no meses), una celda mínima, una máquina, n=1.

lab33 · bare metal  PASS

Saturación a escala máxima y throughput

Pico de 412 celdas vivas simultáneas (52.736 de 53.835 MiB), 0 amplificación y 0 residuo en los 14 niveles; máximo 14,78 celdas/s de creación y 8,98 celdas/s de ciclo completo.

límite  n=1. En el pico la destrucción llega a p99 32 s por la serialización del toolstack; a escala solo una muestra lleva consola con marca.

lab34 · lógica + bare metal  PASS

Fan-out de recursos (Falsación #4)

En lógica: 4.681 nodos y 7.608 aristas con 0 amplificaciones; en metal, con celdas de 128 MiB, todos los escalones completan, 0 amplificaciones y overhead por celda de 1,5 a 2,5 MiB.

límite  Los dos primeros intentos físicos quedaron UNCLEAR (la causa se halló en lab40); el cierre es el tercero. n=1.

lab35 · bare metal  PASS

Frontera mínima observada de la celda bajo carga

Clasificación completa de 18 interfaces y 49 hypercalls; la carga no rompió la frontera, 0 residuo en los 4 escenarios. El aislamiento lo acota Xen, no el runtime.

límite  Caracterización, no veredicto de aislamiento. La observación por traza quedó UNKNOWN en las 49 filas por falta de decodificador (lo cerró lab44).

lab38 · bare metal  PASS

Cadena criptográfica Ed25519 de extremo a extremo dentro de celdas

La cadena celda A, contrafirma, Gate, celda B y recibo de destrucción corre en metal; 4 mutantes (forja por dom0, recibo alterado, B sin evidencia, replay) y 7 vectores negativos rechazados sin crear B.

límite  La verificación de la firma antes de almacenar la modela el harness, no el Gate de referencia. n=1.

lab39 · bare metal  PASS

Humo y regresión de invariantes bajo IOMMU estricto

Los 4 grupos pasan: 60/60 iteraciones, 0 aceptados de 3.444 intentos adversariales con 600/600 legítimos admitidos, y puesta a cero verificada; el create sube cerca de 12 % (unos 23 ms) bajo estricto.

límite  Regresión acotada, no re-prueba exhaustiva; el A/B limpio en un mismo arranque quedó pendiente. n=1.

lab40 · bare metal  PASS

Piso de memoria de arranque de una celda PVH

Piso medido de 96 MiB: con 64 MiB la celda no arranca, con 96 y 128 MiB sí. Cierra la causa raíz del fallo de lab34 y falsa la hipótesis de carrera del harness.

límite  Piso grueso (n=3 por punto), válido para este kernel e initrd de laboratorio; el valor exacto entre 65 y 96 MiB no se afinó.

lab41 · bare metal  PASS

Auditoría de grants de dom0 sobre páginas de la celda

Una celda sin disco, quieta o bajo carga, tiene exactamente 1 grant activo mapeado por dom0 (el anillo de consola), 0 páginas de datos; con disco aparecen 67 refs (control positivo).

límite  La primera corrida fue UNCLEAR por control positivo ciego y se corrigió. Que entren al IOMMU de dom0 es inferencia de la fuente. n=1.

lab42 · bare metal  PASS

Frontera de entropía dentro de la celda

14 de 14 claves distintas; en ningún arranque se generó clave con el generador aleatorio sin listo, ni siquiera con RDRAND apagado (mediana 1,47 s contra 0,14 s).

límite  Prueba disponibilidad y no determinismo, no la calidad de la entropía. El caso sin RDRAND se apoya en lectura de la fuente. n=1.

lab43 · bare metal  PASS

Cadena criptográfica en celda con n mayor que 1

La cadena y el abanico cruzan de punta a punta y todos los mutantes se rechazan con su razón exacta; el abanico llena el envelope exacto (2 por 128 MiB). Hallazgo: la compuerta pura no consume el enlace prev de los recibos de destrucción.

límite  La verificación de la firma y el enlace prev quedan modelados en el arnés (brecha de producción declarada). n=1.

lab44 · bare metal  PASS

Hypercalls observadas por traza real

En los 4 escenarios lo observado es subconjunto del conjunto declarado de 10 hypercalls; la celda rump deja una huella mínima (5 números distintos) y Linux añade 4 más. Control negativo: 0 llamadas con la celda pausada.

límite  Se observa el número de hypercall, no el sub-op ni los argumentos. n=1.

lab45 · bare metal  PASS

Soak bajo IOMMU estricto y fuga de estructuras

400 de 400 ciclos sin fuga: grants activos planos en 2, maptrack plano en 4 y memoria libre de Xen constante (55.883 MiB antes y después).

límite  La memoria libre se lee en MiB (mínimo detectable cercano a 80 KiB por ciclo); los grants activos son un proxy parcial. n=1.

Etapa 4 · 16 labs

Atacar y luego verificar fuera de dom0.

La decisión sale del dominio de control, luego atacamos nuestro propio diseño, medimos lo que dom0 aún puede leer y comparamos contra Linux.

lab46 · lógica  PASS

Conformidad de denegación y agente vivo contra el canal real

12 de 12 checks y oráculo independiente en PASS, cada denegación con su razón exacta; 0 escaladas cruzadas con los 8 modelos de lenguaje probados, incluidos los 3 que sí atacaron por el canal real.

límite  Nivel lógica, no metal ni agente dentro de la celda. Una campaña posterior corrigió sobreafirmaciones del propio lab (mutantes que mutaban datos, no código).

lab47 · bare metal  PASS

Verificador fuera de dom0 (cadena A a B)

El control positivo cruza y los 11 mutantes se deniegan con su razón exacta; creaciones = admisiones = PASS del verificador (1 = 1 = 1); las forjas y repeticiones de dom0 mueren por firma. Re-atestiguado en metal: 3 de 3 corridas PASS.

límite  n=1, un host. dom0 sigue creando dominios y puede leer la RAM de un dominio vivo (residual de hardware). La primera corrida física fue UNCLEAR por un bug del oráculo, corregido.

lab48 · bare metal  PASS

Compuerta endurecida (gate_verified) contra harness

Con el arnés apagado, la compuerta vieja ACEPTA y crea el hijo en los 4 mutantes principales; la nueva los DENIEGA con su razón. Las propiedades pasan de modeladas en el arnés a impuestas por la compuerta.

límite  n=1. dom0 seguía generando las claves privadas en esta corrida (lo ataca lab47); un tap mentiroso no queda cubierto.

lab49-c5 · lógica + bare metal  PASS

Cierre de C5: broker comprometido

El observador viejo acepta ambas falsificaciones; kaf-store niega las dos y el testigo de bytes niega la primera. En metal, el broker comprometido que firmó sin respaldar los bytes deja el medio vacío y no cruza.

límite  El testigo deja pasar un COMMITTED falso sobre bytes honestamente en staging. Esto no zanja el marcador posterior de red team, que registra el cierre de C5 como no ganado. n=1.

lab50-dma · bare metal  PASS

Confinamiento DMA de un dispositivo asignado

dom0 lee el pendrive antes de asignar, pierde el dispositivo al asignar, y el domU lee el mismo número de sectores con 0 faults VT-d en modo estricto. Mide confinamiento, no aislamiento adversarial.

límite  Parcial: no se probó un driver malicioso intentando DMA a memoria de otro dominio. Corrió en HVM porque PVH no soporta passthrough. No citar como aislamiento DMA cerrado.

lab51-netless · bare metal  PASS

lab51-netless · celda con kernel sin pila de red

Una celda PVH con un kernel compilado sin pila de red arranca y corre: /proc/net, /sys/class/net e iplink AUSENTES y la carga cierra con sum=500500, con la consola lista en menos de 1 s de arranque de kernel.

límite  Hecho de construcción de un kernel, no veredicto de aislamiento ni de rendimiento; la evidencia es la consola y la configuración. n=1.

lab52-smt · bare metal  PASS

lab52-smt · placement contra el canal SMT

Un vecino en el hilo hermano ralentiza a la víctima 4,98 a 5,94 veces; en otro core cae a 1,07 a 1,13 veces, casi la línea base. El placement suprime el canal SMT (3 corridas).

límite  Mide interferencia de throughput, no una tasa de bits ni extracción de clave. El residual de ~1,08x es el canal entre núcleos (L3). n=1.

lab53-llc · bare metal  PARTIAL

lab53-llc · canal entre núcleos por la L3

El canal entre núcleos existe y es pequeño: 1,03 a 1,14 veces de interferencia (promedio ~1,08x), un orden de magnitud bajo el canal SMT. Residual medido y acotado, no cerrado.

límite  Sin Intel CAT/RDT no se puede particionar la L3 en esta CPU. Señal pequeña y ruidosa (3 corridas); mide throughput. n=1.

lab54-meminspect · bare metal  CONFIRMED

lab54-meminspect · el dominio de control lee una celda viva

Control positivo: dom0 mapeó 10.076 marcos de la celda viva y encontró el secreto sin cooperación de la celda (2 corridas). La firma autentica, pero no da secrecía frente al dom0.

límite  Mide la exposición, no la cierra. Solo la cierra memoria cifrada por hardware (SEV-SNP/TDX); esta CPU no la tiene. n=1.

lab54b · bare metal  CONFIRMED

lab54b-meminspect-incell · secreto nacido dentro de la celda

3 de 3 corridas con 3 secretos distintos: 1 candidato por corrida, hash verificado, 977 marcos contiguos. El residual de secrecía frente al dom0 queda medido.

límite  No cierra el residual: pide memoria cifrada por hardware (TDX/SEV-SNP), que esta máquina no tiene, y esa prueba de cierre no se ha corrido. El mismo residual existe en toda nube, AWS y GCP incluidas, si no usas computación confidencial: quien controla el host, el proveedor o un atacante que consiga root en él, puede leer la memoria de lo que corre en él. Es una propiedad de la plataforma, no de Kawésqar. Sí probamos VMs confidenciales en ambas nubes (TDX en GCP, SEV-SNP en AWS): la memoria queda cifrada, pero la celda no puede correr adentro (no hay KVM) y una VM confidencial por trabajo tarda de 50 a 58 s en levantar. n=1.

lab55 · bare metal  PASS

lab55-composition · las cinco propiedades juntas

5 de 5 corridas PASS: el gate endurecido decide fuera de dom0, sin clave privada en dom0; las órdenes forjadas o repetidas mueren por firma y los mutantes se deniegan. Cierra el pendiente «Composition».

límite  n=1, 5 repeticiones (reproducibilidad, no estadística). dom0 sigue creando dominios e inspeccionando memoria (residual de lab54).

lab56 · bare metal  PARTIAL

lab56-substrate · NetBSD completo o Linux recortado

Instanciación: NetBSD PVH ~1,65 s, Linux PVH 2,06 s, NetBSD PV 3,19 s. Ambos sustratos sostienen 8 celdas concurrentes con escalado casi plano de N=4 a N=8.

límite  Cifras de etapa temprana, anteriores a las celdas actuales; el texto del lab es inconsistente con su propia tabla, por eso solo se cita la tabla. n=1.

lab57-gate · bare metal  PASS

Celda rump real bajo el Gate

Corrida honesta PASS y 5 mutaciones dan su DENY nombrado antes de correr; cadena de 3 celdas reales, 8/8 celdas gateadas en paralelo, 20/20 en estabilidad. Una celda «eco» sin rump pasaba con testigo solo-bytes y se niega al atestar el kernel.

límite  PASS acotado, NO contención probada: la consola es un medio degradado y dom0 puede leer la RAM de la celda (residual de hardware). n pequeño.

lab58-overhead · bare metal  PASS

Overhead de Xen y de la celda rump frente a Linux

Ciclo sin trabajo de la celda rump ~94 ms frente a ~500 ms de la celda Linux (5,3x); boot 0,9 ms frente a 237 ms; Xen cobra +8 % de cómputo en dom0 y +11 % en una celda Linux PVH frente al metal desnudo; la celda rump queda a +4 % del nativo.

límite  n=50 por celda, un equipo. Una anomalía en la fase sha de rump (+15 a +20 %) queda UNCLEAR; no mide E/S. Cifras de create anteriores al lanzador directo.

lab59 · lógica + bare metal  PARTIAL

Red-team de escape de celda (campañas A a F)

En metal una celda comprometida leyó 0 páginas de un vecino, abrió 0 canales no autorizados y no ganó autoridad; la limpieza por prefijo del lanzador sí mató celdas de otro trabajo (cerrado en lógica); fuzz nativo 85M ejecuciones sin caídas; 15 de 16 ítems del ledger cerrados en lógica.

límite  Lo cerrado es plano lógico y no se ha vuelto a medir en metal. Siguen abiertos: el verificador que no re-hashea los bytes, el canal de caché entre núcleos y dom0 leyendo la RAM de la celda.

redteam-escape · lógica  PARTIAL

redteam-escape · Ola 1 de ataques lógicos (C5, E5, E6, E8, CGAPS)

C5: el cierre no se ganó (un broker que forja la confirmación de almacenamiento cruza; 23/23 comportamientos reproducidos); E6: la limpieza por prefijo mata celdas de otro trabajo; E8: la comprobación de memoria admite 11 celdas de más; E5: DENY. Los fallos se repararon después en lab59.

límite  Plano lógico en un solo proceso, sin metal; este lab registra el estado previo al arreglo.

Etapa 5 · 15 labs

Celdas, cadena firmada y reloj.

Lanzadores directos, el costo del Gate, imágenes selladas, el constructor mínimo en tres fases y la energía por celda. Aquí se refutaron tres hipótesis y falló una comprobación, y todo se conserva.

lab61 · lógica  PASS

Negativos de Secure Boot y mapa adversarial

5 de 5 casos como se esperaba: el control válido verifica y sin firma, clave ajena y certificado desemparejado se rechazan; el artefacto sin SBAT se detecta como tal.

límite  Verificación de firma offline: el rechazo del cargador en el arranque real está descrito como esperado, no medido; 17 dimensiones adversariales quedan pendientes.

lab62 · bare metal  HELD

kz-k2 en KVM frente a Firecracker

Con la celda anterior al recorte, lista desde cero en 4,8 ms (kz-k2 en KVM) contra 8,2 ms de Firecracker con la misma celda; frente a Firecracker con guest Linux la hipótesis pre-registrada se sostiene con 2,14 a 2,34x; a P=8 en celda mínima, 251 contra 147 celdas/s.

límite  n=1. Con la misma celda la ventaja es de 1,04 a 1,10x en trabajos de 100 ms: está en la celda, no en el VMM. Con trabajos de 1000 ms se pierde (~1x). Frente a Firecracker con un Linux sin ajustar, 3,5x medido en la misma máquina.

lab63 · lógica  PASS

lab63 · cadena Gate a launcher con Ed25519

107 pruebas negativas pasan (recibo alterado, contrafirma ausente o ajena, linaje roto, orden mal firmada, replay, replay tras reinicio) y la clausura de las 5 imágenes da 84 de 84 comprobaciones.

límite  Probado en lógica y construcción estática, sin Xen ni VM; la evidencia punta a punta en Xen con Ed25519 está en labs posteriores.

lab64 · bare metal  REFUTED

lab64 · costo del Gate y la orden sobre el ciclo de la celda

La hipótesis pre-registrada «menos de 1 ms» se refuta: Gate más orden suman una mediana de 2,735 ms (P=1, n=200); el delta A/B del tiempo total es +2,574 ms, y +10,815 ms a P=8.

límite  n=1, gobernador ondemand. A P=8 el Gate hace cola por un cerrojo único, así que no es «el Gate es más lento bajo carga». Es un fallo registrado y publicado como rasgo.

lab65 · bare metal  PASS

lab65 · verificador fuera de dom0 sobre celdas rump reales

En metal 16 de 16 escenarios según lo predicho; la extensión con sustitución de artefacto da 19 de 19 en tres corridas, 0 huérfanos; en lógica 62 pruebas de fidelidad y 10 de mutación pasan.

límite  n=1. Un dom0 que miente de forma coherente obtiene PASS, declarado NO-RECLAMO deliberado. El estado del Gate vive en la RAM del verifier. Canal por relé de dom0, sin mTLS. Veredicto acotado, no «contención probada».

lab66 · bare metal  PARTIAL

lab66 · kz-c2 con kernel sellado (cierre del TOCTOU de lab64)

Verificar el hash en cada ciclo cuesta +7,16 ms sobre un ciclo de unos 15 ms y se descarta; la segunda variante cuesta -0,27 ms en total_ms (15,345 a 15,079), dentro del ruido, con 9 pruebas negativas OK.

límite  El ataque TOCTOU se demostró solo con un doble de la carga, no con Xen real. n=1, sin SHA-NI. La cadena completa con Gate se midió en lab70 (+3,21 ms).

lab68 · bare metal  PASS

lab68-e2e-real · evidencia real de A hasta la celda B

10 de 10 cadenas A, Gate, orden, lanzador directo y B en metal (2 tandas de 5), oráculo PASS en cada una; en lógica 128 comprobaciones del lanzador y 54 de preflight sin fallos.

límite  n=1. dom0 sigue siendo base de confianza del tap, y el FETCH de B es un eco firmado, no un FETCH real al broker. Sin mTLS en ese canal. El primer humo en metal fue UNCLEAR y se archivó.

lab70 · bare metal  REFUTED

lab70-gate-fast-fd · costo del Gate rápido sobre lanzar sin Gate

La hipótesis H1 (menos de 1,5 ms de pared) se refuta: el brazo completo añade +3,210 ms de mediana (IC95 3,074 a 3,378) sobre la celda sin Gate; repetido con governor performance: +3,216 ms.

límite  n=1, Xen, P=1. El brazo con descriptor incluye un fsync de nonce durable que los otros no pagan.

lab71 · bare metal  PASS

lab71-gate-evidence · activación única y orden de liberación

1000 de 1000 rondas con exactamente una activación (la línea base sin el predicado cruza 300 de 300); órdenes sin firma, ajenas o repetidas dan 0 creaciones; 10 de 10 ciclos con la liberación confirmada antes del veredicto.

límite  n=1, N hasta 32, sin corte de energía a mitad de escritura. En una corrida un brazo dio FAIL de entorno y una sonda se corrigió tras tres defectos de instrumento: se reportan como están.

lab72 · bare metal  PASS

lab72-b-evidence · evidencia de la Solicitud B

Tras destruir, la enumeración de residuos dio 0 en 100 de 100 celdas, con la destrucción comprobada en 0,093 ms como máximo; libro persistente con kill -9 y reinicio sin violaciones; crear el dominio Xen por lanzador directo 41 ms frente a 115 ms con xl. Incluye un modelo TLA+ con conformidad de trazas.

límite  n=1. fs-verity y dm-verity quedaron BLOCKED en Xen y solo se probaron con celda KVM en otro equipo. Los 41 ms son creación de dominio Xen, no latencia de celda.

kz-f1 · f0 · bare metal  PARTIAL

KZ-F1 fase 0: energía, política CPUID y páginas en cero

C1, lectura de energía desde dom0: FAIL según el criterio pre-registrado; la vía de registros MSR devuelve error de E/S bajo Xen 4.20. Una interfaz alternativa respondió, pero no está validada (n=1, sin medidor externo; dos comprobaciones registradas sobre ella también incumplieron su criterio) y solo afecta a la medición de energía. C2, celda sin política CPUID: revisar (80 de 80 pares con comportamiento idéntico, el volcado difiere). C3, páginas en cero tras reutilizar memoria: PASS (210 dominios de 256 MiB, al menos 51.636 MiB reutilizadas, 0 bytes distintos de cero).

límite  ZBook, n=1, una corrida por comprobación. Lo de C1 no es una cifra de energía citable.

kz-f1 · f1 · bare metal  PASS

KZ-F1 fase 1: ciclo firmado, tiempo de armado y xenstore

Ciclo completo con orden firmada: humo PASS (8 de 8 órdenes válidas; 4 de 4 inválidas rechazadas sin crear dominio), repetido con el frente final: PASS. H2, tiempo de armado frente a kz-c2: DENTRO (diferencia pareada de -2,1 a -3,9 ms; criterio ≤ +1,0 ms), repetida con el frente final: DENTRO. H3, nodos y escrituras de xenstore por celda: PASS dos veces (0 nodos y 0 escrituras en 200 celdas). H7a: PASS. Humo del barrido (el trabajador muere con la celda viva): PASS, 8 de 8.

límite  ZBook, n=1: el humo, H2 y H3 corrieron dos veces en el mismo equipo; H7a y el barrido, una. El tiempo de armado no es el de punta a punta: de punta a punta esta fase fue más lenta que kz-c2 (+7,9 a +16,2 ms).

kz-f1 · f2 · bare metal  PARTIAL

KZ-F1 fase 2: constructor residente, recuperación y escala

Ciclo de 15,1 a 18,1 ms frente a 22,7 a 26,9 ms del brazo de referencia, con el Gate incluido. H1 DENTRO: mejora pareada de -8,672 a -7,243 ms con 200 pares por variante (criterio ≤ -4,0 ms). H6 PASS: 13 estados × 100 y 200 fines limpios, 0 violaciones. Humo 8 de 8 con 3 negativos rechazados; barrido 21 de 21; fallos reales 20 de 20 en las cuatro celdas. No cumplidas: H1b INCONCLUSA (1,022 a 1,059 ms contra umbral ≤ 1,0); H4 ENTRE UMBRALES (entrada 3,443 a 3,647 ms y 224 a 280 celdas/s; criterio ≤ 3 ms y ≥ 243/s); H5 ENTRE UMBRALES (1.800 celdas por variante, sin muro, decil mínimo 74,1/s contra criterio ≥ 78/s); H7b REFUTADA (7.840 líneas de código contra criterio ≤ 1.800).

límite  ZBook, n=1, una sola corrida por prueba, sin sellar. Las definiciones de ciclo difieren de las de KVM y de las del rival (Xen contra KVM, con y sin el Gate de firma) y no se encadenan. «Entre umbrales» e «inconclusa» no son pases.

energía · bare metal  PARTIAL

Energía por celda con contadores RAPL, kernel nativo sin hipervisor

Misma tarea de unos 100 ms, energía de procesador y memoria por celda, con el reposo restado: trabajo nativo 0,74 J, nuestra celda 0,89 J, Firecracker con nuestra misma celda 0,98 J, Firecracker con Linux 2,04 J. Sobrecosto sobre el trabajo nativo: 0,15 J contra 1,30 J. Es menos de la mitad de la energía de procesador que una microVM Firecracker con Linux, para el mismo trabajo. Frente a Firecracker con la misma celda casi empata (0,89 contra 0,98 J), y con un proceso nuevo por celda, 0,99 J, empata. Ciclo de vida completo de una celda (arranca, pasa su prueba y se apaga), sin trabajo: mínima 24 mJ, con red 28, con archivos 32, con red y archivos 38. 5 repeticiones de 400 celdas por brazo, 0 fallos. Coherente con kz-f1 f0: allí la lectura de energía se hacía desde dom0 bajo Xen y falló; aquí el kernel es nativo, sin hipervisor (KVM).

límite  ZBook (una laptop), kernel nativo, carga sintética (un hash en lazo). Es energía del chip (RAPL: procesador y memoria), no del enchufe ni de la cuenta de luz; no cubre red, disco, ventiladores ni fuente. El ahorro frente a Firecracker con Linux viene de reemplazar Linux por la celda, no de la celda frente a Firecracker. Sin cifras en dinero. Falta repetir en un servidor y validar con un medidor externo. Evidencia registrada con hash y sello de tiempo.

kz-c1 · bare metal  PASS

kz-c1, lanzador de celdas con libxl directo

Ciclo de unos 38 ms (create 26 + vida 2 + destroy 9) frente a unos 94 ms más harness con xl; 64/64 celdas ok en cada configuración y techo de 50 a 60 celdas/s por nodo desde P=4, porque la creación se serializa en el toolstack.

límite  Medido en el ZBook, anterior a las generaciones posteriores: no es el ciclo vigente ni un titular de latencia. Sin intervalos de confianza.

Etapa 6 · 2 labs

La nube, en su lugar.

Las corridas anidadas orientan y nada más: los veredictos de rendimiento y aislamiento son los medidos en metal.

kz-gc1 · nube (anidada)  PARTIAL

kz-gc1: la celda KVM en una VM de nube (KVM anidado)

Celda lista desde cero en 6,9 ms frente a 107 ms de Firecracker + Linux sin ajustar. En trabajos de ~100 ms: 2,11x en el peor P, frente a Linux sin ajustar la ventaja se ve; con la misma celda en Firecracker queda en 1,04x (orienta). Ráfaga de 20.000 celdas sin fallos y 13.053 celdas vivas a 1,86 MB cada una.

límite  Exploratorio · en la nube (anidado) · orienta, no es veredicto; los veredictos de rendimiento y aislamiento son los del ZBook. Una ronda por punto; las 13.053 se cortaron por tiempo, no es techo de RAM.

kz-az1 · bare metal + nube anidada  PARTIAL

kz-az1: duelo en metal y corridas en AWS anidado

En metal, la hipótesis de 2x en trabajos de ~100 ms se sostiene solo contra Linux sin ajustar (2,01x, al borde) y se falsa contra Linux ajustado (0,92x a 1,03x, empate técnico). Con la mejor configuración medida del recorte, kz-k2 lista 1,51 contra 2,32 ms de Unikraft y 1,03 contra 1,49 MiB por celda; antes del recorte Unikraft llegaba antes (2,25 contra 4,83 ms).

límite  Una máquina, una ronda por punto. A P=15 Unikraft queda arriba en tasa en 3 de 4 celdas; con los defaults Unikraft va adelante en lista. AWS anidado orienta, no es veredicto. Snapshot-restore queda fuera de los rankings. El mecanismo del recorte no se publica.

Etapa 7 · 11 labs

El código de referencia.

Los componentes mismos, como scripts autónomos que imprimen sus checks y salen con 0 solo si cada vector coincidió con su veredicto esperado. Casi todos en plano lógica.

impl · gate · lógica  PASS

Gate (compuerta de admisión del Job)

42/42 checks sin fallos (rc=0); reserva sin fallos y libro indexado 8/8. En los 31 vectores adversariales ampliados (gate, broker, chain): 19 DENY, 9 GAP declarados, 3 LIMIT por diseño.

límite  Plano lógico. Los 9 GAP se publican sin esconderlos; C5 fue el único de severidad alta. La corrida en metal del Gate es lab71.

impl · broker · lógica  PASS

Broker (E/S mediada con CID)

71/71 checks (68 vectores, pureza, derivación y firmas), 0 fallos, rc=0. Cubre el pin de clave del par: otro certificado de la misma CA con otra clave queda negado.

límite  Plano lógico. Quedan GAP declarados (B2, B5, B6 de diseño y B7, rollback de estado, que depende de la durabilidad del dom0). No es veredicto de aislamiento.

impl · chain · lógica  PASS

Chain (cadena de recibos y linaje)

37/37 escenarios sin fallos, rc=0. La cobertura de firma pasó a atestiguar el conjunto completo de campos (mover cualquiera da BAD_SIGNATURE) y el linaje exige un recibo de destrucción real.

límite  Plano lógico. Hallazgos de partida que se publican: el linaje era vacuo por una constante pública y la cobertura de firma era parcial; ambos se corrigieron en el código real.

impl · observer · lógica  PASS

Observer (contrafirma y recibo de destrucción)

14/14 checks sin fallos y cadena de recibos de destrucción con rc=0. Quitar o sustituir la contrafirma (6+3 variantes) es negado por MISSING_COSIGNATURE.

límite  Plano lógico. El observer contrafirmaba sin un hecho de commit observado (C5); se trató en el código real con un almacén separado. Aquí no se afirma independencia en producción.

impl · ring · lógica  PASS

Ring (framing del canal de petición y respuesta)

62/62 vectores sin fallos, rc=0. El receptor valida solo un snapshot de la página, y los artefactos grandes viajan por chunks.

límite  Plano lógico (loopback en un solo equipo, sin canal Xen real). El fence de Python es de mejor esfuerzo. No es veredicto de aislamiento ni de rendimiento.

impl · transport · lógica  PARTIAL

Transport (mTLS del plano de control)

En lógica: mtls 22/22, serve 21/21, integración 13/13 (handshake real, 0 SKIP), tier2 6/6, ca 18/18, identity 8/8. En metal, una vez: 12/12 suites con handshake TLS 1.3 real. Otro certificado de la misma CA con otra clave queda negado.

límite  Abierto: correrlo entre principales de SO distintos y entre dominios Xen separados. En metal ambos extremos tenían el mismo UID; el canal actual entre dominios es un relay sin mTLS.

impl · crypto · lógica  PASS

kzsig: firma asimétrica Ed25519

15/15 checks, 0 fallos: con solo claves públicas el verificador comprueba pero no forja; 300 firmas aleatorias, firma reutilizada y firma con la pública fallan; un dom0 que firma con su propia clave es rechazado.

límite  Lógica pura. Sigue abierto distribuir de forma auténtica las tres claves públicas; dom0 aún puede descartar o retrasar; las privadas no están protegidas por esta capa.

impl · kaf-store · lógica  PASS

kaf-store: almacenamiento aislado por proceso (cierre de C5)

22/22 checks (rc=0): pedir firma de bytes que nunca se almacenaron, almacenar basura y declarar otro hash, o escribir directo en el medio, es negado; ninguna privada alcanzable en el proceso del broker.

límite  Cierra solo el caso de un proceso broker comprometido. Contra dom0 con root el aislamiento de proceso no basta: es residual de hardware. No se probó en metal en este directorio.

impl · alicebob · lógica  PARTIAL

alicebob: notaría y registro de claves para agentes de consola

Dos revisiones adversariales hallaron y corrigieron una veintena de fallos (cadena reescribible o truncable, dueño por Job, inyección en la consulta de auditoría, entre otros); los tests Go de cadena, firma, servidor y almacén pasan.

límite  Hallazgo NO corregido con código: si daemon y agente corren con el mismo UID no hay frontera agente/humano. Bitácora lineal, no Merkle; sin cifra de rendimiento.

impl · rump-xen · bare metal  PASS

Celda KZ-Rump sobre Xen PVH (y misma imagen en KVM)

Secuencia honesta: T15a enlace PASS; T15b FALLÓ la consola; T15c KZ-RUMP-OK; T16a solo-fs BLOQUEADA (UNCLEAR); T16b solo-fs PASS 3/3; T16c solo-red PASS, contención de 8 hilos sin abortos y un deadlock deliberado que muere observable a los 30 s. Un solo ELF arranca en Xen y en KVM.

límite  n=1, una corrida por prueba. Una primera auditoría halló un watchdog insuficiente y se corrigió. Las cifras de arranque vigentes son las del recorte, no las de estos informes.

mtls-metal · bare metal  PARTIAL

mtls-metal: mTLS del plano de control en el dom0 físico

12/12 suites verdes en metal con handshake TLS 1.3 real (un certificado de la misma CA con otra clave queda negado). El arnés entre principales de UID distinto nunca corrió.

límite  Ambos extremos con el mismo UID y TLS en puro Python: no es separación entre principales de SO ni entre dominios Xen. n=1.

Memoria

La memoria, en toda nube.

Quien controla el host puede leer la memoria de lo que corre en él: el proveedor de nube, o un atacante que consiga root en el host. Eso lo tiene toda nube (AWS, GCP, cualquiera) si no usas computación confidencial. No siempre se dice con estas palabras, pero es así. Lo decimos claro para que nadie piense que es un problema de Kawésqar.

memoria
Nube sin computaciónconfidencialAWS, GCP, cualquieratu VM / tu cargamemoria en clarosecretohipervisor / hostrootleeEl proveedor, o un atacante conroot en el host, lee la memoria.Servidor propio,dom0 de Xenmedido en metal, n=1la celdamemoria en clarosecretodom0rootleeRoot en dom0 lee la memoria de lacelda. Medido: 3 de 3, un secretonacido dentro de la celda.ComputaciónconfidencialTDX · SEV-SNPtu VM / tu cargamemoria cifrada▒▓░ x9#f 7a▒ ░▓hipervisor / hostrootbloqueadaEl host ve solo texto cifrado.La lectura queda bloqueada.Mismo modelo de seguridad en las tres;lo único que cambia es el cifrado de memoria.Nube sin computaciónconfidencialAWS, GCP, cualquieratu VM / tu cargamemoria en clarosecretohipervisor / hostrootleeEl proveedor, o un atacante conroot en el host, lee la memoria.Servidor propio,dom0 de Xenmedido en metal, n=1la celdamemoria en clarosecretodom0rootleeRoot en dom0 lee la memoria de lacelda. Medido: 3 de 3, un secretonacido dentro de la celda.ComputaciónconfidencialTDX · SEV-SNPtu VM / tu cargamemoria cifrada▒▓░ x9#f 7a▒ ░▓hipervisor / hostrootbloqueadaEl host ve solo texto cifrado.La lectura queda bloqueada.Mismo modelo de seguridad en las tres;lo único que cambia es el cifrado de memoria.
memoria   Quién puede leer qué. Flecha ember: lectura desde el host. Solo la tercera pila, con cifrado por hardware, la frena.
plataforma

Es de la plataforma, no de Kawésqar

Cualquier carga en un host sin computación confidencial es legible desde el host. Es una propiedad de la plataforma, igual para todo producto que corre encima.

lo probamos

Lo probamos en nube confidencial

TDX en GCP y SEV-SNP en AWS. La memoria queda cifrada, pero la celda no puede correr adentro (no hay KVM) y una VM confidencial por trabajo tarda de 50 a 58 s en levantar.

lo que pedimos

Por eso se valida en hardware físico

El cierre en la celda se valida en hardware físico (TDX o SEV-SNP), que es lo que pedimos.

Cómo lo usamos

Lo crítico se queda en casa, el resto va rápido.

No proponemos volver atrás frente a la tendencia de moverse a la nube. Usamos la velocidad de Kawésqar en la nube para las cargas menos críticas, o donde se puede tolerar el riesgo, y dejamos en casa solo lo que lo merece: soberano, on-prem o local. Lo más crítico, por ejemplo las llaves de un banco, se queda en hardware propio con cifrado de memoria completo (TDX / SEV-SNP).

escala
Carga toleranteal riesgo→ Nubevelocidad KawésqarCargasensible→ On-premsoberano, en casaLlaves de un banco→ On-prem con cifrado de memoria completoTDX / SEV-SNPse valida en hardware físicomenos críticomás críticoCarga toleranteal riesgo→ Nubevelocidad KawésqarCargasensible→ On-premsoberano, en casaLlaves de un banco→ On-prem con cifrado de memoria completoTDX / SEV-SNPse valida en hardware físicomenos críticomás crítico
escala   Dónde va cada tipo de carga. Solo se deja en casa lo que lo merece.
mismo contrato

Mismo contrato, mismo Gate

A los dos lados rigen el mismo contrato, el mismo Gate y el mismo modelo de seguridad. Solo cambia dónde corre la carga.

medido

Lo que está medido hoy

El cierre con cifrado de memoria completo en la celda se valida en hardware físico (TDX o SEV-SNP), que es lo que pedimos a Intel y AMD. Hoy el residual está medido y declarado (lab54b, 3 de 3), no cerrado.

aún no

Lo que todavía no existe

La federación entre on-prem y nube es una dirección de diseño, todavía sin código.

Sigue leyendo

Evidencia, luego lo que pedimos.

La página de evidencia pone las cifras principales en contexto y nombra las brechas que siguen abiertas. Si quieres reproducir algo de aquí, cuéntanos con claridad.

la regla

No es un certificado ni una afirmación de cumplimiento regulatorio: es un registro fechado y acotado de lo que corrió.

Agent Fabric · caderno de laboratório

Mais de 70 labs. Os que falharam também.

Este é o caderno de laboratório do Agent Fabric: cada experimento que podemos publicar, com o plano em que rodou, o veredito, o número e o limite. As falhas e as refutações estão ao lado dos acertos, porque um laboratório que só mostra o que funcionou mostra um pitch, não evidência.

0+
experimentos no laboratório, e contando
0
mostrados aqui, cada um com seu veredito, plano e limite
0
testados: PASS, HELD ou mitigados por política
0
publicados como FAIL ou REFUTED, não enterrados

patent pending  Pedidos de patente apresentados ao INAPI (Chile); em andamento.

Quadro de vereditos

Cada lab, por plano.

Uma ficha por lab. Verde é um acerto testado, âmbar é parcial, pouco claro, aberto ou um residual medido, laranja é uma falha ou refutação registrada. O plano importa tanto quanto a cor: nossos resultados rodam aninhados (dentro de uma VM, o que não valida isolamento nem desempenho) ou em bare metal, o único lugar onde essas afirmações podem se fechar. Metal é uma estação portátil, n=1, salvo quando o cartão diz outra coisa.

VM aninhada · 16 labs

12 testados · 2 parciais ou abertos · 2 falhos

Nuvem aninhada · 1 lab

0 testados · 1 parciais ou abertos · 0 falhos

O lab72 inclui também um modelo TLA+ confrontado com traces, contado aqui em bare metal. Na nuvem mostramos só o que orienta; ela nunca fecha um veredito.

O que falhou

O que falhou, dito com clareza.

Um laboratório que só publica seus acertos está vendendo. Estes são os pontos em que os dados disseram não, ou disseram menos do que esperávamos, e o que mudamos depois.

VM aninhada  FAIL

lab14 · O guest rump não imprimiu nada

FAIL em Xen aninhado. O guest chegou ao modo longo e executou rump_init, mas duas falhas de console impediram que a marca aparecesse. A falha ditou a ordem das correções que o lab16 fechou.

VM aninhada  FAIL

lab15 · O observador não falhou fechado

F9 FAIL: 38.577 mutações produziram 4.520 crashes, 85 falsas contraassinaturas (só com o stream comprometido) e 56 travamentos. O lab29 reproduziu depois a falha em metal, endureceu o observador e chegou a 0 crashes em 3.000 variantes.

bare metal  FAIL

lab25 · A célula descartável vazou em disco

A RAM resistiu (0 hits), mas um volume reutilizado deixou a célula B ler o padrão de A: 6.882.854 hits. Zerar o volume inteiro ao criar deixou em 0, uma correção verificada no lab antes de chegar ao produto.

bare metal  REFUTED

lab64 · Gate mais ordem custaram mais de 1 ms

Pré-registramos menos de 1 ms. A mediana medida é 2,735 ms (P=1, n=200). A hipótese é refutada e esse número é o que publicamos.

bare metal  REFUTED

lab70 · O Gate mais rápido também não atingiu a meta

H1 pedia menos de 1,5 ms de parede. O braço completo acrescenta +3,210 ms (IC95 3,074 a 3,378), e +3,216 ms de novo com o governor performance. Refutada duas vezes.

bare metal  FAIL

KZ-F1 fase 0 · C1

FAIL segundo o critério pré-registrado: a via de registradores devolve erro de E/S sob Xen 4.20, antes e depois da carga. Uma interface alternativa respondeu e acompanhou a carga, mas não está validada (n=1, sem medidor externo) e duas verificações registradas sobre ela não cumpriram o critério. Afeta apenas a medição de energia, e nenhum número de energia é citado.

bare metal  REFUTED

KZ-F1 fase 2 · H7b

Pré-registramos no máximo 1.800 linhas de código para o caminho de produção. A contagem é 7.840, e acima de 2.000 se dá por refutada. A hipótese está refutada e o número é publicado como foi medido, sem redefinir a contagem.

bare metal  PARTIAL

lab20a · Uma observação que não devemos inflar

Viu-se uma negação de DMA uma vez e não foi possível reproduzi-la; no modo strict o próprio controle positivo foi negado. É uma observação, não um resultado, e não a citamos como «strict nega transferências não mapeadas».

lógica  PARTIAL

C5 · O fechamento não foi ganho

Um broker que forja a confirmação de armazenamento cruzou a porta de admissão na primeira onda lógica (23/23 comportamentos reproduzidos). Um trabalho posterior acrescentou um armazenamento separado; o cartão do kaf-store mostra o que isso fecha e o que não fecha.

bare metal  TIE

lab27 · «Rump expõe menos» era falso

A hipótese foi falseada: rump mantém 2 event channels vivos contra 0 do Linux mínimo, e empatam em grants e nós do xenstore. Concluímos que o rump não justificava, por si só, um investimento profundo.

bare metal + nuvem aninhada  PARTIAL

kz-az1 · O 2x só se sustenta pela metade

Em trabalhos de 100 ms a vantagem de 2x se sustenta contra Linux sem ajuste (2,01x, no limite) e se falseia contra Linux ajustado (0,92x a 1,03x). Antes do recorte, o Unikraft ficava pronto antes. Começamos pelo que não se sustentou.

Linha do tempo

Seis etapas, um fio.

O caderno segue a ordem em que trabalhamos: primeiro o sandbox, depois o Xen aninhado numa VM, depois a máquina física, depois os ataques ao nosso próprio projeto, em seguida as células rápidas e por fim a nuvem. A altura de cada barra é a quantidade de labs mostrados naquela etapa; o código de referência (mais 11 cartões) fica à parte, no final.

linha do tempo
2SANDBOX29 set14XEN ANINHADO29 set – 10 out27BARE METAL16 – 18 out16RED TEAM18 – 24 out15CÉLULAS KZ2NUVEMSANDBOX29 set2XEN ANINHADO29 set – 10 out14BARE METAL16 – 18 out27RED TEAM18 – 24 out16CÉLULAS KZ15NUVEM2
linha do tempo   Labs mostrados por etapa. Um resultado aninhado ou de nuvem nunca conta como veredito de isolamento nem de desempenho: só o metal pode fechá-los.
Etapa 1 · 16 labs

Xen aninhado, as primeiras respostas.

Dentro de uma VM (Xen sob QEMU/KVM). Estas execuções mostram a arquitetura funcionando como lógica sobre domínios reais. Nunca são um veredito de isolamento nem de desempenho, e cada cartão diz isso.

Cada cartão: lab, plano e veredito, depois o que se encontrou e depois o limite que o acompanha. Os números são os que o lab registrou; quando uma campanha posterior os substituiu, o cartão diz isso.

lab03 · VM aninhada  PARTIAL

Rump kernels NetBSD: fechamento das incógnitas

A pilha IP real usa 52 hypercalls (base 46) e 28 de 46 são invocados na inicialização; na ordem oficial a definição real vence em 3 de 3 duplicados; o aviso sysctl é benigno.

limite  Um único guest aninhado; build reproduzível parcial (186/190 idênticas); sem conclusão de desempenho nem de isolamento.

lab04 · VM aninhada  PASS

Não determinismo do build, disco do guest, testes do skill

A diferença eram só caminhos de fonte embutidos: dois builds MKREPRO saem idênticos, 190/190. Corrige a hipótese do lab03; testes do skill 12/12 após reexecutar um.

limite  n=2 builds por condição; a causa exata é inferência forte, não fato; um caso falhou na primeira passada.

lab05 · VM aninhada  PASS

Xen aninhado e célula PVH mínima

A célula imprime sua marca e termina sozinha; 5 ciclos e 1 destroy durante o boot deixam 0 domínios e 0 chaves residuais no xenstore.

limite  Tudo aninhado e sem IOMMU: não vale para desempenho nem isolamento; não é benchmark.

lab06 · VM aninhada  PARTIAL

Validação de skills e célula Xen aninhada (A a D)

A: 10/12 PASS; B: 7 PASS e 1 FAIL (um caso inventou um formato); C: 5/5, 30 ciclos com 0 resíduos e 0 falhas; D: 5/5. A contagem de hypercalls rump é 46 nos dois lados.

limite  Aninhado, sem IOMMU; três casos não podem ser reproduzidos porque o prompt e o critério não estão em disco.

lab07 · VM aninhada  PASS

S4.0 sonda xenstore entre duas células

Sem ACL a leitura cruzada dá EACCES (30/30 no controle); a célula dona concede o acesso e a outra lê o nonce e recebe 3 eventos. O dom0 não é requisito.

limite  Aninhado, N=1 por braço, uma única variante do xenstored; sem cotas nem concorrência.

lab07b · VM aninhada  PASS

S4.2 anel gntalloc+gntdev+evtchn entre duas células

Ring bidirecional com hashes idênticos nas duas pontas, sem privilégio de dom0. A primeira execução FALHOU por um protocolo próprio defeituoso e foi corrigida.

limite  Aninhado, n=1 PASS; N≥30 pendente; não se provou que um terceiro domínio não possa mapear o grant.

lab08 · VM aninhada  PASS

Domínios placeholder de controle, broker e observer

30 ciclos (90 domínios): 0 resíduos no XenStore e 90/90 com código 0; uma Cell de teste teve acesso negado em 15/15 operações nos ramos dos três papéis.

limite  Aninhado, uma execução, domínios placeholder sem lógica real; não é veredicto de isolamento.

lab09 · VM aninhada  PASS

Lógica dos papéis dentro de um initramfs

As três suítes passam dentro de um domínio PVH de 256 MB (17/17, 31/31, 9/9) com ~167-172 MB livres e 0 sinais de falta de memória.

limite  Aninhado, n=1; apenas viabilidade funcional e folga de memória.

lab10 · VM aninhada  PASS

Broker em um domínio e Cell pelo ring

5/5: artefato de 9000 B com hash idêntico nos dois sentidos; CID alheio ou forjado, repetição e destino forçado negados; 0 processos do par no dom0.

limite  Aninhado, n=1; sem conclusão de desempenho nem de resistência no Xen em metal.

lab11 · VM aninhada  PASS

Observer que contraassina e emite recibo de destruição

4/4: contraassinatura só com a Cell viva durante um STORE real; recibo de destruição só depois de destruí-la; 17/17 tentativas de ler, alterar ou apagar o fluxo negadas.

limite  Aninhado, n=1; não cobre uma Cell hostil com capacidades de kernel.

lab12 · VM aninhada  PASS

Cadeia A, recibo, B sobre domínios reais

As 4 condições valem: B não existe antes da ordem, sem contraassinatura o Gate nega, uma Cell que tenta assinar a ordem tem acesso negado (18 vezes) e o hash é idêntico do início ao fim.

limite  PASS restrito à arquitetura: aninhado, n=1, assinatura simétrica da primeira versão; a primeira tentativa falhou por empacotamento.

lab13 · VM aninhada  PASS

B1.8 Falsificações que só o substrato pode refutar

As 3 falsificações dão PASS: CIDs do epoch anterior dão STALE_EPOCH após recriar o broker; um terceiro domínio não mapeia o grant (128 refs negadas); um kill entre STAGED e COMMITTED dá NOT_COMMITTED. Achado: epoch e nonces devem ser duráveis fora do domínio.

limite  Aninhado, n=1, assinatura HMAC v0; o instante do kill é forçado; o dom0 continua na base de confiança.

lab14 · VM aninhada  FAIL

T15b Primeiro boot do rump como guest PVH sob Xen

FAIL instrutivo: o guest entrou por PVH, passou ao modo longo e executou rump_init, mas não imprimiu a marca. Duas falhas de console (hypercalls sem programar e um buffer eliminado pelo compilador) ditaram as correções que o lab16 fechou.

limite  Aninhado, n=1; ficou sem observar se rump_init retornou 0.

lab15 · VM aninhada  FAIL

S9.2 Falsificações 8 e 9 do observador

F8 PASS: o observador sobrevive à queda de uma célula e o Gate não usa sua autodeclaração. F9 FAIL: com 38.577 mutações houve 4.520 crashes, 85 falsas contraassinaturas (só com o stream comprometido) e 56 travamentos. Um crash não assina: é falha de disponibilidade.

limite  Aninhado, n=1, HMAC v0; a mentira é um roteiro fixo; o dom0 continua na base de confiança.

lab16 · VM aninhada  PASS

T15c Fechamento do eixo rump: o guest rump imprime KZ-RUMP-OK sob Xen

Com as duas correções de console, o guest rump imprime a marca, rump_init retorna 0 e desliga por hypercall com código 0. Fecha o bit que o lab14 não conseguiu ler.

limite  Aninhado, n=1: só init e console; sem desempenho nem isolamento forte.

lab17 · VM aninhada  PASS

T16 Guest adversarial: A_effective contido em A_delegated

289 tentativas de mapear grants alheios e 289 de ligar canais de evento deram 0 sucessos, 0 páginas alheias lidas, 0 autoconcessão. Separa-se o que o Xen contém do que o Kawésqar contém (replay, STALE_EPOCH, NOT_COMMITTED).

limite  Aninhado, n=1; a garantia supõe um Xen correto; as negações de grants e canais são do Xen, não do Kawésqar.

Etapa 2 · 12 labs

A primeira campanha em metal.

No domínio de controle físico de uma estação portátil, n=1. Aqui aparecem as primeiras falhas honestas: um canal encoberto que não conseguimos medir e um disco que vazava.

lab18 · bare metal  PASS

Baseline físico e gate do plano de laboratório

Inventário capturado e preflight 11/11. As CPUs deste Xeon não são afetadas por L1TF, MDS nem TAA, então o lab de canal lateral foi redelimitado.

limite  ZBook, n=1, um único dia; é inventário e preparação, não um resultado de isolamento nem de desempenho.

lab19 · bare metal  PASS

Reatestação em metal do núcleo da arquitetura

7/7 veredictos do núcleo PASS no domínio de controle físico. Ciclo de uma célula mínima: p50 0,95 s, p99 1,09 s, 200/200 sem falha, resíduo real 0.

limite  ZBook, n=1, domínio de controle limitado a 4 vCPU: correção, não desempenho; a cadeia sobre domínios reais continua em HMAC.

lab20a · bare metal  PARTIAL

Aplicação do VT-d em bare metal com um dispositivo DMA nomeado

A configuração do VT-d foi atestada e uma negação foi vista uma vez, mas não é reproduzível. No modo strict o instrumento ficou UNCLEAR: o controle positivo foi negado.

limite  ZBook, n=1. Não deve ser citado como «strict nega transferências não mapeadas»: é uma observação, não um resultado testado.

lab21 · bare metal  PASS

Latência base de uma célula em bare metal

200/200 por variante, 0 resíduos: criar com xl 184 ms (p50) e iniciar kernel e init 607 ms (p50, p99 790 ms). Uma célula com trabalho acrescenta um overhead limitado.

limite  Números da geração inicial (kernel Linux e xl), muito anteriores às células atuais: não os compare com elas.

lab22 · bare metal  PASS

Escala e carga

0 falhas do Gate em 256 execuções sob carga; 2.775 ciclos concluídos a 100 %; criação estável em torno de 237 ms de N=1 a N=414; 0 células residuais.

limite  É a lógica do Gate sob pressão de scheduling, não transições reais de célula em escala; o teto de recursos nunca foi estressado.

lab23 · bare metal  UNCLEAR

Canal encoberto por contenção

O controle positivo não cruzou o limiar: melhor caso 9,64 bps com IC 95 % [-33,06 ; 52,31], indistinguível do ruído. Não se afirma nem presença nem ausência de canal.

limite  Instrumento não validado; registrado como canal não medido, nunca como desprezível.

lab23b · bare metal  MITIGATED

Canal encoberto: fechamento por política

Risco fechado por política de posicionamento (não co-agendar células que não confiam entre si em threads irmãs), não por medição: o canal SMT mais forte não existe por construção.

limite  Não é «provamos que não há canal»: os canais por cache compartilhado e memória entre núcleos ficam como risco residual.

lab24 · bare metal  OPEN

XSAs que realmente se aplicam

Três avisos corrigidos, os especulativos inaplicáveis neste silício ou mitigados; resta um aplicável sem correção (um DoS de liveness no daemon do xenstore). Nenhum é escalada nem vazamento entre domínios.

limite  Estado preso a uma versão de pacote; é preciso reexecutar ao trocar de build.

lab25 · bare metal  FAIL

Célula descartável: resíduo de RAM e disco

RAM PASS: 0 hits com sobreposição forçada de ≥548 MiB. Disco FAIL: a célula B vê o padrão de A (6.882.854 hits); zerar o volume inteiro ao criar deixa em 0 hits.

limite  A mitigação foi verificada no lab mas ainda não estava no produto; resíduo por DMA de dispositivo fora de escopo.

lab26 · bare metal  PARTIAL

Fechamento da campanha bare metal

Arquitetura: PASS restrito, fechamento condicional. Sem amplificação de autoridade nem de evidência em Xen real (7/7) e sob carga; célula descartável falseada em disco e PASS em RAM.

limite  Não se afirma isolamento físico verificado, nem célula descartável incondicional, nem a tese completa; canal encoberto UNCLEAR.

lab27 · bare metal  TIE

Shootout de fronteira: KZ-Rump contra Linux mínimo

A hipótese «rump expõe menos» foi falseada: rump mantém 2 event channels vivos (Linux mínimo 0); empatam em grants (0) e em 38 nós do xenstore. Achado: um aborto do adaptador dava reason code 0, igual a um desligamento limpo.

limite  Superfície observada, não alcançável; ambos os runtimes são quase stubs; uma máquina, uma execução.

lab28 · lógica + bare metal  PASS

Árvore de delegação n>1 e não amplificação por indução

Gate real: 1.958 arestas decididas, 0 amplificações; células reais na borda exata do envelope (768 de 768 MiB); um neto que pede mais que o pai é negado com AUTHORITY_NOT_SUBSET; 5 mutantes detectados.

limite  A árvore de parentesco é modelada (o Xen não a conhece); profundidade maior que 2, fan-in e fuzz são lógica. O lab achou defeitos reais do Gate, já corrigidos.

Etapa 3 · 15 labs

Levando o metal ao limite.

Ainda em metal: fuzzing do observador, a cadeia assinada dentro de células reais, saturação até 412 células vivas, IOMMU estrito e hypercalls observados.

lab29 · bare metal  PASS

Falsificação 9: superfície do observador

A primeira execução FALSEOU a propriedade: 47 crashes e 208 violações por brechas alcançáveis via JSON. Após endurecer o observador: 0 crashes e 0 violações em 3.000 variantes, 3 mutantes detectados, 0 resíduos.

limite  PASS restrito ao fio (stream JSON decodificado); restam violações da família inproc, declaradas como GAP. n=1.

lab30 · bare metal  PASS

Falsificação 10: STORE transacional sob kill

Em nenhuma etapa restou um COMMITTED parcial entregável (7 kill -9 reais); o replay após restart deu STALE_EPOCH sempre; 3 mutantes detectados; células reais destruídas.

limite  A persistência do journal é modelada pelo harness: prova a atomicidade do contrato, não o CAS persistente de produção (GAP declarado).

lab31 · bare metal  PASS

Gate adversarial sob carga

0 aceitações em 6.870 tentativas adversariais concorrentes; 1.194 de 1.194 legítimas cruzam sob a mesma carga; a corrida ingênua superreserva 6 de 6 (o instrumento vê a corrida); 0 resíduos.

limite  O lock do ledger é em processo; o CAS persistente de produção fica de fora. Corpus finito, não prova de ausência geral.

lab32 · bare metal  PASS

Soak de estabilidade do ciclo criar-rodar-destruir

400 de 400 iterações: deriva zero (tau de Kendall 0,02 a 0,08), resíduo 0, memória livre do Xen 55.883 a 55.883 MiB, dom0 estável.

limite  Janela limitada (não meses), uma célula mínima, uma máquina, n=1.

lab33 · bare metal  PASS

Saturação em escala máxima e throughput

Pico de 412 células vivas simultâneas (52.736 de 53.835 MiB), 0 amplificação e 0 resíduo nos 14 níveis; máximo de 14,78 células/s de criação e 8,98 células/s de ciclo completo.

limite  n=1. No pico a destruição chega a p99 32 s pela serialização do toolstack; em escala só uma amostra leva console com marca.

lab34 · lógica + bare metal  PASS

Fan-out de recursos (Falsificação #4)

Em lógica: 4.681 nós e 7.608 arestas com 0 amplificações; em metal, com células de 128 MiB, todos os degraus completam, 0 amplificações e overhead por célula de 1,5 a 2,5 MiB.

limite  As duas primeiras tentativas físicas ficaram UNCLEAR (a causa foi achada no lab40); o fechamento é a terceira. n=1.

lab35 · bare metal  PASS

Fronteira mínima observada da célula sob carga

Classificação completa de 18 interfaces e 49 hypercalls; a carga não rompeu a fronteira, 0 resíduo nos 4 cenários. O isolamento é delimitado pelo Xen, não pelo runtime.

limite  Caracterização, não veredicto de isolamento. A observação por trace ficou UNKNOWN nas 49 linhas por falta de decodificador (o lab44 fechou).

lab38 · bare metal  PASS

Cadeia criptográfica Ed25519 de ponta a ponta dentro de células

A cadeia célula A, contraassinatura, Gate, célula B e recibo de destruição roda em metal; 4 mutantes (forja pelo dom0, recibo alterado, B sem evidência, replay) e 7 vetores negativos rejeitados sem criar B.

limite  A verificação da assinatura antes de armazenar é modelada pelo harness, não pelo Gate de referência. n=1.

lab39 · bare metal  PASS

Fumaça e regressão de invariantes sob IOMMU estrito

Os 4 grupos passam: 60/60 iterações, 0 aceitos de 3.444 tentativas adversariais com 600/600 legítimos admitidos, e zeragem verificada; o create sobe cerca de 12 % (uns 23 ms) sob strict.

limite  Regressão limitada, não reprova exaustiva; o A/B limpo no mesmo boot ficou pendente. n=1.

lab40 · bare metal  PASS

Piso de memória de boot de uma célula PVH

Piso medido de 96 MiB: com 64 MiB a célula não inicia, com 96 e 128 MiB sim. Fecha a causa raiz da falha do lab34 e falseia a hipótese de corrida do harness.

limite  Piso grosso (n=3 por ponto), válido para este kernel e initrd de laboratório; o valor exato entre 65 e 96 MiB não foi refinado.

lab41 · bare metal  PASS

Auditoria de grants do dom0 sobre páginas da célula

Uma célula sem disco, parada ou sob carga, tem exatamente 1 grant ativo mapeado pelo dom0 (o anel de console), 0 páginas de dados; com disco aparecem 67 refs (controle positivo).

limite  A primeira execução foi UNCLEAR por controle positivo cego e foi corrigida. Que entrem no IOMMU do dom0 é inferência da fonte. n=1.

lab42 · bare metal  PASS

Fronteira de entropia dentro da célula

14 de 14 chaves distintas; em nenhum boot se gerou chave com o gerador aleatório sem estar pronto, nem com RDRAND desligado (mediana 1,47 s contra 0,14 s).

limite  Prova disponibilidade e não determinismo, não a qualidade da entropia. O caso sem RDRAND se apoia na leitura da fonte. n=1.

lab43 · bare metal  PASS

Cadeia criptográfica em célula com n maior que 1

A cadeia e o leque cruzam de ponta a ponta e todos os mutantes são rejeitados com sua razão exata; o leque preenche o envelope exato (2 por 128 MiB). Achado: o gate puro não consome o elo prev dos recibos de destruição.

limite  A verificação da assinatura e o elo prev ficam modelados no arnês (lacuna de produção declarada). n=1.

lab44 · bare metal  PASS

Hypercalls observados por trace real

Nos 4 cenários o observado é subconjunto do conjunto declarado de 10 hypercalls; a célula rump deixa uma pegada mínima (5 números distintos) e o Linux acrescenta mais 4. Controle negativo: 0 chamadas com a célula pausada.

limite  Observa-se o número do hypercall, não o sub-op nem os argumentos. n=1.

lab45 · bare metal  PASS

Soak sob IOMMU estrito e vazamento de estruturas

400 de 400 ciclos sem vazamento: grants ativos planos em 2, maptrack plano em 4 e memória livre do Xen constante (55.883 MiB antes e depois).

limite  A memória livre é lida em MiB (mínimo detectável perto de 80 KiB por ciclo); os grants ativos são um proxy parcial. n=1.

Etapa 4 · 16 labs

Atacar e depois verificar fora do dom0.

A decisão sai do domínio de controle, depois atacamos o nosso próprio projeto, medimos o que o dom0 ainda pode ler e comparamos com Linux.

lab46 · lógica  PASS

Conformidade de negação e agente vivo contra o canal real

12 de 12 checks e oráculo independente em PASS, cada negação com sua razão exata; 0 escaladas cruzadas com os 8 modelos de linguagem testados, incluindo os 3 que atacaram pelo canal real.

limite  Nível lógica, não metal nem agente dentro da célula. Uma campanha posterior corrigiu exageros do próprio lab (mutantes que mutavam dados, não código).

lab47 · bare metal  PASS

Verificador fora do dom0 (cadeia A a B)

O controle positivo cruza e os 11 mutantes são negados com sua razão exata; criações = admissões = PASS do verificador (1 = 1 = 1); as forjas e repetições do dom0 morrem pela assinatura. Reatestado em metal: 3 de 3 execuções PASS.

limite  n=1, um host. O dom0 ainda cria domínios e pode ler a RAM de um domínio vivo (residual de hardware). A primeira execução física foi UNCLEAR por um bug do oráculo, corrigido.

lab48 · bare metal  PASS

Gate endurecido (gate_verified) contra o harness

Com o arnês desligado, o gate antigo ACEITA e cria o filho nos 4 mutantes principais; o novo os NEGA com sua razão. As propriedades passam de modeladas no arnês a impostas pelo gate.

limite  n=1. O dom0 ainda gerava as chaves privadas nesta execução (o lab47 ataca isso); um tap mentiroso não fica coberto.

lab49-c5 · lógica + bare metal  PASS

Fechamento do C5: broker comprometido

O observador antigo aceita as duas falsificações; o kaf-store nega as duas e a testemunha de bytes nega a primeira. Em metal, o broker comprometido que assinou sem respaldar os bytes deixa o meio vazio e não cruza.

limite  A testemunha deixa passar um COMMITTED falso sobre bytes honestamente em staging. Isto não encerra o marcador posterior de red team, que registra o fechamento de C5 como não ganho. n=1.

lab50-dma · bare metal  PASS

Contenção DMA de um dispositivo atribuído

O dom0 lê o pendrive antes de atribuir, perde o dispositivo ao atribuir, e o domU lê o mesmo número de setores com 0 faults VT-d em modo estrito. Mede confinamento, não isolamento adversarial.

limite  Parcial: não se testou um driver malicioso tentando DMA para a memória de outro domínio. Rodou em HVM porque PVH não suporta passthrough. Não citar como isolamento DMA fechado.

lab51-netless · bare metal  PASS

lab51-netless · célula com kernel sem pilha de rede

Uma célula PVH com um kernel compilado sem pilha de rede inicia e roda: /proc/net, /sys/class/net e iplink AUSENTES e a carga fecha com sum=500500, com o console pronto em menos de 1 s de boot do kernel.

limite  Fato de construção de um kernel, não veredicto de isolamento nem de desempenho; a evidência é o console e a configuração. n=1.

lab52-smt · bare metal  PASS

lab52-smt · placement contra o canal SMT

Um vizinho na thread irmã desacelera a vítima 4,98 a 5,94 vezes; em outro core cai a 1,07 a 1,13 vezes, perto da linha de base. O placement suprime o canal SMT (3 execuções).

limite  Mede interferência de throughput, não uma taxa de bits nem extração de chave. O residual de ~1,08x é o canal entre núcleos (L3). n=1.

lab53-llc · bare metal  PARTIAL

lab53-llc · canal entre núcleos pela L3

O canal entre núcleos existe e é pequeno: 1,03 a 1,14 vezes de interferência (média ~1,08x), uma ordem de grandeza abaixo do canal SMT. Residual medido e delimitado, não fechado.

limite  Sem Intel CAT/RDT não dá para particionar a L3 nesta CPU. Sinal pequeno e ruidoso (3 execuções); mede throughput. n=1.

lab54-meminspect · bare metal  CONFIRMED

lab54-meminspect · o domínio de controle lê uma célula viva

Controle positivo: o dom0 mapeou 10.076 quadros da célula viva e achou o segredo sem cooperação da célula (2 execuções). A assinatura autentica, mas não dá sigilo diante do dom0.

limite  Mede a exposição, não a fecha. Só a fecha memória cifrada por hardware (SEV-SNP/TDX); esta CPU não a tem. n=1.

lab54b · bare metal  CONFIRMED

lab54b-meminspect-incell · segredo nascido dentro da célula

3 de 3 execuções com 3 segredos distintos: 1 candidato por execução, hash verificado, 977 quadros contíguos. O residual de sigilo diante do dom0 fica medido.

limite  Não fecha o residual: pede memória cifrada por hardware (TDX/SEV-SNP), que esta máquina não tem, e essa prova de fechamento não foi executada. O mesmo resíduo existe em toda nuvem, AWS e GCP incluídas, se você não usa computação confidencial: quem controla o host, o provedor ou um atacante que consiga root nele, pode ler a memória do que roda nele. É uma propriedade da plataforma, não do Kawésqar. Testamos VMs confidenciais nas duas nuvens (TDX no GCP, SEV-SNP na AWS): a memória fica cifrada, mas a célula não consegue rodar dentro (não há KVM) e uma VM confidencial por trabalho leva de 50 a 58 s para subir. n=1.

lab55 · bare metal  PASS

lab55-composition · as cinco propriedades juntas

5 de 5 execuções PASS: o gate endurecido decide fora do dom0, sem chave privada no dom0; as ordens forjadas ou repetidas morrem pela assinatura e os mutantes são negados. Fecha a pendência «Composition».

limite  n=1, 5 repetições (reprodutibilidade, não estatística). O dom0 ainda cria domínios e inspeciona memória (residual do lab54).

lab56 · bare metal  PARTIAL

lab56-substrate · NetBSD completo ou Linux enxuto

Instanciação: NetBSD PVH ~1,65 s, Linux PVH 2,06 s, NetBSD PV 3,19 s. Ambos os substratos sustentam 8 células concorrentes com escalonamento quase plano de N=4 a N=8.

limite  Números de etapa inicial, anteriores às células atuais; o texto do lab é inconsistente com a própria tabela, por isso só se cita a tabela. n=1.

lab57-gate · bare metal  PASS

Célula rump real sob o Gate

A execução honesta dá PASS e 5 mutações dão seu DENY nomeado antes de rodar; cadeia de 3 células reais, 8/8 células com gate em paralelo, 20/20 em estabilidade. Uma célula «eco» sem rump passava com testemunha só de bytes e é negada ao atestar o kernel.

limite  PASS restrito, SEM contenção provada: o console é um meio degradado e o dom0 pode ler a RAM da célula (residual de hardware). n pequeno.

lab58-overhead · bare metal  PASS

Overhead do Xen e da célula rump frente ao Linux

Ciclo sem trabalho da célula rump ~94 ms contra ~500 ms da célula Linux (5,3x); boot 0,9 ms contra 237 ms; o Xen cobra +8 % de computação no dom0 e +11 % numa célula Linux PVH frente ao metal puro; a célula rump fica a +4 % do nativo.

limite  n=50 por célula, um equipamento. Uma anomalia na fase sha do rump (+15 a +20 %) fica UNCLEAR; não mede E/S. Números de create anteriores ao lançador direto.

lab59 · lógica + bare metal  PARTIAL

Red team de escape de célula (campanhas A a F)

Em metal uma célula comprometida leu 0 páginas de um vizinho, abriu 0 canais não autorizados e não ganhou autoridade; a limpeza por prefixo do lançador matou células de outro trabalho (fechado em lógica); fuzz nativo 85M execuções sem quedas; 15 de 16 itens do ledger fechados em lógica.

limite  O fechado é plano lógico e não foi remedido em metal. Seguem abertos: o verificador que não re-hasheia os bytes, o canal de cache entre núcleos e o dom0 lendo a RAM da célula.

redteam-escape · lógica  PARTIAL

redteam-escape · Onda 1 de ataques lógicos (C5, E5, E6, E8, CGAPS)

C5: o fechamento não foi ganho (um broker que forja a confirmação de armazenamento cruza; 23/23 comportamentos reproduzidos); E6: a limpeza por prefixo mata células de outro trabalho; E8: a checagem de memória admite 11 células a mais; E5: DENY. As falhas foram reparadas depois no lab59.

limite  Plano lógico em um único processo, sem metal; este lab registra o estado anterior à correção.

Etapa 5 · 15 labs

Células, cadeia assinada e relógio.

Lançadores diretos, o custo do Gate, imagens seladas, o construtor mínimo em três fases e a energia por célula. Aqui três hipóteses foram refutadas e uma verificação falhou, e tudo fica registrado.

lab61 · lógica  PASS

Negativos de Secure Boot e mapa adversarial

5 de 5 casos como esperado: o controle válido verifica e sem assinatura, chave alheia e certificado dessincronizado são rejeitados; o artefato sem SBAT é detectado como tal.

limite  Verificação de assinatura offline: a rejeição do carregador no boot real é descrita como esperada, não medida; 17 dimensões adversariais seguem pendentes.

lab62 · bare metal  HELD

kz-k2 em KVM contra Firecracker

Com a célula anterior ao corte, pronta do zero em 4,8 ms (kz-k2 em KVM) contra 8,2 ms do Firecracker com a mesma célula; frente ao Firecracker com guest Linux a hipótese pré-registrada se sustenta com 2,14 a 2,34x; a P=8 na célula mínima, 251 contra 147 células/s.

limite  n=1. Com a mesma célula a vantagem é de 1,04 a 1,10x em trabalhos de 100 ms: está na célula, não no VMM. Com trabalhos de 1000 ms se perde (~1x). Frente ao Firecracker com um Linux sem ajuste, 3,5x medido na mesma máquina.

lab63 · lógica  PASS

lab63 · cadeia Gate a launcher com Ed25519

107 testes negativos passam (recibo alterado, contraassinatura ausente ou alheia, linhagem quebrada, ordem mal assinada, replay, replay após reinício) e o fechamento das 5 imagens dá 84 de 84 verificações.

limite  Testado em lógica e construção estática, sem Xen nem VM; a evidência ponta a ponta no Xen com Ed25519 está em labs posteriores.

lab64 · bare metal  REFUTED

lab64 · custo do Gate e da ordem sobre o ciclo da célula

A hipótese pré-registrada «menos de 1 ms» é refutada: Gate mais ordem somam uma mediana de 2,735 ms (P=1, n=200); o delta A/B do tempo total é +2,574 ms, e +10,815 ms a P=8.

limite  n=1, governador ondemand. A P=8 o Gate fica em fila por um único lock, então não é «o Gate é mais lento sob carga». É uma falha registrada e publicada como traço.

lab65 · bare metal  PASS

lab65 · verificador fora do dom0 sobre células rump reais

Em metal 16 de 16 cenários como previsto; a extensão com substituição de artefato dá 19 de 19 em três execuções, 0 órfãos; em lógica 62 testes de fidelidade e 10 de mutação passam.

limite  n=1. Um dom0 que mente de forma coerente obtém PASS, declarado NÃO-REIVINDICAÇÃO deliberada. O estado do Gate vive na RAM do verifier. Canal por relé do dom0, sem mTLS. Veredicto restrito, não «contenção provada».

lab66 · bare metal  PARTIAL

lab66 · kz-c2 com kernel selado (fechamento do TOCTOU do lab64)

Verificar o hash a cada ciclo custa +7,16 ms sobre um ciclo de uns 15 ms e é descartado; a segunda variante custa -0,27 ms em total_ms (15,345 a 15,079), dentro do ruído, com 9 testes negativos OK.

limite  O ataque TOCTOU foi demonstrado só com um dublê do carregamento, não com Xen real. n=1, sem SHA-NI. A cadeia completa com Gate foi medida no lab70 (+3,21 ms).

lab68 · bare metal  PASS

lab68-e2e-real · evidência real de A até a célula B

10 de 10 cadeias A, Gate, ordem, lançador direto e B em metal (2 rodadas de 5), oráculo PASS em cada uma; em lógica 128 verificações do lançador e 54 de preflight sem falhas.

limite  n=1. O dom0 segue como base de confiança do tap, e o FETCH de B é um eco assinado, não um FETCH real ao broker. Sem mTLS nesse canal. O primeiro smoke em metal foi UNCLEAR e foi arquivado.

lab70 · bare metal  REFUTED

lab70-gate-fast-fd · custo do Gate rápido sobre lançar sem Gate

A hipótese H1 (menos de 1,5 ms de parede) é refutada: o braço completo acrescenta +3,210 ms de mediana (IC95 3,074 a 3,378) sobre a célula sem Gate; repetido com governor performance: +3,216 ms.

limite  n=1, Xen, P=1. O braço com descritor inclui um fsync de nonce durável que os outros não pagam.

lab71 · bare metal  PASS

lab71-gate-evidence · ativação única e ordem de liberação

1000 de 1000 rodadas com exatamente uma ativação (a linha de base sem o predicado cruza 300 de 300); ordens sem assinatura, alheias ou repetidas dão 0 criações; 10 de 10 ciclos com a liberação confirmada antes do veredito.

limite  n=1, N até 32, sem corte de energia no meio da escrita. Numa execução um braço deu FAIL de ambiente e uma sonda foi corrigida após três defeitos de instrumento: reportados como estão.

lab72 · bare metal  PASS

lab72-b-evidence · evidência do Pedido B

Após destruir, a enumeração de resíduos deu 0 em 100 de 100 células, com a destruição comprovada em no máximo 0,093 ms; livro persistente com kill -9 e reinício sem violações; criar o domínio Xen pelo lançador direto 41 ms contra 115 ms com xl. Inclui um modelo TLA+ com conformidade de traces.

limite  n=1. fs-verity e dm-verity ficaram BLOCKED no Xen e só foram testados com célula KVM em outro equipamento. Os 41 ms são criação de domínio Xen, não latência de célula.

kz-f1 · f0 · bare metal  PARTIAL

KZ-F1 fase 0: energia, política CPUID e páginas zeradas

C1, leitura de energia a partir do dom0: FAIL segundo o critério pré-registrado; a via de registradores MSR devolve erro de E/S sob Xen 4.20. Uma interface alternativa respondeu, mas não está validada (n=1, sem medidor externo; duas verificações registradas sobre ela também não cumpriram o critério) e afeta apenas a medição de energia. C2, célula sem política CPUID: revisar (80 de 80 pares com comportamento idêntico, o despejo difere). C3, páginas zeradas após reutilizar memória: PASS (210 domínios de 256 MiB, ao menos 51.636 MiB reutilizadas, 0 bytes diferentes de zero).

limite  ZBook, n=1, uma execução por verificação. O resultado de C1 não é um número de energia citável.

kz-f1 · f1 · bare metal  PASS

KZ-F1 fase 1: ciclo assinado, tempo de montagem e xenstore

Ciclo completo com ordem assinada: fumaça PASS (8 de 8 ordens válidas; 4 de 4 inválidas rejeitadas sem criar domínio), repetida com a frente final: PASS. H2, tempo de montagem contra kz-c2: DENTRO (diferença pareada de -2,1 a -3,9 ms; critério ≤ +1,0 ms), repetida com a frente final: DENTRO. H3, nós e escritas de xenstore por célula: PASS duas vezes (0 nós e 0 escritas em 200 células). H7a: PASS. Fumaça da varredura (o trabalhador morre com a célula viva): PASS, 8 de 8.

limite  ZBook, n=1: a fumaça, H2 e H3 rodaram duas vezes na mesma máquina; H7a e a varredura, uma. O tempo de montagem não é o de ponta a ponta: de ponta a ponta esta fase foi mais lenta que o kz-c2 (+7,9 a +16,2 ms).

kz-f1 · f2 · bare metal  PARTIAL

KZ-F1 fase 2: construtor residente, recuperação e escala

Ciclo de 15,1 a 18,1 ms contra 22,7 a 26,9 ms do braço de referência, com o Gate incluído. H1 DENTRO: melhora pareada de -8,672 a -7,243 ms com 200 pares por variante (critério ≤ -4,0 ms). H6 PASS: 13 estados × 100 e 200 términos limpos, 0 violações. Fumaça 8 de 8 com 3 negativos rejeitados; varredura 21 de 21; falhas reais 20 de 20 nas quatro células. Não cumpridas: H1b INCONCLUSIVA (1,022 a 1,059 ms contra limiar ≤ 1,0); H4 ENTRE LIMIARES (entrada 3,443 a 3,647 ms e 224 a 280 células/s; critério ≤ 3 ms e ≥ 243/s); H5 ENTRE LIMIARES (1.800 células por variante, sem muro, decil mínimo 74,1/s contra critério ≥ 78/s); H7b REFUTADA (7.840 linhas de código contra critério ≤ 1.800).

limite  ZBook, n=1, uma única execução por teste, sem selar. As definições de ciclo diferem das do KVM e das do rival (Xen contra KVM, com e sem o Gate de assinatura) e não se encadeiam. «Entre limiares» e «inconclusiva» não são aprovações.

energia · bare metal  PARTIAL

Energia por célula com contadores RAPL, kernel nativo sem hipervisor

Mesma tarefa de cerca de 100 ms, energia de processador e memória por célula, com o repouso subtraído: trabalho nativo 0,74 J, nossa célula 0,89 J, Firecracker com a nossa mesma célula 0,98 J, Firecracker com Linux 2,04 J. Sobrecusto sobre o trabalho nativo: 0,15 J contra 1,30 J. É menos da metade da energia de processador de uma microVM Firecracker com Linux, para o mesmo trabalho. Frente ao Firecracker com a mesma célula quase empata (0,89 contra 0,98 J), e com um processo novo por célula, 0,99 J, empata. Ciclo de vida completo de uma célula (inicia, passa no seu teste e desliga), sem trabalho: mínima 24 mJ, com rede 28, com arquivos 32, com rede e arquivos 38. 5 repetições de 400 células por braço, 0 falhas. Coerente com kz-f1 f0: ali a leitura de energia era feita a partir do dom0 sob Xen e falhou; aqui o kernel é nativo, sem hipervisor (KVM).

limite  ZBook (um laptop), kernel nativo, carga sintética (um hash em laço). É energia do chip (RAPL: processador e memória), não da tomada nem da conta de luz; não cobre rede, disco, ventiladores nem fonte. A economia frente ao Firecracker com Linux vem de trocar o Linux pela célula, não da célula frente ao Firecracker. Sem valores em dinheiro. Falta repetir em um servidor e validar com um medidor externo. Evidência registrada com hash e carimbo de tempo.

kz-c1 · bare metal  PASS

kz-c1, lançador de células com libxl direto

Ciclo de uns 38 ms (create 26 + vida 2 + destroy 9) contra uns 94 ms mais harness com xl; 64/64 células ok em cada configuração e teto de 50 a 60 células/s por nó a partir de P=4, porque a criação se serializa no toolstack.

limite  Medido no ZBook, anterior às gerações posteriores: não é o ciclo vigente nem um destaque de latência. Sem intervalos de confiança.

Etapa 6 · 2 labs

A nuvem, no seu lugar.

As execuções aninhadas orientam e nada mais: os vereditos de desempenho e isolamento são os medidos em metal.

kz-gc1 · nuvem (aninhada)  PARTIAL

kz-gc1: a célula KVM em uma VM de nuvem (KVM aninhado)

Célula pronta do zero em 6,9 ms contra 107 ms do Firecracker + Linux sem ajuste. Em trabalhos de ~100 ms: 2,11x no pior P, contra Linux sem ajuste a vantagem aparece; com a mesma célula no Firecracker fica em 1,04x (orienta). Rajada de 20.000 células sem falhas e 13.053 células vivas a 1,86 MB cada.

limite  Exploratório · na nuvem (aninhado) · orienta, não é veredito; os veredictos de desempenho e isolamento são os do ZBook. Uma rodada por ponto; as 13.053 foram cortadas por tempo, não é teto de RAM.

kz-az1 · bare metal + nuvem aninhada  PARTIAL

kz-az1: duelo em metal e execuções na AWS aninhada

Em metal, a hipótese de 2x em trabalhos de ~100 ms se sustenta só contra Linux sem ajuste (2,01x, no limite) e se falseia contra Linux ajustado (0,92x a 1,03x, empate técnico). Com a melhor configuração medida do recorte, kz-k2 pronta em 1,51 contra 2,32 ms do Unikraft e 1,03 contra 1,49 MiB por célula; antes do recorte o Unikraft chegava antes (2,25 contra 4,83 ms).

limite  Uma máquina, uma rodada por ponto. A P=15 o Unikraft fica à frente em taxa em 3 de 4 células; com os defaults o Unikraft vai à frente em tempo de pronta. AWS aninhado orienta, não é veredito. Snapshot-restore fica fora dos rankings. O mecanismo do recorte não é publicado.

Etapa 7 · 11 labs

O código de referência.

Os próprios componentes, como scripts autônomos que imprimem seus checks e saem com 0 só se cada vetor coincidiu com o veredito esperado. Quase todos no plano lógico.

impl · gate · lógica  PASS

Gate (portão de admissão do Job)

42/42 checks sem falhas (rc=0); reserva sem falhas e livro indexado 8/8. Nos 31 vetores adversariais ampliados (gate, broker, chain): 19 DENY, 9 GAP declarados, 3 LIMIT por projeto.

limite  Plano lógico. Os 9 GAP são publicados sem escondê-los; C5 foi o único de severidade alta. A execução em metal do Gate é o lab71.

impl · broker · lógica  PASS

Broker (E/S mediada com CID)

71/71 checks (68 vetores, pureza, derivação e assinaturas), 0 falhas, rc=0. Cobre o pin de chave do par: outro certificado da mesma CA com outra chave é negado.

limite  Plano lógico. Restam GAP declarados (B2, B5, B6 de projeto e B7, rollback de estado, que depende da durabilidade do dom0). Não é veredito de isolamento.

impl · chain · lógica  PASS

Chain (cadeia de recibos e linhagem)

37/37 cenários sem falhas, rc=0. A cobertura de assinatura passou a atestar o conjunto completo de campos (mover qualquer um dá BAD_SIGNATURE) e a linhagem exige um recibo de destruição real.

limite  Plano lógico. Achados de partida que são publicados: a linhagem era vazia por uma constante pública e a cobertura de assinatura era parcial; ambos foram corrigidos no código real.

impl · observer · lógica  PASS

Observer (contraassinatura e recibo de destruição)

14/14 checks sem falhas e cadeia de recibos de destruição com rc=0. Remover ou substituir a contraassinatura (6+3 variantes) é negado por MISSING_COSIGNATURE.

limite  Plano lógico. O observer contraassinava sem um fato de commit observado (C5); foi tratado no código real com um armazenamento separado. Aqui não se afirma independência em produção.

impl · ring · lógica  PASS

Ring (framing do canal de requisição e resposta)

62/62 vetores sem falhas, rc=0. O receptor valida só um snapshot da página, e os artefatos grandes viajam em chunks.

limite  Plano lógico (loopback em um só equipamento, sem canal Xen real). O fence do Python é de melhor esforço. Não é veredito de isolamento nem de desempenho.

impl · transport · lógica  PARTIAL

Transport (mTLS do plano de controle)

Em lógica: mtls 22/22, serve 21/21, integração 13/13 (handshake real, 0 SKIP), tier2 6/6, ca 18/18, identity 8/8. Em metal, uma vez: 12/12 suítes com handshake TLS 1.3 real. Outro certificado da mesma CA com outra chave é negado.

limite  Aberto: rodá-lo entre principais de SO distintos e entre domínios Xen separados. Em metal as duas pontas tinham o mesmo UID; o canal atual entre domínios é um relay sem mTLS.

impl · crypto · lógica  PASS

kzsig: assinatura assimétrica Ed25519

15/15 checks, 0 falhas: com só chaves públicas o verificador confere mas não forja; 300 assinaturas aleatórias, assinatura reutilizada e assinatura com a pública falham; um dom0 que assina com sua própria chave é rejeitado.

limite  Lógica pura. Segue aberto distribuir de forma autêntica as três chaves públicas; o dom0 ainda pode descartar ou atrasar; as privadas não são protegidas por esta camada.

impl · kaf-store · lógica  PASS

kaf-store: armazenamento isolado por processo (fechamento do C5)

22/22 checks (rc=0): pedir assinatura de bytes que nunca foram armazenados, armazenar lixo e declarar outro hash, ou escrever direto no meio, é negado; nenhuma privada alcançável no processo do broker.

limite  Fecha só o caso de um processo broker comprometido. Contra dom0 com root o isolamento de processo não basta: é residual de hardware. Não foi testado em metal neste diretório.

impl · alicebob · lógica  PARTIAL

alicebob: cartório e registro de chaves para agentes de console

Duas revisões adversariais acharam e corrigiram umas vinte falhas (cadeia reescrevível ou truncável, dono por Job, injeção na consulta de auditoria, entre outras); os testes Go de cadeia, assinatura, servidor e armazenamento passam.

limite  Achado NÃO corrigido com código: se daemon e agente rodam com o mesmo UID não há fronteira agente/humano. Bitácora linear, não Merkle; sem número de desempenho.

impl · rump-xen · bare metal  PASS

Célula KZ-Rump sobre Xen PVH (e a mesma imagem em KVM)

Sequência honesta: T15a link PASS; T15b o console FALHOU; T15c KZ-RUMP-OK; T16a só-fs BLOQUEADA (UNCLEAR); T16b só-fs PASS 3/3; T16c só-rede PASS, contenção de 8 threads sem abortos e um deadlock deliberado que morre observável aos 30 s. Um único ELF inicia em Xen e em KVM.

limite  n=1, uma execução por teste. Uma primeira auditoria achou um watchdog insuficiente e foi corrigido. Os números de boot vigentes são os do recorte, não os destes relatórios.

mtls-metal · bare metal  PARTIAL

mtls-metal: mTLS do plano de controle no dom0 físico

12/12 suítes verdes em metal com handshake TLS 1.3 real (um certificado da mesma CA com outra chave é negado). O arnês entre principais de UID distinto nunca rodou.

limite  As duas pontas com o mesmo UID e TLS em Python puro: não é separação entre principais de SO nem entre domínios Xen. n=1.

Memória

A memória, em toda nuvem.

Quem controla o host pode ler a memória do que roda nele: o provedor de nuvem, ou um atacante que consiga root no host. Toda nuvem tem isso (AWS, GCP, qualquer uma) se você não usa computação confidencial. Nem sempre se diz com estas palavras, mas é assim. Dizemos com clareza para que ninguém pense que é um problema da Kawésqar.

memória
Nuvem sem computaçãoconfidencialAWS, GCP, qualquer umasua VM / sua cargamemória em clarosegredohipervisor / hostrootlêO provedor, ou um atacante comroot no host, lê a memória.Servidor próprio,dom0 do Xenmedido em metal, n=1a célulamemória em clarosegredodom0rootlêRoot no dom0 lê a memória dacélula. Medido: 3 de 3, um segredonascido dentro da célula.ComputaçãoconfidencialTDX · SEV-SNPsua VM / sua cargamemória cifrada▒▓░ x9#f 7a▒ ░▓hipervisor / hostrootbloqueadaO host vê só texto cifrado.A leitura fica bloqueada.Mesmo modelo de segurança nas três;só muda a cifragem da memória.Nuvem sem computaçãoconfidencialAWS, GCP, qualquer umasua VM / sua cargamemória em clarosegredohipervisor / hostrootlêO provedor, ou um atacante comroot no host, lê a memória.Servidor próprio,dom0 do Xenmedido em metal, n=1a célulamemória em clarosegredodom0rootlêRoot no dom0 lê a memória dacélula. Medido: 3 de 3, um segredonascido dentro da célula.ComputaçãoconfidencialTDX · SEV-SNPsua VM / sua cargamemória cifrada▒▓░ x9#f 7a▒ ░▓hipervisor / hostrootbloqueadaO host vê só texto cifrado.A leitura fica bloqueada.Mesmo modelo de segurança nas três;só muda a cifragem da memória.
memória   Quem pode ler o quê. Seta ember: leitura a partir do host. Só a terceira pilha, com cifragem por hardware, a detém.
plataforma

É da plataforma, não da Kawésqar

Qualquer carga em um host sem computação confidencial é legível a partir do host. É uma propriedade da plataforma, igual para todo produto que roda nela.

testamos

Testamos em nuvem confidencial

TDX no GCP e SEV-SNP na AWS. A memória fica cifrada, mas a célula não consegue rodar lá dentro (não há KVM) e uma VM confidencial por trabalho leva de 50 a 58 s para subir.

o que pedimos

Por isso se valida em hardware físico

O fechamento na célula é validado em hardware físico (TDX ou SEV-SNP), que é o que pedimos.

Como usamos

O crítico fica em casa, o resto vai rápido.

Não propomos voltar atrás diante da tendência de ir para a nuvem. Usamos a velocidade da Kawésqar na nuvem para as cargas menos críticas, ou onde o risco pode ser tolerado, e deixamos em casa só o que merece: soberano, on-prem ou local. O mais crítico, por exemplo as chaves de um banco, fica em hardware próprio com cifragem de memória completa (TDX / SEV-SNP).

escada
Carga toleranteao risco→ Nuvemvelocidade KawésqarCargasensível→ On-premsoberano, em casaChaves de um banco→ On-prem com cifragem de memória completaTDX / SEV-SNPvalidado em hardware físicomenos críticomais críticoCarga toleranteao risco→ Nuvemvelocidade KawésqarCargasensível→ On-premsoberano, em casaChaves de um banco→ On-prem com cifragem de memória completaTDX / SEV-SNPvalidado em hardware físicomenos críticomais crítico
escada   Para onde vai cada tipo de carga. Só fica em casa o que merece.
mesmo contrato

Mesmo contrato, mesmo Gate

Dos dois lados valem o mesmo contrato, o mesmo Gate e o mesmo modelo de segurança. Só muda onde a carga roda.

medido

O que está medido hoje

O fechamento com cifragem de memória completa na célula é validado em hardware físico (TDX ou SEV-SNP), que é o que pedimos à Intel e à AMD. Hoje o residual está medido e declarado (lab54b, 3 de 3), não fechado.

ainda não

O que ainda não existe

A federação entre on-prem e nuvem é uma direção de projeto, ainda sem código.

Continue lendo

Evidência, depois o que pedimos.

A página de evidência põe os números principais em contexto e nomeia as lacunas que seguem abertas. Se quiser reproduzir algo daqui, conte-nos com clareza.

a regra

Não é um certificado nem uma afirmação de conformidade regulatória: é um registro datado e delimitado do que rodou.

Agent Fabric · lab notebook

More than 70 labs. The ones that failed too.

This is the lab notebook of the Agent Fabric: every experiment we can publish, with the plane it ran on, the verdict, the figure and the limit. Failures and refutations sit next to the passes, because a lab that only shows what worked is showing you a pitch, not evidence.

0+
experiments in the lab, and counting
0
shown here, each with its verdict, plane and limit
0
tested: PASS, HELD or mitigated by policy
0
published as FAIL or REFUTED, not buried

patent pending  Patent applications filed with INAPI (Chile); in process.

Verdict board

Every lab, by plane.

One chip per lab. Green is a tested pass, amber is partial, unclear, open or a measured residual, orange is a recorded failure or refutation. The plane matters as much as the colour: our results run nested (inside a VM, which validates neither isolation nor performance) or on bare metal, the only place those claims can close. Metal means one laptop workstation, n=1, unless a card says otherwise.

Nested VM · 16 labs

12 tested · 2 partial or open · 2 failed

Nested cloud · 1 lab

0 tested · 1 partial or open · 0 failed

Lab72 also carries a TLA+ model checked against traces, counted here under bare metal. In the cloud we show only what orients; it never closes a verdict.

What failed

What failed, said plainly.

A lab that only publishes its passes is selling. These are the places where the data said no, or said less than we hoped, and what we changed after.

nested VM  FAIL

lab14 · The rump guest printed nothing

FAIL on nested Xen. The guest reached long mode and ran rump_init, yet two console faults kept the marker from ever appearing. The failure dictated the order of fixes that lab16 closed.

nested VM  FAIL

lab15 · The observer did not fail closed

F9 FAIL: 38,577 mutations produced 4,520 crashes, 85 false countersignatures (only with a compromised stream) and 56 hangs. lab29 later reproduced the failure on metal, hardened the observer and reached 0 crashes in 3,000 variants.

bare metal  FAIL

lab25 · The disposable cell leaked on disk

RAM held (0 hits) but a reused volume let cell B read cell A's pattern: 6,882,854 hits. Zeroing the whole volume on create brought it to 0, a fix verified in the lab before it reached the product.

bare metal  REFUTED

lab64 · Gate plus order cost more than 1 ms

We pre-registered under 1 ms. The measured median is 2.735 ms (P=1, n=200). The hypothesis is refuted and the number is the one we publish.

bare metal  REFUTED

lab70 · The faster Gate still missed its target

H1 asked for under 1.5 ms of wall time. The full arm adds +3.210 ms (95% CI 3.074 to 3.378), and +3.216 ms again with the performance governor. Refuted twice.

bare metal  FAIL

KZ-F1 phase 0 · C1

FAIL against the pre-registered criterion: the register path returns an I/O error under Xen 4.20, before and after load. An alternative interface answered and followed the load, but it is not validated (n=1, no external meter) and two registered checks of it missed their criteria. It affects only energy measurement, and no energy figure is cited.

bare metal  REFUTED

KZ-F1 phase 2 · H7b

We pre-registered at most 1,800 lines of code for the production path. The count is 7,840, and above 2,000 counts as refuted. The hypothesis is refuted and the number is published as measured, without redefining the count.

bare metal  PARTIAL

lab20a · An observation we must not oversell

One DMA denial was seen once and could not be reproduced; under strict mode the positive control was itself denied. It is an observation, not a result, and we do not cite it as “strict denies unmapped transfers”.

logic  PARTIAL

C5 · The close was not won

A broker that forges the storage confirmation crossed the gate in the first logic wave (23/23 behaviors reproduced). Later work added a separate store; the card for kaf-store shows what that closes and what it does not.

bare metal  TIE

lab27 · “Rump exposes less” was false

The hypothesis was falsified: rump keeps 2 live event channels against 0 for minimal Linux, and they tie on grants and xenstore nodes. We concluded rump did not justify deep investment on that ground alone.

bare metal + nested cloud  PARTIAL

kz-az1 · The 2x claim only half holds

On 100 ms jobs the 2x edge holds against untuned Linux (2.01x, borderline) and is falsified against tuned Linux (0.92x to 1.03x). Before the trim, Unikraft was ready sooner. We lead with the part that did not hold.

Timeline

Six stages, one thread.

The notebook follows the order we worked in: sandbox first, then Xen nested in a VM, then the physical machine, then attacks on our own design, then the fast cells, and last the cloud. Bar height is the number of labs shown for each stage; the reference code (11 more cards) sits apart at the end.

timeline
2SANDBOX14NESTED XEN27BARE METAL16RED TEAM15KZ CELLS2CLOUDSANDBOX2NESTED XEN14BARE METAL27RED TEAM16KZ CELLS15CLOUD2
timeline   Labs shown per stage. A nested or cloud result never counts as an isolation or performance verdict: only bare metal can close those.
Stage 1 · 16 labs

Nested Xen, the first answers.

Inside a VM (Xen under QEMU/KVM). These runs show the architecture working as logic on real domains. They are never an isolation or performance verdict, and each card says so.

Each card: lab, plane and verdict, then what was found, then the limit that comes with it. Figures are the ones the lab recorded; where a later campaign superseded them, the card says so.

lab03 · nested VM  PARTIAL

NetBSD rump kernels: closing the unknowns

The real IP stack uses 52 hypercalls (base 46) and 28 of 46 are invoked at startup; in the official order the real definition wins in 3 of 3 duplicates; the sysctl warning is benign.

limit  One nested guest; partially reproducible build (186/190 identical); no conclusion on performance or isolation.

lab04 · nested VM  PASS

Build non-determinism, guest disk, skill tests

The difference was only embedded source paths: two MKREPRO builds come out identical, 190/190. It corrects the lab03 hypothesis; skill tests 12/12 after rerunning one.

limit  n=2 builds per condition; the exact cause is strong inference, not fact; one case failed on the first pass.

lab05 · nested VM  PASS

Nested Xen and a minimal PVH cell

The cell prints its marker and exits on its own; 5 cycles and 1 destroy during boot leave 0 domains and 0 residual xenstore keys.

limit  All nested and without IOMMU: not valid for performance or isolation; not a benchmark.

lab06 · nested VM  PARTIAL

Skill validation and nested Xen cell (A to D)

A: 10/12 PASS; B: 7 PASS and 1 FAIL (one case invented a format); C: 5/5, 30 cycles with 0 residue and 0 failures; D: 5/5. The rump hypercall count is 46 on both sides.

limit  Nested, no IOMMU; three cases cannot be reproduced because their prompt and criterion are not on disk.

lab07 · nested VM  PASS

S4.0 xenstore probe between two cells

Without an ACL the cross read returns EACCES (30/30 in the control); the owning cell grants access and the other reads the nonce and gets 3 events. dom0 is not required.

limit  Nested, N=1 per arm, a single xenstored variant; no quotas or concurrency.

lab07b · nested VM  PASS

S4.2 gntalloc+gntdev+evtchn ring between two cells

Bidirectional ring with identical hashes at both ends, no dom0 privilege. The first run FAILED because of our own faulty protocol and was fixed.

limit  Nested, n=1 PASS; N≥30 pending; it was not shown that a third domain cannot map the grant.

lab08 · nested VM  PASS

Placeholder domains for control, broker and observer

30 cycles (90 domains): 0 XenStore residue and 90/90 exit code 0; a test Cell was denied in 15/15 operations on the three role branches.

limit  Nested, one run, placeholder domains with no real logic; not an isolation verdict.

lab09 · nested VM  PASS

Role logic inside an initramfs

The three suites pass inside a 256 MB PVH domain (17/17, 31/31, 9/9) with ~167-172 MB free and 0 out-of-memory signals.

limit  Nested, n=1; functional viability and memory headroom only.

lab10 · nested VM  PASS

Broker in one domain and a Cell over the ring

5/5: a 9000 B artifact with identical hash both ways; foreign or forged CID, replay and forced destination denied; 0 processes of the pair in dom0.

limit  Nested, n=1; no conclusion on performance or resistance on Xen bare metal.

lab11 · nested VM  PASS

Observer that countersigns and issues a destruction receipt

4/4: countersignature only while the Cell was alive during a real STORE; destruction receipt only after destroying it; 17/17 attempts to read, alter or delete the stream denied.

limit  Nested, n=1; does not cover a hostile Cell with kernel capabilities.

lab12 · nested VM  PASS

Chain A, receipt, B across real domains

All 4 conditions hold: B does not exist before the order, without a countersignature the Gate denies, a Cell trying to sign the order is denied (18 times) and the hash is identical end to end.

limit  PASS bounded to architecture: nested, n=1, first-generation symmetric signature; the first attempt failed on packaging.

lab13 · nested VM  PASS

B1.8 Falsifications only the substrate can refute

All 3 falsifications PASS: CIDs from the previous epoch give STALE_EPOCH after recreating the broker; a third domain cannot map the grant (128 refs denied); a kill between STAGED and COMMITTED gives NOT_COMMITTED. Finding: epoch and nonces must be durable outside the domain.

limit  Nested, n=1, HMAC v0 signature; the kill instant is forced; dom0 remains in the trust base.

lab14 · nested VM  FAIL

T15b First boot of rump as a PVH guest under Xen

Instructive FAIL: the guest entered via PVH, reached long mode and ran rump_init, but did not print the marker. Two console faults (unprogrammed hypercalls and a buffer removed by the compiler) dictated the fixes that lab16 closed.

limit  Nested, n=1; it stayed unobserved whether rump_init returned 0.

lab15 · nested VM  FAIL

S9.2 Observer falsifications 8 and 9

F8 PASS: the observer survives a cell crash and the Gate does not use its self-report. F9 FAIL: with 38,577 mutations there were 4,520 crashes, 85 false countersignatures (only with a compromised stream) and 56 hangs. A crash signs nothing: it is an availability failure.

limit  Nested, n=1, HMAC v0; the lie is a fixed script; dom0 remains in the trust base.

lab16 · nested VM  PASS

T15c Closing the rump axis: the rump guest prints KZ-RUMP-OK under Xen

With the two console fixes, the rump guest prints the marker, rump_init returns 0 and it shuts down by hypercall with code 0. It closes the bit lab14 could not read.

limit  Nested, n=1: init and console only; no performance or hard isolation.

lab17 · nested VM  PASS

T16 Adversarial guest: A_effective contained in A_delegated

289 attempts to map foreign grants and 289 to bind event channels gave 0 successes, 0 foreign pages read, 0 self-grants. What Xen contains is separated from what Kawésqar contains (replay, STALE_EPOCH, NOT_COMMITTED).

limit  Nested, n=1; the guarantee assumes a correct Xen; grant and channel denials are Xen's, not Kawésqar's.

Stage 2 · 12 labs

The first campaign on metal.

On the physical control domain of one laptop workstation, n=1. Here the first honest failures appear: a covert channel we could not measure and a disk that leaked.

lab18 · bare metal  PASS

Physical baseline and gate for the lab plane

Inventory captured and preflight 11/11. The CPUs in this Xeon are not affected by L1TF, MDS or TAA, so the side-channel lab was re-scoped.

limit  ZBook, n=1, a single day; this is inventory and preparation, not an isolation or performance result.

lab19 · bare metal  PASS

Re-attesting the architecture core on metal

7/7 core verdicts PASS on the physical control domain. Minimal cell cycle: p50 0.95 s, p99 1.09 s, 200/200 without failure, real residue 0.

limit  ZBook, n=1, control domain capped at 4 vCPU: correctness, not performance; the chain on real domains is still HMAC.

lab20a · bare metal  PARTIAL

VT-d enforcement on bare metal with a named DMA device

VT-d configuration was attested and a denial was seen once, but it is not reproducible. Under strict mode the instrument ended UNCLEAR: the positive control was denied.

limit  ZBook, n=1. It must not be cited as “strict denies unmapped transfers”: it is an observation, not a tested result.

lab21 · bare metal  PASS

Baseline latency of a cell on bare metal

200/200 per variant, 0 residue: create with xl 184 ms (p50) and kernel plus init boot 607 ms (p50, p99 790 ms). A cell with work adds a bounded overhead.

limit  First-generation figures (Linux kernel and xl), long before today's cells: do not compare them.

lab22 · bare metal  PASS

Scale and load

0 Gate failures in 256 runs under load; 2,775 cycles completed at 100%; creation steady around 237 ms from N=1 to N=414; 0 residual cells.

limit  This is the Gate logic under scheduling pressure, not real cell transitions at scale; the resource ceiling was never stressed.

lab23 · bare metal  UNCLEAR

Contention covert channel

The positive control did not cross the threshold: best case 9.64 bps with 95% CI [-33.06 ; 52.31], indistinguishable from noise. Neither presence nor absence of a channel is claimed.

limit  Instrument not validated; recorded as an unmeasured channel, never as negligible.

lab23b · bare metal  MITIGATED

Covert channel: closed by policy

Risk closed by a placement policy (do not co-schedule mutually distrusting cells on sibling threads), not by measurement: the strongest SMT channel does not exist by construction.

limit  It is not “we showed there is no channel”: shared-cache and cross-core memory channels remain as residual risk.

lab24 · bare metal  OPEN

XSAs that really apply

Three advisories patched, the speculative ones inapplicable on this silicon or mitigated; one applicable advisory remains unpatched (a liveness DoS in the xenstore daemon). None is an escalation or cross-domain leak.

limit  State tied to a package version; it must be rerun when the build changes.

lab25 · bare metal  FAIL

Disposable cell: RAM and disk residue

RAM PASS: 0 hits with forced overlap of ≥548 MiB. Disk FAIL: cell B sees A's pattern (6,882,854 hits); zeroing the whole volume on create brings it to 0 hits.

limit  The mitigation was verified in the lab but was not yet in the product; device DMA residue out of scope.

lab26 · bare metal  PARTIAL

Closing the bare metal campaign

Architecture: bounded PASS, conditional close. No amplification of authority or evidence on real Xen (7/7) and under load; disposable cell falsified on disk and PASS on RAM.

limit  No verified physical isolation is claimed, nor unconditional disposable cell, nor the full thesis; covert channel UNCLEAR.

lab27 · bare metal  TIE

Boundary shootout: KZ-Rump against minimal Linux

The hypothesis “rump exposes less” was falsified: rump keeps 2 live event channels (minimal Linux 0); they tie on grants (0) and on 38 xenstore nodes. Finding: an adapter abort gave reason code 0, same as a clean shutdown.

limit  Observed surface, not reachable; both runtimes are near-stubs; one machine, one run.

lab28 · logic + bare metal  PASS

Delegation tree n>1 and non-amplification by induction

Real Gate: 1,958 edges decided, 0 amplifications; real cells at the exact envelope edge (768 of 768 MiB); a grandchild asking for more than its parent is denied with AUTHORITY_NOT_SUBSET; 5 mutants detected.

limit  The parentage tree is modeled (Xen does not know it); depth above 2, fan-in and fuzz are logic. The lab found real Gate defects, since fixed.

Stage 3 · 15 labs

Pushing the limits.

Still on metal: observer fuzzing, the signed chain inside real cells, saturation to 412 live cells, strict IOMMU and observed hypercalls.

lab29 · bare metal  PASS

Falsification 9: observer surface

The first run FALSIFIED the property: 47 crashes and 208 violations through holes reachable via JSON. After hardening the observer: 0 crashes and 0 violations across 3,000 variants, 3 mutants detected, 0 residue.

limit  PASS bounded to the wire (decoded JSON stream); in-process violations remain, declared as a GAP. n=1.

lab30 · bare metal  PASS

Falsification 10: transactional STORE under kill

At no stage was a partial COMMITTED deliverable left (7 real kill -9); replay after restart always gave STALE_EPOCH; 3 mutants detected; real cells destroyed.

limit  Journal persistence is modeled by the harness: it tests the contract's atomicity, not the production persistent CAS (declared GAP).

lab31 · bare metal  PASS

Adversarial Gate under load

0 acceptances out of 6,870 concurrent adversarial attempts; 1,194 of 1,194 legitimate ones cross under the same load; the naive race over-reserves 6 of 6 (the instrument sees the race); 0 residue.

limit  The ledger lock is in-process; production persistent CAS is out. A finite corpus, not a general proof of absence.

lab32 · bare metal  PASS

Stability soak of the create-run-destroy cycle

400 of 400 iterations: zero drift (Kendall tau 0.02 to 0.08), residue 0, Xen free memory 55,883 to 55,883 MiB, dom0 stable.

limit  Bounded window (not months), one minimal cell, one machine, n=1.

lab33 · bare metal  PASS

Saturation at maximum scale and throughput

Peak of 412 simultaneously live cells (52,736 of 53,835 MiB), 0 amplification and 0 residue across 14 levels; at most 14.78 cells/s creation and 8.98 cells/s full cycle.

limit  n=1. At the peak, destruction reaches p99 32 s because the toolstack serializes; at scale only a sample carries a console marker.

lab34 · logic + bare metal  PASS

Resource fan-out (Falsification #4)

In logic: 4,681 nodes and 7,608 edges with 0 amplifications; on metal, with 128 MiB cells, every step completes, 0 amplifications and per-cell overhead of 1.5 to 2.5 MiB.

limit  The first two physical attempts were UNCLEAR (cause found in lab40); the close is the third. n=1.

lab35 · bare metal  PASS

Minimal observed boundary of the cell under load

Full classification of 18 interfaces and 49 hypercalls; load did not break the boundary, 0 residue in all 4 scenarios. Xen, not the runtime, bounds isolation.

limit  Characterization, not an isolation verdict. Trace observation stayed UNKNOWN on all 49 rows for lack of a decoder (lab44 closed it).

lab38 · bare metal  PASS

End-to-end Ed25519 cryptographic chain inside cells

The chain of cell A, countersignature, Gate, cell B and destruction receipt runs on metal; 4 mutants (dom0 forgery, altered receipt, B without evidence, replay) and 7 negative vectors rejected without creating B.

limit  Verifying the signature before storing is modeled by the harness, not by the reference Gate. n=1.

lab39 · bare metal  PASS

Smoke and invariant regression under strict IOMMU

All 4 groups pass: 60/60 iterations, 0 accepted out of 3,444 adversarial attempts with 600/600 legitimate admitted, and zeroing verified; create rises about 12% (some 23 ms) under strict.

limit  Bounded regression, not exhaustive re-testing; the clean A/B in the same boot is pending. n=1.

lab40 · bare metal  PASS

Boot memory floor of a PVH cell

Measured floor of 96 MiB: with 64 MiB the cell does not boot, with 96 and 128 MiB it does. Closes the root cause of the lab34 failure and falsifies the harness race hypothesis.

limit  Coarse floor (n=3 per point), valid for this lab kernel and initrd; the exact value between 65 and 96 MiB was not refined.

lab41 · bare metal  PASS

Audit of dom0 grants over cell pages

A diskless cell, idle or under load, has exactly 1 active grant mapped by dom0 (the console ring), 0 data pages; with a disk 67 refs appear (positive control).

limit  The first run was UNCLEAR because of a blind positive control and was fixed. That they enter dom0's IOMMU is inference from the source. n=1.

lab42 · bare metal  PASS

Entropy boundary inside the cell

14 of 14 distinct keys; in no boot was a key generated while the random generator was not ready, even with RDRAND off (median 1.47 s versus 0.14 s).

limit  It shows availability and non-determinism, not entropy quality. The no-RDRAND case rests on reading the source. n=1.

lab43 · bare metal  PASS

Cryptographic chain in a cell with n greater than 1

The chain and the fan-out cross end to end and every mutant is rejected with its exact reason; the fan-out fills the exact envelope (2 by 128 MiB). Finding: the pure gate does not consume the prev link of destruction receipts.

limit  Signature verification and the prev link are modeled in the harness (production gap declared). n=1.

lab44 · bare metal  PASS

Hypercalls observed through a real trace

In all 4 scenarios what is observed is a subset of the declared set of 10 hypercalls; the rump cell leaves a minimal footprint (5 distinct numbers) and Linux adds 4 more. Negative control: 0 calls with the cell paused.

limit  The hypercall number is observed, not the sub-op or the arguments. n=1.

lab45 · bare metal  PASS

Soak under strict IOMMU and structure leaks

400 of 400 cycles without leaks: active grants flat at 2, maptrack flat at 4 and Xen free memory constant (55,883 MiB before and after).

limit  Free memory is read in MiB (detectable minimum near 80 KiB per cycle); active grants are a partial proxy. n=1.

Stage 4 · 16 labs

Attack, then verify outside dom0.

The decision moves out of the control domain, then we attack our own design, measure what dom0 can still read and compare against Linux.

lab46 · logic  PASS

Denial conformance and live agent against the real channel

12 of 12 checks and the independent oracle PASS, each denial with its exact reason; 0 escalations crossed with the 8 language models tested, including the 3 that did attack through the real channel.

limit  Logic level, not metal nor an agent inside the cell. A later campaign corrected the lab's own overclaims (mutants that mutated data, not code).

lab47 · bare metal  PASS

Verifier outside dom0 (chain A to B)

The positive control crosses and all 11 mutants are denied with their exact reason; creations = admissions = verifier PASS (1 = 1 = 1); dom0 forgeries and replays die by signature. Re-attested on metal: 3 of 3 runs PASS.

limit  n=1, one host. dom0 still creates domains and can read the RAM of a live domain (hardware residual). The first physical run was UNCLEAR because of an oracle bug, fixed.

lab48 · bare metal  PASS

Hardened gate (gate_verified) against the harness

With the harness switched off, the old gate ACCEPTS and creates the child in the 4 main mutants; the new one DENIES them with its reason. The properties move from modeled in the harness to enforced by the gate.

limit  n=1. dom0 still generated the private keys in this run (lab47 attacks that); a lying tap is not covered.

lab49-c5 · logic + bare metal  PASS

Closing C5: compromised broker

The old observer accepts both forgeries; kaf-store denies both and the byte witness denies the first. On metal, a compromised broker that signed without backing the bytes leaves the medium empty and does not cross.

limit  The witness lets through a false COMMITTED over bytes honestly in staging. This does not settle the later red-team marker, which records the C5 close as not won. n=1.

lab50-dma · bare metal  PASS

DMA confinement of an assigned device

dom0 reads the USB stick before assignment, loses the device on assignment, and the domU reads the same number of sectors with 0 VT-d faults in strict mode. It measures confinement, not adversarial isolation.

limit  Partial: a malicious driver attempting DMA into another domain's memory was not tested. It ran in HVM because PVH does not support passthrough. Do not cite as closed DMA isolation.

lab51-netless · bare metal  PASS

lab51-netless · cell with a kernel without a network stack

A PVH cell with a kernel built without a network stack boots and runs: /proc/net, /sys/class/net and iplink ABSENT and the load closes with sum=500500, with the console ready in under 1 s of kernel boot.

limit  A fact about how a kernel was built, not an isolation or performance verdict; the evidence is the console and the configuration. n=1.

lab52-smt · bare metal  PASS

lab52-smt · placement against the SMT channel

A neighbor on the sibling thread slows the victim 4.98 to 5.94 times; on another core it drops to 1.07 to 1.13 times, near baseline. Placement suppresses the SMT channel (3 runs).

limit  It measures throughput interference, not a bit rate or key extraction. The ~1.08x residual is the cross-core channel (L3). n=1.

lab53-llc · bare metal  PARTIAL

lab53-llc · cross-core channel through the L3

The cross-core channel exists and is small: 1.03 to 1.14 times interference (average ~1.08x), an order of magnitude below the SMT channel. A residual measured and bounded, not closed.

limit  Without Intel CAT/RDT the L3 cannot be partitioned on this CPU. Small, noisy signal (3 runs); it measures throughput. n=1.

lab54-meminspect · bare metal  CONFIRMED

lab54-meminspect · the control domain reads a live cell

Positive control: dom0 mapped 10,076 frames of the live cell and found the secret without the cell's cooperation (2 runs). The signature authenticates, but gives no secrecy against dom0.

limit  It measures the exposure, it does not close it. Only hardware-encrypted memory (SEV-SNP/TDX) closes it; this CPU lacks it. n=1.

lab54b · bare metal  CONFIRMED

lab54b-meminspect-incell · secret born inside the cell

3 of 3 runs with 3 distinct secrets: 1 candidate per run, hash verified, 977 contiguous frames. The secrecy residual against dom0 is now measured.

limit  It does not close the residual: that needs hardware-encrypted memory (TDX/SEV-SNP), which this machine lacks, and that closing test has not been run. The same residual exists in every cloud, AWS and GCP included, unless confidential computing is on: whoever controls the host, the provider or an attacker who gets root on it, can read the memory of what runs on it. It is a property of the platform, not of Kawésqar. We did try confidential VMs in both clouds (TDX on GCP, SEV-SNP on AWS): memory is encrypted, but the cell cannot run inside (no KVM) and a confidential VM per job takes 50 to 58 s to come up. n=1.

lab55 · bare metal  PASS

lab55-composition · the five properties together

5 of 5 runs PASS: the hardened gate decides outside dom0, with no private key in dom0; forged or repeated orders die by signature and mutants are denied. Closes the “Composition” item.

limit  n=1, 5 repetitions (reproducibility, not statistics). dom0 still creates domains and can inspect memory (lab54 residual).

lab56 · bare metal  PARTIAL

lab56-substrate · full NetBSD or trimmed Linux

Instantiation: NetBSD PVH ~1.65 s, Linux PVH 2.06 s, NetBSD PV 3.19 s. Both substrates sustain 8 concurrent cells with near-flat scaling from N=4 to N=8.

limit  Early-stage figures, earlier than today's cells; the lab's text disagrees with its own table, so only the table is cited. n=1.

lab57-gate · bare metal  PASS

Real rump cell under the Gate

The honest run PASSES and 5 mutations give their named DENY before running; a chain of 3 real cells, 8/8 cells gated in parallel, 20/20 in stability. An “echo” cell without rump passed with a bytes-only witness and is denied once the kernel is attested.

limit  Bounded PASS, NOT containment shown: the console is a degraded medium and dom0 can read the cell's RAM (hardware residual). Small n.

lab58-overhead · bare metal  PASS

Xen and rump cell overhead against Linux

Rump cell cycle without work ~94 ms versus ~500 ms for the Linux cell (5.3x); boot 0.9 ms versus 237 ms; Xen charges +8% compute in dom0 and +11% in a Linux PVH cell versus bare metal; the rump cell sits at +4% of native.

limit  n=50 per cell, one machine. An anomaly in the rump sha phase (+15 to +20%) stays UNCLEAR; no I/O measured. Create figures predate the direct launcher.

lab59 · logic + bare metal  PARTIAL

Cell-escape red team (campaigns A to F)

On metal a compromised cell read 0 pages of a neighbor, opened 0 unauthorized channels and gained no authority; the launcher's prefix cleanup did kill cells of another job (closed in logic); native fuzz 85M executions without crashes; 15 of 16 ledger items closed in logic.

limit  What is closed is the logic plane and has not been re-measured on metal. Still open: the verifier that does not re-hash the bytes, the cross-core cache channel and dom0 reading cell RAM.

redteam-escape · logic  PARTIAL

redteam-escape · Wave 1 of logic attacks (C5, E5, E6, E8, CGAPS)

C5: the close was not won (a broker that forges the storage confirmation crosses; 23/23 behaviors reproduced); E6: prefix cleanup kills cells of another job; E8: the memory check admits 11 extra cells; E5: DENY. The failures were repaired later in lab59.

limit  Logic plane in a single process, no metal; this lab records the state before the fix.

Stage 5 · 15 labs

Cells, a signed chain and the clock.

Direct launchers, the cost of the Gate, sealed images, the minimal builder in three phases and the energy per cell. Three hypotheses were refuted and one check failed here, and all are kept.

lab61 · logic  PASS

Secure Boot negatives and adversarial map

5 of 5 cases as expected: the valid control verifies, and unsigned, foreign-key and mismatched-certificate artifacts are rejected; the artifact without SBAT is detected as such.

limit  Offline signature verification: loader rejection at real boot is described as expected, not measured; 17 adversarial dimensions remain pending.

lab62 · bare metal  HELD

kz-k2 on KVM against Firecracker

With the cell before the trim, ready from zero in 4.8 ms (kz-k2 on KVM) against 8.2 ms for Firecracker with the same cell; against Firecracker with a Linux guest the pre-registered hypothesis holds at 2.14 to 2.34x; at P=8 on the minimal cell, 251 versus 147 cells/s.

limit  n=1. With the same cell the edge is 1.04 to 1.10x on 100 ms jobs: it lives in the cell, not the VMM. On 1000 ms jobs it is lost (~1x). Against Firecracker with an untuned Linux, 3.5x measured on the same machine.

lab63 · logic  PASS

lab63 · Gate to launcher chain with Ed25519

107 negative tests pass (altered receipt, missing or foreign countersignature, broken lineage, badly signed order, replay, replay after restart) and the closure of the 5 images gives 84 of 84 checks.

limit  Tested in logic and static build, without Xen or a VM; the end-to-end Xen evidence with Ed25519 is in later labs.

lab64 · bare metal  REFUTED

lab64 · Gate and order cost over the cell cycle

The pre-registered “under 1 ms” hypothesis is refuted: Gate plus order add a median of 2.735 ms (P=1, n=200); the A/B delta in total time is +2.574 ms, and +10.815 ms at P=8.

limit  n=1, ondemand governor. At P=8 the Gate queues behind a single lock, so this is not “the Gate is slower under load”. A recorded failure, published as a trait.

lab65 · bare metal  PASS

lab65 · verifier outside dom0 over real rump cells

On metal 16 of 16 scenarios as predicted; the extension with artifact substitution gives 19 of 19 in three runs, 0 orphans; in logic 62 fidelity tests and 10 mutation tests pass.

limit  n=1. A dom0 that lies coherently gets PASS, declared a deliberate NO-CLAIM. Gate state lives in the verifier's RAM. dom0 relay channel, no mTLS. A bounded verdict, not “containment shown”.

lab66 · bare metal  PARTIAL

lab66 · kz-c2 with a sealed kernel (closing the lab64 TOCTOU)

Verifying the hash on every cycle costs +7.16 ms on a cycle of about 15 ms and is discarded; the second variant costs -0.27 ms in total_ms (15.345 to 15.079), within noise, with 9 negative tests OK.

limit  The TOCTOU attack was shown only with a stand-in for the loader, not real Xen. n=1, no SHA-NI. The full chain with Gate was measured in lab70 (+3.21 ms).

lab68 · bare metal  PASS

lab68-e2e-real · real evidence from A to cell B

10 of 10 chains of A, Gate, order, direct launcher and B on metal (2 batches of 5), oracle PASS in each; in logic 128 launcher checks and 54 preflight checks without failures.

limit  n=1. dom0 remains the trust base for the tap, and B's FETCH is a signed echo, not a real FETCH to the broker. No mTLS on that channel. The first smoke on metal was UNCLEAR and was archived.

lab70 · bare metal  REFUTED

lab70-gate-fast-fd · cost of the fast Gate over launching without a Gate

Hypothesis H1 (under 1.5 ms of wall time) is refuted: the full arm adds +3.210 ms median (95% CI 3.074 to 3.378) over the cell without Gate; repeated with the performance governor: +3.216 ms.

limit  n=1, Xen, P=1. The descriptor arm includes a durable nonce fsync that the others do not pay.

lab71 · bare metal  PASS

lab71-gate-evidence · single activation and release order

1000 of 1000 rounds with exactly one activation (the baseline without the predicate crosses 300 of 300); unsigned, foreign or repeated orders give 0 creations; 10 of 10 cycles with the release confirmed before the verdict.

limit  n=1, N up to 32, no power cut mid-write. In one run an arm gave an environment FAIL and a probe was fixed after three instrument defects: reported as they are.

lab72 · bare metal  PASS

lab72-b-evidence · evidence for Application B

After destroy, residue enumeration gave 0 in 100 of 100 cells, with destruction verified in at most 0.093 ms; persistent ledger with kill -9 and restart without violations; creating the Xen domain through the direct launcher 41 ms versus 115 ms with xl. It includes a TLA+ model with trace conformance.

limit  n=1. fs-verity and dm-verity were BLOCKED on Xen and were only tested with a KVM cell on another machine. The 41 ms is Xen domain creation, not cell latency.

kz-f1 · f0 · bare metal  PARTIAL

KZ-F1 phase 0: energy, CPUID policy and zeroed pages

C1, energy reading from dom0: FAIL against the pre-registered criterion; the MSR register path returns an I/O error under Xen 4.20. An alternative interface answered, but it is not validated (n=1, no external meter; two registered checks of it also missed their criteria) and it affects only energy measurement. C2, cell without CPUID policy: REVIEW (80 of 80 pairs with identical behavior, the dump differs). C3, zeroed pages after reusing memory: PASS (210 domains of 256 MiB, at least 51,636 MiB reused, 0 non-zero bytes).

limit  ZBook, n=1, one run per check. C1 is not a citable energy figure.

kz-f1 · f1 · bare metal  PASS

KZ-F1 phase 1: signed cycle, build time and xenstore

Full cycle with a signed order: smoke PASS (8 of 8 valid orders; 4 of 4 invalid orders rejected without creating a domain), repeated with the final front: PASS. H2, build time against kz-c2: INSIDE (paired difference of -2.1 to -3.9 ms; criterion ≤ +1.0 ms), repeated with the final front: INSIDE. H3, xenstore nodes and writes per cell: PASS twice (0 nodes and 0 writes across 200 cells). H7a: PASS. Sweep smoke (the worker dies with the cell alive): PASS, 8 of 8.

limit  ZBook, n=1: the smoke, H2 and H3 ran twice on the same machine; H7a and the sweep once. Build time is not end to end: end to end this phase was slower than kz-c2 (+7.9 to +16.2 ms).

kz-f1 · f2 · bare metal  PARTIAL

KZ-F1 phase 2: resident builder, recovery and scale

Cycle of 15.1 to 18.1 ms against 22.7 to 26.9 ms for the reference arm, Gate included. H1 INSIDE: paired improvement of -8.672 to -7.243 ms with 200 pairs per variant (criterion ≤ -4.0 ms). H6 PASS: 13 states × 100 and 200 clean endings, 0 violations. Smoke 8 of 8 with 3 negatives rejected; sweep 21 of 21; real failures 20 of 20 across the four cells. Not met: H1b INCONCLUSIVE (1.022 to 1.059 ms against a threshold of ≤ 1.0); H4 BETWEEN THRESHOLDS (entry 3.443 to 3.647 ms and 224 to 280 cells/s; criterion ≤ 3 ms and ≥ 243/s); H5 BETWEEN THRESHOLDS (1,800 cells per variant, no wall, lowest decile 74.1/s against a criterion of ≥ 78/s); H7b REFUTED (7,840 lines of code against a criterion of ≤ 1,800).

limit  ZBook, n=1, one run per test, unsealed. The cycle definitions differ from the KVM ones and from the rival's (Xen against KVM, with and without the signing Gate) and are not chained. “Between thresholds” and “inconclusive” are not passes.

energy · bare metal  PARTIAL

Energy per cell with RAPL counters, native kernel without hypervisor

Same task of about 100 ms, processor and memory energy per cell, with idle subtracted: native work 0.74 J, our cell 0.89 J, Firecracker with our same cell 0.98 J, Firecracker with Linux 2.04 J. Overhead over the native work: 0.15 J against 1.30 J. That is less than half the processor energy of a Firecracker microVM with Linux, for the same work. Against Firecracker with the same cell it is close to a tie (0.89 against 0.98 J), and with a new process per cell, 0.99 J, it ties. Full life cycle of a cell (boots, passes its test, shuts down), with no work: minimal 24 mJ, with network 28, with files 32, with network and files 38. 5 repetitions of 400 cells per arm, 0 failures. Coherent with kz-f1 f0: there the energy reading was taken from dom0 under Xen and failed; here the kernel is native, without a hypervisor (KVM).

limit  ZBook (one laptop), native kernel, synthetic load (a hash in a loop). This is chip energy (RAPL: processor and memory), not the wall plug or the electricity bill; it does not cover network, disk, fans or power supply. The saving against Firecracker with Linux comes from replacing Linux with the cell, not from the cell against Firecracker. No money figures. Still to repeat on a server and to validate against an external meter. Evidence recorded with hash and timestamp.

kz-c1 · bare metal  PASS

kz-c1, cell launcher with direct libxl

A cycle of about 38 ms (create 26 + life 2 + destroy 9) against about 94 ms plus harness with xl; 64/64 cells ok in each configuration and a ceiling of 50 to 60 cells/s per node from P=4, because creation serializes in the toolstack.

limit  Measured on the ZBook, before later generations: not the current cycle nor a latency headline. No confidence intervals.

Stage 6 · 2 labs

The cloud, kept in its place.

Nested runs orient and nothing else: the performance and isolation verdicts are the ones measured on metal.

kz-gc1 · cloud (nested)  PARTIAL

kz-gc1: the KVM cell in a cloud VM (nested KVM)

Cell ready from zero in 6.9 ms against 107 ms for untuned Firecracker + Linux. On ~100 ms jobs: 2.11x at the worst P, against untuned Linux the edge shows; with the same cell on Firecracker it drops to 1.04x (it orients). A burst of 20,000 cells without failures and 13,053 live cells at 1.86 MB each.

limit  Exploratory · in the cloud (nested) · it orients, it is not a verdict; the performance and isolation verdicts are the ZBook's. One round per point; the 13,053 was cut by time, not a RAM ceiling.

kz-az1 · bare metal + nested cloud  PARTIAL

kz-az1: duel on metal and runs on nested AWS

On metal, the 2x hypothesis on ~100 ms jobs holds only against untuned Linux (2.01x, borderline) and is falsified against tuned Linux (0.92x to 1.03x, technical tie). With the best measured configuration after the trim, kz-k2 ready 1.51 versus 2.32 ms for Unikraft and 1.03 versus 1.49 MiB per cell; before the trim Unikraft arrived first (2.25 versus 4.83 ms).

limit  One machine, one round per point. At P=15 Unikraft stays ahead on rate in 3 of 4 cells; with defaults Unikraft leads on ready time. Nested AWS orients, it is not a verdict. Snapshot-restore stays out of the rankings. The trim mechanism is not published.

Stage 7 · 11 labs

The reference code.

The components themselves, as standalone scripts that print their checks and exit 0 only if every vector matched its expected verdict. Mostly logic plane.

impl · gate · logic  PASS

Gate (admission gate for the Job)

42/42 checks without failures (rc=0); reservation without failures and indexed ledger 8/8. Across the 31 extended adversarial vectors (gate, broker, chain): 19 DENY, 9 declared GAPs, 3 LIMIT by design.

limit  Logic plane. The 9 GAPs are published without hiding them; C5 was the only high-severity one. The Gate's metal run is lab71.

impl · broker · logic  PASS

Broker (mediated I/O with CID)

71/71 checks (68 vectors, purity, derivation and signatures), 0 failures, rc=0. Covers the peer key pin: another certificate from the same CA with another key is denied.

limit  Logic plane. Declared GAPs remain (B2, B5, B6 by design, and B7, state rollback, which depends on dom0 durability). Not an isolation verdict.

impl · chain · logic  PASS

Chain (receipt and lineage chain)

37/37 scenarios without failures, rc=0. Signature coverage now attests the full set of fields (moving any gives BAD_SIGNATURE) and lineage requires a real destruction receipt.

limit  Logic plane. Starting findings that are published: lineage was vacuous because of a public constant and signature coverage was partial; both were fixed in the real code.

impl · observer · logic  PASS

Observer (countersignature and destruction receipt)

14/14 checks without failures and the destruction-receipt chain with rc=0. Removing or replacing the countersignature (6+3 variants) is denied with MISSING_COSIGNATURE.

limit  Logic plane. The observer countersigned without an observed commit fact (C5); it was addressed in the real code with a separate store. Independence in production is not claimed here.

impl · ring · logic  PASS

Ring (framing of the request and response channel)

62/62 vectors without failures, rc=0. The receiver validates only a snapshot of the page, and large artifacts travel in chunks.

limit  Logic plane (loopback on one machine, no real Xen channel). Python's fence is best-effort. Not an isolation or performance verdict.

impl · transport · logic  PARTIAL

Transport (mTLS for the control plane)

In logic: mtls 22/22, serve 21/21, integration 13/13 (real handshake, 0 SKIP), tier2 6/6, ca 18/18, identity 8/8. On metal, once: 12/12 suites with a real TLS 1.3 handshake. Another certificate from the same CA with another key is denied.

limit  Open: running it between distinct OS principals and between separate Xen domains. On metal both ends had the same UID; the current inter-domain channel is a relay without mTLS.

impl · crypto · logic  PASS

kzsig: Ed25519 asymmetric signing

15/15 checks, 0 failures: with public keys only the verifier checks but cannot forge; 300 random signatures, a reused signature and a signature made with the public key all fail; a dom0 signing with its own key is rejected.

limit  Pure logic. Authentically distributing the three public keys remains open; dom0 can still drop or delay; the private keys are not protected by this layer.

impl · kaf-store · logic  PASS

kaf-store: per-process isolated storage (closing C5)

22/22 checks (rc=0): asking for a signature over bytes that were never stored, storing garbage and declaring another hash, or writing straight to the medium, is denied; no private key reachable from the broker process.

limit  It closes only the compromised-broker-process case. Against dom0 with root, process isolation is not enough: that is a hardware residual. Not tested on metal in this directory.

impl · alicebob · logic  PARTIAL

alicebob: notary and key registry for console agents

Two adversarial reviews found and fixed about twenty defects (rewritable or truncatable chain, per-Job owner, injection in the audit query, among others); the Go tests for chain, signature, server and store pass.

limit  Finding NOT fixed in code: if daemon and agent run under the same UID there is no agent/human boundary. Linear log, not Merkle; no performance figure.

impl · rump-xen · bare metal  PASS

KZ-Rump cell on Xen PVH (and the same image on KVM)

Honest sequence: T15a link PASS; T15b the console FAILED; T15c KZ-RUMP-OK; T16a fs-only BLOCKED (UNCLEAR); T16b fs-only PASS 3/3; T16c net-only PASS, 8-thread contention without aborts and a deliberate deadlock that dies observably at 30 s. A single ELF boots on Xen and on KVM.

limit  n=1, one run per test. A first audit found an insufficient watchdog and it was fixed. Current boot figures are those of the trim, not of these reports.

mtls-metal · bare metal  PARTIAL

mtls-metal: control-plane mTLS on the physical dom0

12/12 suites green on metal with a real TLS 1.3 handshake (a certificate from the same CA with another key is denied). The harness between principals with distinct UIDs never ran.

limit  Both ends with the same UID and pure-Python TLS: not separation between OS principals nor between Xen domains. n=1.

Mémoire

La mémoire, dans tout cloud.

Whoever controls the host can read the memory of what runs on it: the cloud provider, or an attacker who gets root on the host. Every cloud has this (AWS, GCP, any other) unless you use confidential computing. It is not always said in these words, but that is how it works. We say it plainly so nobody takes it for a Kawésqar problem.

mémoire
Cloud sansinformatique confidentielleAWS, GCP, tout cloudvotre VM / votre chargemémoire en clairsecrethyperviseur / hôterootlitLe fournisseur, ou un attaquant avecle root sur l’hôte, lit la mémoire.Serveur propre,Xen dom0measured on metal, n=1la cellulemémoire en clairsecretdom0rootlitRoot in dom0 reads the cell memory.Mesuré : 3 sur 3, un secret néà l’intérieur de la cellule.InformatiqueconfidentielleTDX · SEV-SNPvotre VM / votre chargemémoire chiffrée▒▓░ x9#f 7a▒ ░▓hyperviseur / hôterootblockedThe host sees only ciphertext.The read is blocked.Même modèle de sécurité dans les trois ;la seule chose qui change est le chiffrement de la mémoire.Cloud sansinformatique confidentielleAWS, GCP, tout cloudvotre VM / votre chargemémoire en clairsecrethyperviseur / hôterootlitLe fournisseur, ou un attaquant avecle root sur l’hôte, lit la mémoire.Serveur propre,Xen dom0measured on metal, n=1la cellulemémoire en clairsecretdom0rootlitRoot in dom0 reads the cell memory.Mesuré : 3 sur 3, un secret néà l’intérieur de la cellule.InformatiqueconfidentielleTDX · SEV-SNPvotre VM / votre chargemémoire chiffrée▒▓░ x9#f 7a▒ ░▓hyperviseur / hôterootblockedThe host sees only ciphertext.The read is blocked.Même modèle de sécurité dans les trois ;la seule chose qui change est le chiffrement de la mémoire.
memory   Who can read what. Orange arrow: a read by the host. Only the third stack, with hardware encryption, stops it.
plateforme

It belongs to the platform, not to Kawésqar

Any workload on any host without confidential computing is readable from the host. It is a property of the platform, the same for every product that runs on it.

we tried it

Nous l’avons testé dans le cloud confidentiel

TDX sur GCP et SEV-SNP sur AWS. La mémoire reste chiffrée, mais la cellule ne peut pas s’y exécuter (il n’y a pas de KVM) et une VM confidentielle par tâche met de 50 à 58 s à démarrer.

what we ask

So the closure is validated on physical hardware

The cell closure is validated on physical hardware (TDX or SEV-SNP), which is what we are asking for.

Comment nous l’utilisons

Le critique reste chez soi, le reste va vite.

Nous ne nous opposons pas au passage au cloud. Nous utilisons la vitesse de Kawésqar dans le cloud pour les charges moins critiques, ou là où le risque est tolérable, et nous ne gardons chez soi que ce qui le mérite : le souverain, le sur site ou le local. Le plus critique, comme les clés d’une banque, reste sur votre propre matériel avec chiffrement complet de la mémoire (TDX / SEV-SNP).

échelle
Tolérant au risquecharge→ Cloudvitesse KawésqarSensiblecharge→ Sur sitesouverain, chez soiLes clés d’une banque→ Sur site avec chiffrement complet de la mémoireTDX / SEV-SNPvalidé sur du matériel physiquemoins critiqueplus critiqueTolérant au risquecharge→ Cloudvitesse KawésqarSensiblecharge→ Sur sitesouverain, chez soiLes clés d’une banque→ Sur site avec chiffrement complet de la mémoireTDX / SEV-SNPvalidé sur du matériel physiquemoins critiqueplus critique
ladder   Where each kind of workload goes. Only what deserves it stays at home.
same contract

Same contract, same Gate

On both sides the same contract, the same Gate and the same security model apply. Only where the workload runs changes.

measured

What is measured today

Closing the cell with full memory encryption is validated on physical hardware (TDX or SEV-SNP), which is what we ask Intel and AMD for. Today the residual is measured and declared (lab54b, 3 of 3), not closed.

not yet

What does not exist yet

Federation between on-prem and cloud is a design direction, still without code.

Keep reading

Evidence, then the ask.

The evidence page puts the headline figures in context and names the gaps still open. If you want to reproduce something here, tell us plainly.

the rule

Not a certificate, not a claim of regulatory compliance: a dated, bounded record of what ran.

Agent Fabric · lab notebook

More than 70 labs. The ones that failed too.

This is the lab notebook of the Agent Fabric: every experiment we can publish, with the plane it ran on, the verdict, the figure and the limit. Failures and refutations sit next to the passes, because a lab that only shows what worked is showing you a pitch, not evidence.

0+
experiments in the lab, and counting
0
shown here, each with its verdict, plane and limit
0
tested: PASS, HELD or mitigated by policy
0
published as FAIL or REFUTED, not buried

patent pending  Patent applications filed with INAPI (Chile); in process.

Verdict board

Every lab, by plane.

One chip per lab. Green is a tested pass, amber is partial, unclear, open or a measured residual, orange is a recorded failure or refutation. The plane matters as much as the colour: our results run nested (inside a VM, which validates neither isolation nor performance) or on bare metal, the only place those claims can close. Metal means one laptop workstation, n=1, unless a card says otherwise.

Nested VM · 16 labs

12 tested · 2 partial or open · 2 failed

Nested cloud · 1 lab

0 tested · 1 partial or open · 0 failed

Lab72 also carries a TLA+ model checked against traces, counted here under bare metal. In the cloud we show only what orients; it never closes a verdict.

What failed

What failed, said plainly.

A lab that only publishes its passes is selling. These are the places where the data said no, or said less than we hoped, and what we changed after.

nested VM  FAIL

lab14 · The rump guest printed nothing

FAIL on nested Xen. The guest reached long mode and ran rump_init, yet two console faults kept the marker from ever appearing. The failure dictated the order of fixes that lab16 closed.

nested VM  FAIL

lab15 · The observer did not fail closed

F9 FAIL: 38,577 mutations produced 4,520 crashes, 85 false countersignatures (only with a compromised stream) and 56 hangs. lab29 later reproduced the failure on metal, hardened the observer and reached 0 crashes in 3,000 variants.

bare metal  FAIL

lab25 · The disposable cell leaked on disk

RAM held (0 hits) but a reused volume let cell B read cell A's pattern: 6,882,854 hits. Zeroing the whole volume on create brought it to 0, a fix verified in the lab before it reached the product.

bare metal  REFUTED

lab64 · Gate plus order cost more than 1 ms

We pre-registered under 1 ms. The measured median is 2.735 ms (P=1, n=200). The hypothesis is refuted and the number is the one we publish.

bare metal  REFUTED

lab70 · The faster Gate still missed its target

H1 asked for under 1.5 ms of wall time. The full arm adds +3.210 ms (95% CI 3.074 to 3.378), and +3.216 ms again with the performance governor. Refuted twice.

bare metal  FAIL

KZ-F1 phase 0 · C1

FAIL against the pre-registered criterion: the register path returns an I/O error under Xen 4.20, before and after load. An alternative interface answered and followed the load, but it is not validated (n=1, no external meter) and two registered checks of it missed their criteria. It affects only energy measurement, and no energy figure is cited.

bare metal  REFUTED

KZ-F1 phase 2 · H7b

We pre-registered at most 1,800 lines of code for the production path. The count is 7,840, and above 2,000 counts as refuted. The hypothesis is refuted and the number is published as measured, without redefining the count.

bare metal  PARTIAL

lab20a · An observation we must not oversell

One DMA denial was seen once and could not be reproduced; under strict mode the positive control was itself denied. It is an observation, not a result, and we do not cite it as “strict denies unmapped transfers”.

logic  PARTIAL

C5 · The close was not won

A broker that forges the storage confirmation crossed the gate in the first logic wave (23/23 behaviors reproduced). Later work added a separate store; the card for kaf-store shows what that closes and what it does not.

bare metal  TIE

lab27 · “Rump exposes less” was false

The hypothesis was falsified: rump keeps 2 live event channels against 0 for minimal Linux, and they tie on grants and xenstore nodes. We concluded rump did not justify deep investment on that ground alone.

bare metal + nested cloud  PARTIAL

kz-az1 · The 2x claim only half holds

On 100 ms jobs the 2x edge holds against untuned Linux (2.01x, borderline) and is falsified against tuned Linux (0.92x to 1.03x). Before the trim, Unikraft was ready sooner. We lead with the part that did not hold.

Timeline

Six stages, one thread.

The notebook follows the order we worked in: sandbox first, then Xen nested in a VM, then the physical machine, then attacks on our own design, then the fast cells, and last the cloud. Bar height is the number of labs shown for each stage; the reference code (11 more cards) sits apart at the end.

timeline
2SANDBOX14NESTED XEN27BARE METAL16RED TEAM15KZ CELLS2CLOUDSANDBOX2NESTED XEN14BARE METAL27RED TEAM16KZ CELLS15CLOUD2
timeline   Labs shown per stage. A nested or cloud result never counts as an isolation or performance verdict: only bare metal can close those.
Stage 1 · 16 labs

Nested Xen, the first answers.

Inside a VM (Xen under QEMU/KVM). These runs show the architecture working as logic on real domains. They are never an isolation or performance verdict, and each card says so.

Each card: lab, plane and verdict, then what was found, then the limit that comes with it. Figures are the ones the lab recorded; where a later campaign superseded them, the card says so.

lab03 · nested VM  PARTIAL

NetBSD rump kernels: closing the unknowns

The real IP stack uses 52 hypercalls (base 46) and 28 of 46 are invoked at startup; in the official order the real definition wins in 3 of 3 duplicates; the sysctl warning is benign.

limit  One nested guest; partially reproducible build (186/190 identical); no conclusion on performance or isolation.

lab04 · nested VM  PASS

Build non-determinism, guest disk, skill tests

The difference was only embedded source paths: two MKREPRO builds come out identical, 190/190. It corrects the lab03 hypothesis; skill tests 12/12 after rerunning one.

limit  n=2 builds per condition; the exact cause is strong inference, not fact; one case failed on the first pass.

lab05 · nested VM  PASS

Nested Xen and a minimal PVH cell

The cell prints its marker and exits on its own; 5 cycles and 1 destroy during boot leave 0 domains and 0 residual xenstore keys.

limit  All nested and without IOMMU: not valid for performance or isolation; not a benchmark.

lab06 · nested VM  PARTIAL

Skill validation and nested Xen cell (A to D)

A: 10/12 PASS; B: 7 PASS and 1 FAIL (one case invented a format); C: 5/5, 30 cycles with 0 residue and 0 failures; D: 5/5. The rump hypercall count is 46 on both sides.

limit  Nested, no IOMMU; three cases cannot be reproduced because their prompt and criterion are not on disk.

lab07 · nested VM  PASS

S4.0 xenstore probe between two cells

Without an ACL the cross read returns EACCES (30/30 in the control); the owning cell grants access and the other reads the nonce and gets 3 events. dom0 is not required.

limit  Nested, N=1 per arm, a single xenstored variant; no quotas or concurrency.

lab07b · nested VM  PASS

S4.2 gntalloc+gntdev+evtchn ring between two cells

Bidirectional ring with identical hashes at both ends, no dom0 privilege. The first run FAILED because of our own faulty protocol and was fixed.

limit  Nested, n=1 PASS; N≥30 pending; it was not shown that a third domain cannot map the grant.

lab08 · nested VM  PASS

Placeholder domains for control, broker and observer

30 cycles (90 domains): 0 XenStore residue and 90/90 exit code 0; a test Cell was denied in 15/15 operations on the three role branches.

limit  Nested, one run, placeholder domains with no real logic; not an isolation verdict.

lab09 · nested VM  PASS

Role logic inside an initramfs

The three suites pass inside a 256 MB PVH domain (17/17, 31/31, 9/9) with ~167-172 MB free and 0 out-of-memory signals.

limit  Nested, n=1; functional viability and memory headroom only.

lab10 · nested VM  PASS

Broker in one domain and a Cell over the ring

5/5: a 9000 B artifact with identical hash both ways; foreign or forged CID, replay and forced destination denied; 0 processes of the pair in dom0.

limit  Nested, n=1; no conclusion on performance or resistance on Xen bare metal.

lab11 · nested VM  PASS

Observer that countersigns and issues a destruction receipt

4/4: countersignature only while the Cell was alive during a real STORE; destruction receipt only after destroying it; 17/17 attempts to read, alter or delete the stream denied.

limit  Nested, n=1; does not cover a hostile Cell with kernel capabilities.

lab12 · nested VM  PASS

Chain A, receipt, B across real domains

All 4 conditions hold: B does not exist before the order, without a countersignature the Gate denies, a Cell trying to sign the order is denied (18 times) and the hash is identical end to end.

limit  PASS bounded to architecture: nested, n=1, first-generation symmetric signature; the first attempt failed on packaging.

lab13 · nested VM  PASS

B1.8 Falsifications only the substrate can refute

All 3 falsifications PASS: CIDs from the previous epoch give STALE_EPOCH after recreating the broker; a third domain cannot map the grant (128 refs denied); a kill between STAGED and COMMITTED gives NOT_COMMITTED. Finding: epoch and nonces must be durable outside the domain.

limit  Nested, n=1, HMAC v0 signature; the kill instant is forced; dom0 remains in the trust base.

lab14 · nested VM  FAIL

T15b First boot of rump as a PVH guest under Xen

Instructive FAIL: the guest entered via PVH, reached long mode and ran rump_init, but did not print the marker. Two console faults (unprogrammed hypercalls and a buffer removed by the compiler) dictated the fixes that lab16 closed.

limit  Nested, n=1; it stayed unobserved whether rump_init returned 0.

lab15 · nested VM  FAIL

S9.2 Observer falsifications 8 and 9

F8 PASS: the observer survives a cell crash and the Gate does not use its self-report. F9 FAIL: with 38,577 mutations there were 4,520 crashes, 85 false countersignatures (only with a compromised stream) and 56 hangs. A crash signs nothing: it is an availability failure.

limit  Nested, n=1, HMAC v0; the lie is a fixed script; dom0 remains in the trust base.

lab16 · nested VM  PASS

T15c Closing the rump axis: the rump guest prints KZ-RUMP-OK under Xen

With the two console fixes, the rump guest prints the marker, rump_init returns 0 and it shuts down by hypercall with code 0. It closes the bit lab14 could not read.

limit  Nested, n=1: init and console only; no performance or hard isolation.

lab17 · nested VM  PASS

T16 Adversarial guest: A_effective contained in A_delegated

289 attempts to map foreign grants and 289 to bind event channels gave 0 successes, 0 foreign pages read, 0 self-grants. What Xen contains is separated from what Kawésqar contains (replay, STALE_EPOCH, NOT_COMMITTED).

limit  Nested, n=1; the guarantee assumes a correct Xen; grant and channel denials are Xen's, not Kawésqar's.

Stage 2 · 12 labs

The first campaign on metal.

On the physical control domain of one laptop workstation, n=1. Here the first honest failures appear: a covert channel we could not measure and a disk that leaked.

lab18 · bare metal  PASS

Physical baseline and gate for the lab plane

Inventory captured and preflight 11/11. The CPUs in this Xeon are not affected by L1TF, MDS or TAA, so the side-channel lab was re-scoped.

limit  ZBook, n=1, a single day; this is inventory and preparation, not an isolation or performance result.

lab19 · bare metal  PASS

Re-attesting the architecture core on metal

7/7 core verdicts PASS on the physical control domain. Minimal cell cycle: p50 0.95 s, p99 1.09 s, 200/200 without failure, real residue 0.

limit  ZBook, n=1, control domain capped at 4 vCPU: correctness, not performance; the chain on real domains is still HMAC.

lab20a · bare metal  PARTIAL

VT-d enforcement on bare metal with a named DMA device

VT-d configuration was attested and a denial was seen once, but it is not reproducible. Under strict mode the instrument ended UNCLEAR: the positive control was denied.

limit  ZBook, n=1. It must not be cited as “strict denies unmapped transfers”: it is an observation, not a tested result.

lab21 · bare metal  PASS

Baseline latency of a cell on bare metal

200/200 per variant, 0 residue: create with xl 184 ms (p50) and kernel plus init boot 607 ms (p50, p99 790 ms). A cell with work adds a bounded overhead.

limit  First-generation figures (Linux kernel and xl), long before today's cells: do not compare them.

lab22 · bare metal  PASS

Scale and load

0 Gate failures in 256 runs under load; 2,775 cycles completed at 100%; creation steady around 237 ms from N=1 to N=414; 0 residual cells.

limit  This is the Gate logic under scheduling pressure, not real cell transitions at scale; the resource ceiling was never stressed.

lab23 · bare metal  UNCLEAR

Contention covert channel

The positive control did not cross the threshold: best case 9.64 bps with 95% CI [-33.06 ; 52.31], indistinguishable from noise. Neither presence nor absence of a channel is claimed.

limit  Instrument not validated; recorded as an unmeasured channel, never as negligible.

lab23b · bare metal  MITIGATED

Covert channel: closed by policy

Risk closed by a placement policy (do not co-schedule mutually distrusting cells on sibling threads), not by measurement: the strongest SMT channel does not exist by construction.

limit  It is not “we showed there is no channel”: shared-cache and cross-core memory channels remain as residual risk.

lab24 · bare metal  OPEN

XSAs that really apply

Three advisories patched, the speculative ones inapplicable on this silicon or mitigated; one applicable advisory remains unpatched (a liveness DoS in the xenstore daemon). None is an escalation or cross-domain leak.

limit  State tied to a package version; it must be rerun when the build changes.

lab25 · bare metal  FAIL

Disposable cell: RAM and disk residue

RAM PASS: 0 hits with forced overlap of ≥548 MiB. Disk FAIL: cell B sees A's pattern (6,882,854 hits); zeroing the whole volume on create brings it to 0 hits.

limit  The mitigation was verified in the lab but was not yet in the product; device DMA residue out of scope.

lab26 · bare metal  PARTIAL

Closing the bare metal campaign

Architecture: bounded PASS, conditional close. No amplification of authority or evidence on real Xen (7/7) and under load; disposable cell falsified on disk and PASS on RAM.

limit  No verified physical isolation is claimed, nor unconditional disposable cell, nor the full thesis; covert channel UNCLEAR.

lab27 · bare metal  TIE

Boundary shootout: KZ-Rump against minimal Linux

The hypothesis “rump exposes less” was falsified: rump keeps 2 live event channels (minimal Linux 0); they tie on grants (0) and on 38 xenstore nodes. Finding: an adapter abort gave reason code 0, same as a clean shutdown.

limit  Observed surface, not reachable; both runtimes are near-stubs; one machine, one run.

lab28 · logic + bare metal  PASS

Delegation tree n>1 and non-amplification by induction

Real Gate: 1,958 edges decided, 0 amplifications; real cells at the exact envelope edge (768 of 768 MiB); a grandchild asking for more than its parent is denied with AUTHORITY_NOT_SUBSET; 5 mutants detected.

limit  The parentage tree is modeled (Xen does not know it); depth above 2, fan-in and fuzz are logic. The lab found real Gate defects, since fixed.

Stage 3 · 15 labs

Pushing the limits.

Still on metal: observer fuzzing, the signed chain inside real cells, saturation to 412 live cells, strict IOMMU and observed hypercalls.

lab29 · bare metal  PASS

Falsification 9: observer surface

The first run FALSIFIED the property: 47 crashes and 208 violations through holes reachable via JSON. After hardening the observer: 0 crashes and 0 violations across 3,000 variants, 3 mutants detected, 0 residue.

limit  PASS bounded to the wire (decoded JSON stream); in-process violations remain, declared as a GAP. n=1.

lab30 · bare metal  PASS

Falsification 10: transactional STORE under kill

At no stage was a partial COMMITTED deliverable left (7 real kill -9); replay after restart always gave STALE_EPOCH; 3 mutants detected; real cells destroyed.

limit  Journal persistence is modeled by the harness: it tests the contract's atomicity, not the production persistent CAS (declared GAP).

lab31 · bare metal  PASS

Adversarial Gate under load

0 acceptances out of 6,870 concurrent adversarial attempts; 1,194 of 1,194 legitimate ones cross under the same load; the naive race over-reserves 6 of 6 (the instrument sees the race); 0 residue.

limit  The ledger lock is in-process; production persistent CAS is out. A finite corpus, not a general proof of absence.

lab32 · bare metal  PASS

Stability soak of the create-run-destroy cycle

400 of 400 iterations: zero drift (Kendall tau 0.02 to 0.08), residue 0, Xen free memory 55,883 to 55,883 MiB, dom0 stable.

limit  Bounded window (not months), one minimal cell, one machine, n=1.

lab33 · bare metal  PASS

Saturation at maximum scale and throughput

Peak of 412 simultaneously live cells (52,736 of 53,835 MiB), 0 amplification and 0 residue across 14 levels; at most 14.78 cells/s creation and 8.98 cells/s full cycle.

limit  n=1. At the peak, destruction reaches p99 32 s because the toolstack serializes; at scale only a sample carries a console marker.

lab34 · logic + bare metal  PASS

Resource fan-out (Falsification #4)

In logic: 4,681 nodes and 7,608 edges with 0 amplifications; on metal, with 128 MiB cells, every step completes, 0 amplifications and per-cell overhead of 1.5 to 2.5 MiB.

limit  The first two physical attempts were UNCLEAR (cause found in lab40); the close is the third. n=1.

lab35 · bare metal  PASS

Minimal observed boundary of the cell under load

Full classification of 18 interfaces and 49 hypercalls; load did not break the boundary, 0 residue in all 4 scenarios. Xen, not the runtime, bounds isolation.

limit  Characterization, not an isolation verdict. Trace observation stayed UNKNOWN on all 49 rows for lack of a decoder (lab44 closed it).

lab38 · bare metal  PASS

End-to-end Ed25519 cryptographic chain inside cells

The chain of cell A, countersignature, Gate, cell B and destruction receipt runs on metal; 4 mutants (dom0 forgery, altered receipt, B without evidence, replay) and 7 negative vectors rejected without creating B.

limit  Verifying the signature before storing is modeled by the harness, not by the reference Gate. n=1.

lab39 · bare metal  PASS

Smoke and invariant regression under strict IOMMU

All 4 groups pass: 60/60 iterations, 0 accepted out of 3,444 adversarial attempts with 600/600 legitimate admitted, and zeroing verified; create rises about 12% (some 23 ms) under strict.

limit  Bounded regression, not exhaustive re-testing; the clean A/B in the same boot is pending. n=1.

lab40 · bare metal  PASS

Boot memory floor of a PVH cell

Measured floor of 96 MiB: with 64 MiB the cell does not boot, with 96 and 128 MiB it does. Closes the root cause of the lab34 failure and falsifies the harness race hypothesis.

limit  Coarse floor (n=3 per point), valid for this lab kernel and initrd; the exact value between 65 and 96 MiB was not refined.

lab41 · bare metal  PASS

Audit of dom0 grants over cell pages

A diskless cell, idle or under load, has exactly 1 active grant mapped by dom0 (the console ring), 0 data pages; with a disk 67 refs appear (positive control).

limit  The first run was UNCLEAR because of a blind positive control and was fixed. That they enter dom0's IOMMU is inference from the source. n=1.

lab42 · bare metal  PASS

Entropy boundary inside the cell

14 of 14 distinct keys; in no boot was a key generated while the random generator was not ready, even with RDRAND off (median 1.47 s versus 0.14 s).

limit  It shows availability and non-determinism, not entropy quality. The no-RDRAND case rests on reading the source. n=1.

lab43 · bare metal  PASS

Cryptographic chain in a cell with n greater than 1

The chain and the fan-out cross end to end and every mutant is rejected with its exact reason; the fan-out fills the exact envelope (2 by 128 MiB). Finding: the pure gate does not consume the prev link of destruction receipts.

limit  Signature verification and the prev link are modeled in the harness (production gap declared). n=1.

lab44 · bare metal  PASS

Hypercalls observed through a real trace

In all 4 scenarios what is observed is a subset of the declared set of 10 hypercalls; the rump cell leaves a minimal footprint (5 distinct numbers) and Linux adds 4 more. Negative control: 0 calls with the cell paused.

limit  The hypercall number is observed, not the sub-op or the arguments. n=1.

lab45 · bare metal  PASS

Soak under strict IOMMU and structure leaks

400 of 400 cycles without leaks: active grants flat at 2, maptrack flat at 4 and Xen free memory constant (55,883 MiB before and after).

limit  Free memory is read in MiB (detectable minimum near 80 KiB per cycle); active grants are a partial proxy. n=1.

Stage 4 · 16 labs

Attack, then verify outside dom0.

The decision moves out of the control domain, then we attack our own design, measure what dom0 can still read and compare against Linux.

lab46 · logic  PASS

Denial conformance and live agent against the real channel

12 of 12 checks and the independent oracle PASS, each denial with its exact reason; 0 escalations crossed with the 8 language models tested, including the 3 that did attack through the real channel.

limit  Logic level, not metal nor an agent inside the cell. A later campaign corrected the lab's own overclaims (mutants that mutated data, not code).

lab47 · bare metal  PASS

Verifier outside dom0 (chain A to B)

The positive control crosses and all 11 mutants are denied with their exact reason; creations = admissions = verifier PASS (1 = 1 = 1); dom0 forgeries and replays die by signature. Re-attested on metal: 3 of 3 runs PASS.

limit  n=1, one host. dom0 still creates domains and can read the RAM of a live domain (hardware residual). The first physical run was UNCLEAR because of an oracle bug, fixed.

lab48 · bare metal  PASS

Hardened gate (gate_verified) against the harness

With the harness switched off, the old gate ACCEPTS and creates the child in the 4 main mutants; the new one DENIES them with its reason. The properties move from modeled in the harness to enforced by the gate.

limit  n=1. dom0 still generated the private keys in this run (lab47 attacks that); a lying tap is not covered.

lab49-c5 · logic + bare metal  PASS

Closing C5: compromised broker

The old observer accepts both forgeries; kaf-store denies both and the byte witness denies the first. On metal, a compromised broker that signed without backing the bytes leaves the medium empty and does not cross.

limit  The witness lets through a false COMMITTED over bytes honestly in staging. This does not settle the later red-team marker, which records the C5 close as not won. n=1.

lab50-dma · bare metal  PASS

DMA confinement of an assigned device

dom0 reads the USB stick before assignment, loses the device on assignment, and the domU reads the same number of sectors with 0 VT-d faults in strict mode. It measures confinement, not adversarial isolation.

limit  Partial: a malicious driver attempting DMA into another domain's memory was not tested. It ran in HVM because PVH does not support passthrough. Do not cite as closed DMA isolation.

lab51-netless · bare metal  PASS

lab51-netless · cell with a kernel without a network stack

A PVH cell with a kernel built without a network stack boots and runs: /proc/net, /sys/class/net and iplink ABSENT and the load closes with sum=500500, with the console ready in under 1 s of kernel boot.

limit  A fact about how a kernel was built, not an isolation or performance verdict; the evidence is the console and the configuration. n=1.

lab52-smt · bare metal  PASS

lab52-smt · placement against the SMT channel

A neighbor on the sibling thread slows the victim 4.98 to 5.94 times; on another core it drops to 1.07 to 1.13 times, near baseline. Placement suppresses the SMT channel (3 runs).

limit  It measures throughput interference, not a bit rate or key extraction. The ~1.08x residual is the cross-core channel (L3). n=1.

lab53-llc · bare metal  PARTIAL

lab53-llc · cross-core channel through the L3

The cross-core channel exists and is small: 1.03 to 1.14 times interference (average ~1.08x), an order of magnitude below the SMT channel. A residual measured and bounded, not closed.

limit  Without Intel CAT/RDT the L3 cannot be partitioned on this CPU. Small, noisy signal (3 runs); it measures throughput. n=1.

lab54-meminspect · bare metal  CONFIRMED

lab54-meminspect · the control domain reads a live cell

Positive control: dom0 mapped 10,076 frames of the live cell and found the secret without the cell's cooperation (2 runs). The signature authenticates, but gives no secrecy against dom0.

limit  It measures the exposure, it does not close it. Only hardware-encrypted memory (SEV-SNP/TDX) closes it; this CPU lacks it. n=1.

lab54b · bare metal  CONFIRMED

lab54b-meminspect-incell · secret born inside the cell

3 of 3 runs with 3 distinct secrets: 1 candidate per run, hash verified, 977 contiguous frames. The secrecy residual against dom0 is now measured.

limit  It does not close the residual: that needs hardware-encrypted memory (TDX/SEV-SNP), which this machine lacks, and that closing test has not been run. The same residual exists in every cloud, AWS and GCP included, unless confidential computing is on: whoever controls the host, the provider or an attacker who gets root on it, can read the memory of what runs on it. It is a property of the platform, not of Kawésqar. We did try confidential VMs in both clouds (TDX on GCP, SEV-SNP on AWS): memory is encrypted, but the cell cannot run inside (no KVM) and a confidential VM per job takes 50 to 58 s to come up. n=1.

lab55 · bare metal  PASS

lab55-composition · the five properties together

5 of 5 runs PASS: the hardened gate decides outside dom0, with no private key in dom0; forged or repeated orders die by signature and mutants are denied. Closes the “Composition” item.

limit  n=1, 5 repetitions (reproducibility, not statistics). dom0 still creates domains and can inspect memory (lab54 residual).

lab56 · bare metal  PARTIAL

lab56-substrate · full NetBSD or trimmed Linux

Instantiation: NetBSD PVH ~1.65 s, Linux PVH 2.06 s, NetBSD PV 3.19 s. Both substrates sustain 8 concurrent cells with near-flat scaling from N=4 to N=8.

limit  Early-stage figures, earlier than today's cells; the lab's text disagrees with its own table, so only the table is cited. n=1.

lab57-gate · bare metal  PASS

Real rump cell under the Gate

The honest run PASSES and 5 mutations give their named DENY before running; a chain of 3 real cells, 8/8 cells gated in parallel, 20/20 in stability. An “echo” cell without rump passed with a bytes-only witness and is denied once the kernel is attested.

limit  Bounded PASS, NOT containment shown: the console is a degraded medium and dom0 can read the cell's RAM (hardware residual). Small n.

lab58-overhead · bare metal  PASS

Xen and rump cell overhead against Linux

Rump cell cycle without work ~94 ms versus ~500 ms for the Linux cell (5.3x); boot 0.9 ms versus 237 ms; Xen charges +8% compute in dom0 and +11% in a Linux PVH cell versus bare metal; the rump cell sits at +4% of native.

limit  n=50 per cell, one machine. An anomaly in the rump sha phase (+15 to +20%) stays UNCLEAR; no I/O measured. Create figures predate the direct launcher.

lab59 · logic + bare metal  PARTIAL

Cell-escape red team (campaigns A to F)

On metal a compromised cell read 0 pages of a neighbor, opened 0 unauthorized channels and gained no authority; the launcher's prefix cleanup did kill cells of another job (closed in logic); native fuzz 85M executions without crashes; 15 of 16 ledger items closed in logic.

limit  What is closed is the logic plane and has not been re-measured on metal. Still open: the verifier that does not re-hash the bytes, the cross-core cache channel and dom0 reading cell RAM.

redteam-escape · logic  PARTIAL

redteam-escape · Wave 1 of logic attacks (C5, E5, E6, E8, CGAPS)

C5: the close was not won (a broker that forges the storage confirmation crosses; 23/23 behaviors reproduced); E6: prefix cleanup kills cells of another job; E8: the memory check admits 11 extra cells; E5: DENY. The failures were repaired later in lab59.

limit  Logic plane in a single process, no metal; this lab records the state before the fix.

Stage 5 · 15 labs

Cells, a signed chain and the clock.

Direct launchers, the cost of the Gate, sealed images, the minimal builder in three phases and the energy per cell. Three hypotheses were refuted and one check failed here, and all are kept.

lab61 · logic  PASS

Secure Boot negatives and adversarial map

5 of 5 cases as expected: the valid control verifies, and unsigned, foreign-key and mismatched-certificate artifacts are rejected; the artifact without SBAT is detected as such.

limit  Offline signature verification: loader rejection at real boot is described as expected, not measured; 17 adversarial dimensions remain pending.

lab62 · bare metal  HELD

kz-k2 on KVM against Firecracker

With the cell before the trim, ready from zero in 4.8 ms (kz-k2 on KVM) against 8.2 ms for Firecracker with the same cell; against Firecracker with a Linux guest the pre-registered hypothesis holds at 2.14 to 2.34x; at P=8 on the minimal cell, 251 versus 147 cells/s.

limit  n=1. With the same cell the edge is 1.04 to 1.10x on 100 ms jobs: it lives in the cell, not the VMM. On 1000 ms jobs it is lost (~1x). Against Firecracker with an untuned Linux, 3.5x measured on the same machine.

lab63 · logic  PASS

lab63 · Gate to launcher chain with Ed25519

107 negative tests pass (altered receipt, missing or foreign countersignature, broken lineage, badly signed order, replay, replay after restart) and the closure of the 5 images gives 84 of 84 checks.

limit  Tested in logic and static build, without Xen or a VM; the end-to-end Xen evidence with Ed25519 is in later labs.

lab64 · bare metal  REFUTED

lab64 · Gate and order cost over the cell cycle

The pre-registered “under 1 ms” hypothesis is refuted: Gate plus order add a median of 2.735 ms (P=1, n=200); the A/B delta in total time is +2.574 ms, and +10.815 ms at P=8.

limit  n=1, ondemand governor. At P=8 the Gate queues behind a single lock, so this is not “the Gate is slower under load”. A recorded failure, published as a trait.

lab65 · bare metal  PASS

lab65 · verifier outside dom0 over real rump cells

On metal 16 of 16 scenarios as predicted; the extension with artifact substitution gives 19 of 19 in three runs, 0 orphans; in logic 62 fidelity tests and 10 mutation tests pass.

limit  n=1. A dom0 that lies coherently gets PASS, declared a deliberate NO-CLAIM. Gate state lives in the verifier's RAM. dom0 relay channel, no mTLS. A bounded verdict, not “containment shown”.

lab66 · bare metal  PARTIAL

lab66 · kz-c2 with a sealed kernel (closing the lab64 TOCTOU)

Verifying the hash on every cycle costs +7.16 ms on a cycle of about 15 ms and is discarded; the second variant costs -0.27 ms in total_ms (15.345 to 15.079), within noise, with 9 negative tests OK.

limit  The TOCTOU attack was shown only with a stand-in for the loader, not real Xen. n=1, no SHA-NI. The full chain with Gate was measured in lab70 (+3.21 ms).

lab68 · bare metal  PASS

lab68-e2e-real · real evidence from A to cell B

10 of 10 chains of A, Gate, order, direct launcher and B on metal (2 batches of 5), oracle PASS in each; in logic 128 launcher checks and 54 preflight checks without failures.

limit  n=1. dom0 remains the trust base for the tap, and B's FETCH is a signed echo, not a real FETCH to the broker. No mTLS on that channel. The first smoke on metal was UNCLEAR and was archived.

lab70 · bare metal  REFUTED

lab70-gate-fast-fd · cost of the fast Gate over launching without a Gate

Hypothesis H1 (under 1.5 ms of wall time) is refuted: the full arm adds +3.210 ms median (95% CI 3.074 to 3.378) over the cell without Gate; repeated with the performance governor: +3.216 ms.

limit  n=1, Xen, P=1. The descriptor arm includes a durable nonce fsync that the others do not pay.

lab71 · bare metal  PASS

lab71-gate-evidence · single activation and release order

1000 of 1000 rounds with exactly one activation (the baseline without the predicate crosses 300 of 300); unsigned, foreign or repeated orders give 0 creations; 10 of 10 cycles with the release confirmed before the verdict.

limit  n=1, N up to 32, no power cut mid-write. In one run an arm gave an environment FAIL and a probe was fixed after three instrument defects: reported as they are.

lab72 · bare metal  PASS

lab72-b-evidence · evidence for Application B

After destroy, residue enumeration gave 0 in 100 of 100 cells, with destruction verified in at most 0.093 ms; persistent ledger with kill -9 and restart without violations; creating the Xen domain through the direct launcher 41 ms versus 115 ms with xl. It includes a TLA+ model with trace conformance.

limit  n=1. fs-verity and dm-verity were BLOCKED on Xen and were only tested with a KVM cell on another machine. The 41 ms is Xen domain creation, not cell latency.

kz-f1 · f0 · bare metal  PARTIAL

KZ-F1 phase 0: energy, CPUID policy and zeroed pages

C1, energy reading from dom0: FAIL against the pre-registered criterion; the MSR register path returns an I/O error under Xen 4.20. An alternative interface answered, but it is not validated (n=1, no external meter; two registered checks of it also missed their criteria) and it affects only energy measurement. C2, cell without CPUID policy: REVIEW (80 of 80 pairs with identical behavior, the dump differs). C3, zeroed pages after reusing memory: PASS (210 domains of 256 MiB, at least 51,636 MiB reused, 0 non-zero bytes).

limit  ZBook, n=1, one run per check. C1 is not a citable energy figure.

kz-f1 · f1 · bare metal  PASS

KZ-F1 phase 1: signed cycle, build time and xenstore

Full cycle with a signed order: smoke PASS (8 of 8 valid orders; 4 of 4 invalid orders rejected without creating a domain), repeated with the final front: PASS. H2, build time against kz-c2: INSIDE (paired difference of -2.1 to -3.9 ms; criterion ≤ +1.0 ms), repeated with the final front: INSIDE. H3, xenstore nodes and writes per cell: PASS twice (0 nodes and 0 writes across 200 cells). H7a: PASS. Sweep smoke (the worker dies with the cell alive): PASS, 8 of 8.

limit  ZBook, n=1: the smoke, H2 and H3 ran twice on the same machine; H7a and the sweep once. Build time is not end to end: end to end this phase was slower than kz-c2 (+7.9 to +16.2 ms).

kz-f1 · f2 · bare metal  PARTIAL

KZ-F1 phase 2: resident builder, recovery and scale

Cycle of 15.1 to 18.1 ms against 22.7 to 26.9 ms for the reference arm, Gate included. H1 INSIDE: paired improvement of -8.672 to -7.243 ms with 200 pairs per variant (criterion ≤ -4.0 ms). H6 PASS: 13 states × 100 and 200 clean endings, 0 violations. Smoke 8 of 8 with 3 negatives rejected; sweep 21 of 21; real failures 20 of 20 across the four cells. Not met: H1b INCONCLUSIVE (1.022 to 1.059 ms against a threshold of ≤ 1.0); H4 BETWEEN THRESHOLDS (entry 3.443 to 3.647 ms and 224 to 280 cells/s; criterion ≤ 3 ms and ≥ 243/s); H5 BETWEEN THRESHOLDS (1,800 cells per variant, no wall, lowest decile 74.1/s against a criterion of ≥ 78/s); H7b REFUTED (7,840 lines of code against a criterion of ≤ 1,800).

limit  ZBook, n=1, one run per test, unsealed. The cycle definitions differ from the KVM ones and from the rival's (Xen against KVM, with and without the signing Gate) and are not chained. “Between thresholds” and “inconclusive” are not passes.

energy · bare metal  PARTIAL

Energy per cell with RAPL counters, native kernel without hypervisor

Same task of about 100 ms, processor and memory energy per cell, with idle subtracted: native work 0.74 J, our cell 0.89 J, Firecracker with our same cell 0.98 J, Firecracker with Linux 2.04 J. Overhead over the native work: 0.15 J against 1.30 J. That is less than half the processor energy of a Firecracker microVM with Linux, for the same work. Against Firecracker with the same cell it is close to a tie (0.89 against 0.98 J), and with a new process per cell, 0.99 J, it ties. Full life cycle of a cell (boots, passes its test, shuts down), with no work: minimal 24 mJ, with network 28, with files 32, with network and files 38. 5 repetitions of 400 cells per arm, 0 failures. Coherent with kz-f1 f0: there the energy reading was taken from dom0 under Xen and failed; here the kernel is native, without a hypervisor (KVM).

limit  ZBook (one laptop), native kernel, synthetic load (a hash in a loop). This is chip energy (RAPL: processor and memory), not the wall plug or the electricity bill; it does not cover network, disk, fans or power supply. The saving against Firecracker with Linux comes from replacing Linux with the cell, not from the cell against Firecracker. No money figures. Still to repeat on a server and to validate against an external meter. Evidence recorded with hash and timestamp.

kz-c1 · bare metal  PASS

kz-c1, cell launcher with direct libxl

A cycle of about 38 ms (create 26 + life 2 + destroy 9) against about 94 ms plus harness with xl; 64/64 cells ok in each configuration and a ceiling of 50 to 60 cells/s per node from P=4, because creation serializes in the toolstack.

limit  Measured on the ZBook, before later generations: not the current cycle nor a latency headline. No confidence intervals.

Stage 6 · 2 labs

The cloud, kept in its place.

Nested runs orient and nothing else: the performance and isolation verdicts are the ones measured on metal.

kz-gc1 · cloud (nested)  PARTIAL

kz-gc1: the KVM cell in a cloud VM (nested KVM)

Cell ready from zero in 6.9 ms against 107 ms for untuned Firecracker + Linux. On ~100 ms jobs: 2.11x at the worst P, against untuned Linux the edge shows; with the same cell on Firecracker it drops to 1.04x (it orients). A burst of 20,000 cells without failures and 13,053 live cells at 1.86 MB each.

limit  Exploratory · in the cloud (nested) · it orients, it is not a verdict; the performance and isolation verdicts are the ZBook's. One round per point; the 13,053 was cut by time, not a RAM ceiling.

kz-az1 · bare metal + nested cloud  PARTIAL

kz-az1: duel on metal and runs on nested AWS

On metal, the 2x hypothesis on ~100 ms jobs holds only against untuned Linux (2.01x, borderline) and is falsified against tuned Linux (0.92x to 1.03x, technical tie). With the best measured configuration after the trim, kz-k2 ready 1.51 versus 2.32 ms for Unikraft and 1.03 versus 1.49 MiB per cell; before the trim Unikraft arrived first (2.25 versus 4.83 ms).

limit  One machine, one round per point. At P=15 Unikraft stays ahead on rate in 3 of 4 cells; with defaults Unikraft leads on ready time. Nested AWS orients, it is not a verdict. Snapshot-restore stays out of the rankings. The trim mechanism is not published.

Stage 7 · 11 labs

The reference code.

The components themselves, as standalone scripts that print their checks and exit 0 only if every vector matched its expected verdict. Mostly logic plane.

impl · gate · logic  PASS

Gate (admission gate for the Job)

42/42 checks without failures (rc=0); reservation without failures and indexed ledger 8/8. Across the 31 extended adversarial vectors (gate, broker, chain): 19 DENY, 9 declared GAPs, 3 LIMIT by design.

limit  Logic plane. The 9 GAPs are published without hiding them; C5 was the only high-severity one. The Gate's metal run is lab71.

impl · broker · logic  PASS

Broker (mediated I/O with CID)

71/71 checks (68 vectors, purity, derivation and signatures), 0 failures, rc=0. Covers the peer key pin: another certificate from the same CA with another key is denied.

limit  Logic plane. Declared GAPs remain (B2, B5, B6 by design, and B7, state rollback, which depends on dom0 durability). Not an isolation verdict.

impl · chain · logic  PASS

Chain (receipt and lineage chain)

37/37 scenarios without failures, rc=0. Signature coverage now attests the full set of fields (moving any gives BAD_SIGNATURE) and lineage requires a real destruction receipt.

limit  Logic plane. Starting findings that are published: lineage was vacuous because of a public constant and signature coverage was partial; both were fixed in the real code.

impl · observer · logic  PASS

Observer (countersignature and destruction receipt)

14/14 checks without failures and the destruction-receipt chain with rc=0. Removing or replacing the countersignature (6+3 variants) is denied with MISSING_COSIGNATURE.

limit  Logic plane. The observer countersigned without an observed commit fact (C5); it was addressed in the real code with a separate store. Independence in production is not claimed here.

impl · ring · logic  PASS

Ring (framing of the request and response channel)

62/62 vectors without failures, rc=0. The receiver validates only a snapshot of the page, and large artifacts travel in chunks.

limit  Logic plane (loopback on one machine, no real Xen channel). Python's fence is best-effort. Not an isolation or performance verdict.

impl · transport · logic  PARTIAL

Transport (mTLS for the control plane)

In logic: mtls 22/22, serve 21/21, integration 13/13 (real handshake, 0 SKIP), tier2 6/6, ca 18/18, identity 8/8. On metal, once: 12/12 suites with a real TLS 1.3 handshake. Another certificate from the same CA with another key is denied.

limit  Open: running it between distinct OS principals and between separate Xen domains. On metal both ends had the same UID; the current inter-domain channel is a relay without mTLS.

impl · crypto · logic  PASS

kzsig: Ed25519 asymmetric signing

15/15 checks, 0 failures: with public keys only the verifier checks but cannot forge; 300 random signatures, a reused signature and a signature made with the public key all fail; a dom0 signing with its own key is rejected.

limit  Pure logic. Authentically distributing the three public keys remains open; dom0 can still drop or delay; the private keys are not protected by this layer.

impl · kaf-store · logic  PASS

kaf-store: per-process isolated storage (closing C5)

22/22 checks (rc=0): asking for a signature over bytes that were never stored, storing garbage and declaring another hash, or writing straight to the medium, is denied; no private key reachable from the broker process.

limit  It closes only the compromised-broker-process case. Against dom0 with root, process isolation is not enough: that is a hardware residual. Not tested on metal in this directory.

impl · alicebob · logic  PARTIAL

alicebob: notary and key registry for console agents

Two adversarial reviews found and fixed about twenty defects (rewritable or truncatable chain, per-Job owner, injection in the audit query, among others); the Go tests for chain, signature, server and store pass.

limit  Finding NOT fixed in code: if daemon and agent run under the same UID there is no agent/human boundary. Linear log, not Merkle; no performance figure.

impl · rump-xen · bare metal  PASS

KZ-Rump cell on Xen PVH (and the same image on KVM)

Honest sequence: T15a link PASS; T15b the console FAILED; T15c KZ-RUMP-OK; T16a fs-only BLOCKED (UNCLEAR); T16b fs-only PASS 3/3; T16c net-only PASS, 8-thread contention without aborts and a deliberate deadlock that dies observably at 30 s. A single ELF boots on Xen and on KVM.

limit  n=1, one run per test. A first audit found an insufficient watchdog and it was fixed. Current boot figures are those of the trim, not of these reports.

mtls-metal · bare metal  PARTIAL

mtls-metal: control-plane mTLS on the physical dom0

12/12 suites green on metal with a real TLS 1.3 handshake (a certificate from the same CA with another key is denied). The harness between principals with distinct UIDs never ran.

limit  Both ends with the same UID and pure-Python TLS: not separation between OS principals nor between Xen domains. n=1.

Speicher

Der Speicher, in jeder Cloud.

Whoever controls the host can read the memory of what runs on it: the cloud provider, or an attacker who gets root on the host. Every cloud has this (AWS, GCP, any other) unless you use confidential computing. It is not always said in these words, but that is how it works. We say it plainly so nobody takes it for a Kawésqar problem.

Speicher
Cloud ohneConfidential ComputingAWS, GCP, jede CloudIhre VM / Ihre LastSpeicher im KlartextGeheimnisHypervisor / HostrootliestDer Anbieter oder ein Angreifer mitroot auf dem Host liest den Speicher.Eigener Server,Xen dom0measured on metal, n=1die ZelleSpeicher im KlartextGeheimnisdom0rootliestRoot in dom0 reads the cell memory.Gemessen: 3 von 3, ein in der Zelleentstandenes Geheimnis.ConfidentialComputingTDX · SEV-SNPIhre VM / Ihre LastSpeicher verschlüsselt▒▓░ x9#f 7a▒ ░▓Hypervisor / HostrootblockedThe host sees only ciphertext.The read is blocked.Gleiches Sicherheitsmodell in allen dreien;das Einzige, was sich ändert, ist die Speicherverschlüsselung.Cloud ohneConfidential ComputingAWS, GCP, jede CloudIhre VM / Ihre LastSpeicher im KlartextGeheimnisHypervisor / HostrootliestDer Anbieter oder ein Angreifer mitroot auf dem Host liest den Speicher.Eigener Server,Xen dom0measured on metal, n=1die ZelleSpeicher im KlartextGeheimnisdom0rootliestRoot in dom0 reads the cell memory.Gemessen: 3 von 3, ein in der Zelleentstandenes Geheimnis.ConfidentialComputingTDX · SEV-SNPIhre VM / Ihre LastSpeicher verschlüsselt▒▓░ x9#f 7a▒ ░▓Hypervisor / HostrootblockedThe host sees only ciphertext.The read is blocked.Gleiches Sicherheitsmodell in allen dreien;das Einzige, was sich ändert, ist die Speicherverschlüsselung.
memory   Who can read what. Orange arrow: a read by the host. Only the third stack, with hardware encryption, stops it.
Plattform

It belongs to the platform, not to Kawésqar

Any workload on any host without confidential computing is readable from the host. It is a property of the platform, the same for every product that runs on it.

we tried it

Wir haben es in der vertraulichen Cloud getestet

TDX auf GCP und SEV-SNP auf AWS. Der Speicher bleibt verschlüsselt, aber die Zelle kann darin nicht laufen (es gibt kein KVM), und eine vertrauliche VM pro Auftrag braucht 50 bis 58 s zum Hochfahren.

what we ask

So the closure is validated on physical hardware

The cell closure is validated on physical hardware (TDX or SEV-SNP), which is what we are asking for.

Wie wir es einsetzen

Das Kritische bleibt zu Hause, der Rest geht schnell.

Wir wenden uns nicht gegen den Weg in die Cloud. Wir nutzen die Geschwindigkeit von Kawésqar in der Cloud für weniger kritische Lasten oder dort, wo das Risiko tragbar ist, und behalten zu Hause nur, was es verdient: Souveränes, On-Prem oder Lokales. Das Kritischste, etwa die Schlüssel einer Bank, bleibt auf Ihrer eigenen Hardware mit vollständiger Speicherverschlüsselung (TDX / SEV-SNP).

Stufenleiter
RisikotoleranteLast→ CloudGeschwindigkeit von KawésqarSensibelLast→ On-Premsouverän, zu HauseDie Schlüssel einer Bank→ On-Prem mit vollständiger SpeicherverschlüsselungTDX / SEV-SNPauf physischer Hardware validiertweniger kritischkritischerRisikotoleranteLast→ CloudGeschwindigkeit von KawésqarSensibelLast→ On-Premsouverän, zu HauseDie Schlüssel einer Bank→ On-Prem mit vollständiger SpeicherverschlüsselungTDX / SEV-SNPauf physischer Hardware validiertweniger kritischkritischer
ladder   Where each kind of workload goes. Only what deserves it stays at home.
same contract

Same contract, same Gate

On both sides the same contract, the same Gate and the same security model apply. Only where the workload runs changes.

measured

What is measured today

Closing the cell with full memory encryption is validated on physical hardware (TDX or SEV-SNP), which is what we ask Intel and AMD for. Today the residual is measured and declared (lab54b, 3 of 3), not closed.

not yet

What does not exist yet

Federation between on-prem and cloud is a design direction, still without code.

Keep reading

Evidence, then the ask.

The evidence page puts the headline figures in context and names the gaps still open. If you want to reproduce something here, tell us plainly.

the rule

Not a certificate, not a claim of regulatory compliance: a dated, bounded record of what ran.

Agent Fabric · lab notebook

More than 70 labs. The ones that failed too.

This is the lab notebook of the Agent Fabric: every experiment we can publish, with the plane it ran on, the verdict, the figure and the limit. Failures and refutations sit next to the passes, because a lab that only shows what worked is showing you a pitch, not evidence.

0+
experiments in the lab, and counting
0
shown here, each with its verdict, plane and limit
0
tested: PASS, HELD or mitigated by policy
0
published as FAIL or REFUTED, not buried

patent pending  Patent applications filed with INAPI (Chile); in process.

Verdict board

Every lab, by plane.

One chip per lab. Green is a tested pass, amber is partial, unclear, open or a measured residual, orange is a recorded failure or refutation. The plane matters as much as the colour: our results run nested (inside a VM, which validates neither isolation nor performance) or on bare metal, the only place those claims can close. Metal means one laptop workstation, n=1, unless a card says otherwise.

Nested VM · 16 labs

12 tested · 2 partial or open · 2 failed

Nested cloud · 1 lab

0 tested · 1 partial or open · 0 failed

Lab72 also carries a TLA+ model checked against traces, counted here under bare metal. In the cloud we show only what orients; it never closes a verdict.

What failed

What failed, said plainly.

A lab that only publishes its passes is selling. These are the places where the data said no, or said less than we hoped, and what we changed after.

nested VM  FAIL

lab14 · The rump guest printed nothing

FAIL on nested Xen. The guest reached long mode and ran rump_init, yet two console faults kept the marker from ever appearing. The failure dictated the order of fixes that lab16 closed.

nested VM  FAIL

lab15 · The observer did not fail closed

F9 FAIL: 38,577 mutations produced 4,520 crashes, 85 false countersignatures (only with a compromised stream) and 56 hangs. lab29 later reproduced the failure on metal, hardened the observer and reached 0 crashes in 3,000 variants.

bare metal  FAIL

lab25 · The disposable cell leaked on disk

RAM held (0 hits) but a reused volume let cell B read cell A's pattern: 6,882,854 hits. Zeroing the whole volume on create brought it to 0, a fix verified in the lab before it reached the product.

bare metal  REFUTED

lab64 · Gate plus order cost more than 1 ms

We pre-registered under 1 ms. The measured median is 2.735 ms (P=1, n=200). The hypothesis is refuted and the number is the one we publish.

bare metal  REFUTED

lab70 · The faster Gate still missed its target

H1 asked for under 1.5 ms of wall time. The full arm adds +3.210 ms (95% CI 3.074 to 3.378), and +3.216 ms again with the performance governor. Refuted twice.

bare metal  FAIL

KZ-F1 phase 0 · C1

FAIL against the pre-registered criterion: the register path returns an I/O error under Xen 4.20, before and after load. An alternative interface answered and followed the load, but it is not validated (n=1, no external meter) and two registered checks of it missed their criteria. It affects only energy measurement, and no energy figure is cited.

bare metal  REFUTED

KZ-F1 phase 2 · H7b

We pre-registered at most 1,800 lines of code for the production path. The count is 7,840, and above 2,000 counts as refuted. The hypothesis is refuted and the number is published as measured, without redefining the count.

bare metal  PARTIAL

lab20a · An observation we must not oversell

One DMA denial was seen once and could not be reproduced; under strict mode the positive control was itself denied. It is an observation, not a result, and we do not cite it as “strict denies unmapped transfers”.

logic  PARTIAL

C5 · The close was not won

A broker that forges the storage confirmation crossed the gate in the first logic wave (23/23 behaviors reproduced). Later work added a separate store; the card for kaf-store shows what that closes and what it does not.

bare metal  TIE

lab27 · “Rump exposes less” was false

The hypothesis was falsified: rump keeps 2 live event channels against 0 for minimal Linux, and they tie on grants and xenstore nodes. We concluded rump did not justify deep investment on that ground alone.

bare metal + nested cloud  PARTIAL

kz-az1 · The 2x claim only half holds

On 100 ms jobs the 2x edge holds against untuned Linux (2.01x, borderline) and is falsified against tuned Linux (0.92x to 1.03x). Before the trim, Unikraft was ready sooner. We lead with the part that did not hold.

Timeline

Six stages, one thread.

The notebook follows the order we worked in: sandbox first, then Xen nested in a VM, then the physical machine, then attacks on our own design, then the fast cells, and last the cloud. Bar height is the number of labs shown for each stage; the reference code (11 more cards) sits apart at the end.

timeline
2SANDBOX14NESTED XEN27BARE METAL16RED TEAM15KZ CELLS2SKYSANDBOX2NESTED XEN14BARE METAL27RED TEAM16KZ CELLS15SKY2
timeline   Labs shown per stage. A nested or cloud result never counts as an isolation or performance verdict: only bare metal can close those.
Stage 1 · 16 labs

Nested Xen, the first answers.

Inside a VM (Xen under QEMU/KVM). These runs show the architecture working as logic on real domains. They are never an isolation or performance verdict, and each card says so.

Each card: lab, plane and verdict, then what was found, then the limit that comes with it. Figures are the ones the lab recorded; where a later campaign superseded them, the card says so.

lab03 · nested VM  PARTIAL

NetBSD rump kernels: closing the unknowns

The real IP stack uses 52 hypercalls (base 46) and 28 of 46 are invoked at startup; in the official order the real definition wins in 3 of 3 duplicates; the sysctl warning is benign.

limit  One nested guest; partially reproducible build (186/190 identical); no conclusion on performance or isolation.

lab04 · nested VM  PASS

Build non-determinism, guest disk, skill tests

The difference was only embedded source paths: two MKREPRO builds come out identical, 190/190. It corrects the lab03 hypothesis; skill tests 12/12 after rerunning one.

limit  n=2 builds per condition; the exact cause is strong inference, not fact; one case failed on the first pass.

lab05 · nested VM  PASS

Nested Xen and a minimal PVH cell

The cell prints its marker and exits on its own; 5 cycles and 1 destroy during boot leave 0 domains and 0 residual xenstore keys.

limit  All nested and without IOMMU: not valid for performance or isolation; not a benchmark.

lab06 · nested VM  PARTIAL

Skill validation and nested Xen cell (A to D)

A: 10/12 PASS; B: 7 PASS and 1 FAIL (one case invented a format); C: 5/5, 30 cycles with 0 residue and 0 failures; D: 5/5. The rump hypercall count is 46 on both sides.

limit  Nested, no IOMMU; three cases cannot be reproduced because their prompt and criterion are not on disk.

lab07 · nested VM  PASS

S4.0 xenstore probe between two cells

Without an ACL the cross read returns EACCES (30/30 in the control); the owning cell grants access and the other reads the nonce and gets 3 events. dom0 is not required.

limit  Nested, N=1 per arm, a single xenstored variant; no quotas or concurrency.

lab07b · nested VM  PASS

S4.2 gntalloc+gntdev+evtchn ring between two cells

Bidirectional ring with identical hashes at both ends, no dom0 privilege. The first run FAILED because of our own faulty protocol and was fixed.

limit  Nested, n=1 PASS; N≥30 pending; it was not shown that a third domain cannot map the grant.

lab08 · nested VM  PASS

Placeholder domains for control, broker and observer

30 cycles (90 domains): 0 XenStore residue and 90/90 exit code 0; a test Cell was denied in 15/15 operations on the three role branches.

limit  Nested, one run, placeholder domains with no real logic; not an isolation verdict.

lab09 · nested VM  PASS

Role logic inside an initramfs

The three suites pass inside a 256 MB PVH domain (17/17, 31/31, 9/9) with ~167-172 MB free and 0 out-of-memory signals.

limit  Nested, n=1; functional viability and memory headroom only.

lab10 · nested VM  PASS

Broker in one domain and a Cell over the ring

5/5: a 9000 B artifact with identical hash both ways; foreign or forged CID, replay and forced destination denied; 0 processes of the pair in dom0.

limit  Nested, n=1; no conclusion on performance or resistance on Xen bare metal.

lab11 · nested VM  PASS

Observer that countersigns and issues a destruction receipt

4/4: countersignature only while the Cell was alive during a real STORE; destruction receipt only after destroying it; 17/17 attempts to read, alter or delete the stream denied.

limit  Nested, n=1; does not cover a hostile Cell with kernel capabilities.

lab12 · nested VM  PASS

Chain A, receipt, B across real domains

All 4 conditions hold: B does not exist before the order, without a countersignature the Gate denies, a Cell trying to sign the order is denied (18 times) and the hash is identical end to end.

limit  PASS bounded to architecture: nested, n=1, first-generation symmetric signature; the first attempt failed on packaging.

lab13 · nested VM  PASS

B1.8 Falsifications only the substrate can refute

All 3 falsifications PASS: CIDs from the previous epoch give STALE_EPOCH after recreating the broker; a third domain cannot map the grant (128 refs denied); a kill between STAGED and COMMITTED gives NOT_COMMITTED. Finding: epoch and nonces must be durable outside the domain.

limit  Nested, n=1, HMAC v0 signature; the kill instant is forced; dom0 remains in the trust base.

lab14 · nested VM  FAIL

T15b First boot of rump as a PVH guest under Xen

Instructive FAIL: the guest entered via PVH, reached long mode and ran rump_init, but did not print the marker. Two console faults (unprogrammed hypercalls and a buffer removed by the compiler) dictated the fixes that lab16 closed.

limit  Nested, n=1; it stayed unobserved whether rump_init returned 0.

lab15 · nested VM  FAIL

S9.2 Observer falsifications 8 and 9

F8 PASS: the observer survives a cell crash and the Gate does not use its self-report. F9 FAIL: with 38,577 mutations there were 4,520 crashes, 85 false countersignatures (only with a compromised stream) and 56 hangs. A crash signs nothing: it is an availability failure.

limit  Nested, n=1, HMAC v0; the lie is a fixed script; dom0 remains in the trust base.

lab16 · nested VM  PASS

T15c Closing the rump axis: the rump guest prints KZ-RUMP-OK under Xen

With the two console fixes, the rump guest prints the marker, rump_init returns 0 and it shuts down by hypercall with code 0. It closes the bit lab14 could not read.

limit  Nested, n=1: init and console only; no performance or hard isolation.

lab17 · nested VM  PASS

T16 Adversarial guest: A_effective contained in A_delegated

289 attempts to map foreign grants and 289 to bind event channels gave 0 successes, 0 foreign pages read, 0 self-grants. What Xen contains is separated from what Kawésqar contains (replay, STALE_EPOCH, NOT_COMMITTED).

limit  Nested, n=1; the guarantee assumes a correct Xen; grant and channel denials are Xen's, not Kawésqar's.

Stage 2 · 12 labs

The first campaign on metal.

On the physical control domain of one laptop workstation, n=1. Here the first honest failures appear: a covert channel we could not measure and a disk that leaked.

lab18 · bare metal  PASS

Physical baseline and gate for the lab plane

Inventory captured and preflight 11/11. The CPUs in this Xeon are not affected by L1TF, MDS or TAA, so the side-channel lab was re-scoped.

limit  ZBook, n=1, a single day; this is inventory and preparation, not an isolation or performance result.

lab19 · bare metal  PASS

Re-attesting the architecture core on metal

7/7 core verdicts PASS on the physical control domain. Minimal cell cycle: p50 0.95 s, p99 1.09 s, 200/200 without failure, real residue 0.

limit  ZBook, n=1, control domain capped at 4 vCPU: correctness, not performance; the chain on real domains is still HMAC.

lab20a · bare metal  PARTIAL

VT-d enforcement on bare metal with a named DMA device

VT-d configuration was attested and a denial was seen once, but it is not reproducible. Under strict mode the instrument ended UNCLEAR: the positive control was denied.

limit  ZBook, n=1. It must not be cited as “strict denies unmapped transfers”: it is an observation, not a tested result.

lab21 · bare metal  PASS

Baseline latency of a cell on bare metal

200/200 per variant, 0 residue: create with xl 184 ms (p50) and kernel plus init boot 607 ms (p50, p99 790 ms). A cell with work adds a bounded overhead.

limit  First-generation figures (Linux kernel and xl), long before today's cells: do not compare them.

lab22 · bare metal  PASS

Scale and load

0 Gate failures in 256 runs under load; 2,775 cycles completed at 100%; creation steady around 237 ms from N=1 to N=414; 0 residual cells.

limit  This is the Gate logic under scheduling pressure, not real cell transitions at scale; the resource ceiling was never stressed.

lab23 · bare metal  UNCLEAR

Contention covert channel

The positive control did not cross the threshold: best case 9.64 bps with 95% CI [-33.06 ; 52.31], indistinguishable from noise. Neither presence nor absence of a channel is claimed.

limit  Instrument not validated; recorded as an unmeasured channel, never as negligible.

lab23b · bare metal  MITIGATED

Covert channel: closed by policy

Risk closed by a placement policy (do not co-schedule mutually distrusting cells on sibling threads), not by measurement: the strongest SMT channel does not exist by construction.

limit  It is not “we showed there is no channel”: shared-cache and cross-core memory channels remain as residual risk.

lab24 · bare metal  OPEN

XSAs that really apply

Three advisories patched, the speculative ones inapplicable on this silicon or mitigated; one applicable advisory remains unpatched (a liveness DoS in the xenstore daemon). None is an escalation or cross-domain leak.

limit  State tied to a package version; it must be rerun when the build changes.

lab25 · bare metal  FAIL

Disposable cell: RAM and disk residue

RAM PASS: 0 hits with forced overlap of ≥548 MiB. Disk FAIL: cell B sees A's pattern (6,882,854 hits); zeroing the whole volume on create brings it to 0 hits.

limit  The mitigation was verified in the lab but was not yet in the product; device DMA residue out of scope.

lab26 · bare metal  PARTIAL

Closing the bare metal campaign

Architecture: bounded PASS, conditional close. No amplification of authority or evidence on real Xen (7/7) and under load; disposable cell falsified on disk and PASS on RAM.

limit  No verified physical isolation is claimed, nor unconditional disposable cell, nor the full thesis; covert channel UNCLEAR.

lab27 · bare metal  TIE

Boundary shootout: KZ-Rump against minimal Linux

The hypothesis “rump exposes less” was falsified: rump keeps 2 live event channels (minimal Linux 0); they tie on grants (0) and on 38 xenstore nodes. Finding: an adapter abort gave reason code 0, same as a clean shutdown.

limit  Observed surface, not reachable; both runtimes are near-stubs; one machine, one run.

lab28 · logic + bare metal  PASS

Delegation tree n>1 and non-amplification by induction

Real Gate: 1,958 edges decided, 0 amplifications; real cells at the exact envelope edge (768 of 768 MiB); a grandchild asking for more than its parent is denied with AUTHORITY_NOT_SUBSET; 5 mutants detected.

limit  The parentage tree is modeled (Xen does not know it); depth above 2, fan-in and fuzz are logic. The lab found real Gate defects, since fixed.

Stage 3 · 15 labs

Pushing the limits.

Still on metal: observer fuzzing, the signed chain inside real cells, saturation to 412 live cells, strict IOMMU and observed hypercalls.

lab29 · bare metal  PASS

Falsification 9: observer surface

The first run FALSIFIED the property: 47 crashes and 208 violations through holes reachable via JSON. After hardening the observer: 0 crashes and 0 violations across 3,000 variants, 3 mutants detected, 0 residue.

limit  PASS bounded to the wire (decoded JSON stream); in-process violations remain, declared as a GAP. n=1.

lab30 · bare metal  PASS

Falsification 10: transactional STORE under kill

At no stage was a partial COMMITTED deliverable left (7 real kill -9); replay after restart always gave STALE_EPOCH; 3 mutants detected; real cells destroyed.

limit  Journal persistence is modeled by the harness: it tests the contract's atomicity, not the production persistent CAS (declared GAP).

lab31 · bare metal  PASS

Adversarial Gate under load

0 acceptances out of 6,870 concurrent adversarial attempts; 1,194 of 1,194 legitimate ones cross under the same load; the naive race over-reserves 6 of 6 (the instrument sees the race); 0 residue.

limit  The ledger lock is in-process; production persistent CAS is out. A finite corpus, not a general proof of absence.

lab32 · bare metal  PASS

Stability soak of the create-run-destroy cycle

400 of 400 iterations: zero drift (Kendall tau 0.02 to 0.08), residue 0, Xen free memory 55,883 to 55,883 MiB, dom0 stable.

limit  Bounded window (not months), one minimal cell, one machine, n=1.

lab33 · bare metal  PASS

Saturation at maximum scale and throughput

Peak of 412 simultaneously live cells (52,736 of 53,835 MiB), 0 amplification and 0 residue across 14 levels; at most 14.78 cells/s creation and 8.98 cells/s full cycle.

limit  n=1. At the peak, destruction reaches p99 32 s because the toolstack serializes; at scale only a sample carries a console marker.

lab34 · logic + bare metal  PASS

Resource fan-out (Falsification #4)

In logic: 4,681 nodes and 7,608 edges with 0 amplifications; on metal, with 128 MiB cells, every step completes, 0 amplifications and per-cell overhead of 1.5 to 2.5 MiB.

limit  The first two physical attempts were UNCLEAR (cause found in lab40); the close is the third. n=1.

lab35 · bare metal  PASS

Minimal observed boundary of the cell under load

Full classification of 18 interfaces and 49 hypercalls; load did not break the boundary, 0 residue in all 4 scenarios. Xen, not the runtime, bounds isolation.

limit  Characterization, not an isolation verdict. Trace observation stayed UNKNOWN on all 49 rows for lack of a decoder (lab44 closed it).

lab38 · bare metal  PASS

End-to-end Ed25519 cryptographic chain inside cells

The chain of cell A, countersignature, Gate, cell B and destruction receipt runs on metal; 4 mutants (dom0 forgery, altered receipt, B without evidence, replay) and 7 negative vectors rejected without creating B.

limit  Verifying the signature before storing is modeled by the harness, not by the reference Gate. n=1.

lab39 · bare metal  PASS

Smoke and invariant regression under strict IOMMU

All 4 groups pass: 60/60 iterations, 0 accepted out of 3,444 adversarial attempts with 600/600 legitimate admitted, and zeroing verified; create rises about 12% (some 23 ms) under strict.

limit  Bounded regression, not exhaustive re-testing; the clean A/B in the same boot is pending. n=1.

lab40 · bare metal  PASS

Boot memory floor of a PVH cell

Measured floor of 96 MiB: with 64 MiB the cell does not boot, with 96 and 128 MiB it does. Closes the root cause of the lab34 failure and falsifies the harness race hypothesis.

limit  Coarse floor (n=3 per point), valid for this lab kernel and initrd; the exact value between 65 and 96 MiB was not refined.

lab41 · bare metal  PASS

Audit of dom0 grants over cell pages

A diskless cell, idle or under load, has exactly 1 active grant mapped by dom0 (the console ring), 0 data pages; with a disk 67 refs appear (positive control).

limit  The first run was UNCLEAR because of a blind positive control and was fixed. That they enter dom0's IOMMU is inference from the source. n=1.

lab42 · bare metal  PASS

Entropy boundary inside the cell

14 of 14 distinct keys; in no boot was a key generated while the random generator was not ready, even with RDRAND off (median 1.47 s versus 0.14 s).

limit  It shows availability and non-determinism, not entropy quality. The no-RDRAND case rests on reading the source. n=1.

lab43 · bare metal  PASS

Cryptographic chain in a cell with n greater than 1

The chain and the fan-out cross end to end and every mutant is rejected with its exact reason; the fan-out fills the exact envelope (2 by 128 MiB). Finding: the pure gate does not consume the prev link of destruction receipts.

limit  Signature verification and the prev link are modeled in the harness (production gap declared). n=1.

lab44 · bare metal  PASS

Hypercalls observed through a real trace

In all 4 scenarios what is observed is a subset of the declared set of 10 hypercalls; the rump cell leaves a minimal footprint (5 distinct numbers) and Linux adds 4 more. Negative control: 0 calls with the cell paused.

limit  The hypercall number is observed, not the sub-op or the arguments. n=1.

lab45 · bare metal  PASS

Soak under strict IOMMU and structure leaks

400 of 400 cycles without leaks: active grants flat at 2, maptrack flat at 4 and Xen free memory constant (55,883 MiB before and after).

limit  Free memory is read in MiB (detectable minimum near 80 KiB per cycle); active grants are a partial proxy. n=1.

Stage 4 · 16 labs

Attack, then verify outside dom0.

The decision moves out of the control domain, then we attack our own design, measure what dom0 can still read and compare against Linux.

lab46 · logic  PASS

Denial conformance and live agent against the real channel

12 of 12 checks and the independent oracle PASS, each denial with its exact reason; 0 escalations crossed with the 8 language models tested, including the 3 that did attack through the real channel.

limit  Logic level, not metal nor an agent inside the cell. A later campaign corrected the lab's own overclaims (mutants that mutated data, not code).

lab47 · bare metal  PASS

Verifier outside dom0 (chain A to B)

The positive control crosses and all 11 mutants are denied with their exact reason; creations = admissions = verifier PASS (1 = 1 = 1); dom0 forgeries and replays die by signature. Re-attested on metal: 3 of 3 runs PASS.

limit  n=1, one host. dom0 still creates domains and can read the RAM of a live domain (hardware residual). The first physical run was UNCLEAR because of an oracle bug, fixed.

lab48 · bare metal  PASS

Hardened gate (gate_verified) against the harness

With the harness switched off, the old gate ACCEPTS and creates the child in the 4 main mutants; the new one DENIES them with its reason. The properties move from modeled in the harness to enforced by the gate.

limit  n=1. dom0 still generated the private keys in this run (lab47 attacks that); a lying tap is not covered.

lab49-c5 · logic + bare metal  PASS

Closing C5: compromised broker

The old observer accepts both forgeries; kaf-store denies both and the byte witness denies the first. On metal, a compromised broker that signed without backing the bytes leaves the medium empty and does not cross.

limit  The witness lets through a false COMMITTED over bytes honestly in staging. This does not settle the later red-team marker, which records the C5 close as not won. n=1.

lab50-dma · bare metal  PASS

DMA confinement of an assigned device

dom0 reads the USB stick before assignment, loses the device on assignment, and the domU reads the same number of sectors with 0 VT-d faults in strict mode. It measures confinement, not adversarial isolation.

limit  Partial: a malicious driver attempting DMA into another domain's memory was not tested. It ran in HVM because PVH does not support passthrough. Do not cite as closed DMA isolation.

lab51-netless · bare metal  PASS

lab51-netless · cell with a kernel without a network stack

A PVH cell with a kernel built without a network stack boots and runs: /proc/net, /sys/class/net and iplink ABSENT and the load closes with sum=500500, with the console ready in under 1 s of kernel boot.

limit  A fact about how a kernel was built, not an isolation or performance verdict; the evidence is the console and the configuration. n=1.

lab52-smt · bare metal  PASS

lab52-smt · placement against the SMT channel

A neighbor on the sibling thread slows the victim 4.98 to 5.94 times; on another core it drops to 1.07 to 1.13 times, near baseline. Placement suppresses the SMT channel (3 runs).

limit  It measures throughput interference, not a bit rate or key extraction. The ~1.08x residual is the cross-core channel (L3). n=1.

lab53-llc · bare metal  PARTIAL

lab53-llc · cross-core channel through the L3

The cross-core channel exists and is small: 1.03 to 1.14 times interference (average ~1.08x), an order of magnitude below the SMT channel. A residual measured and bounded, not closed.

limit  Without Intel CAT/RDT the L3 cannot be partitioned on this CPU. Small, noisy signal (3 runs); it measures throughput. n=1.

lab54-meminspect · bare metal  CONFIRMED

lab54-meminspect · the control domain reads a live cell

Positive control: dom0 mapped 10,076 frames of the live cell and found the secret without the cell's cooperation (2 runs). The signature authenticates, but gives no secrecy against dom0.

limit  It measures the exposure, it does not close it. Only hardware-encrypted memory (SEV-SNP/TDX) closes it; this CPU lacks it. n=1.

lab54b · bare metal  CONFIRMED

lab54b-meminspect-incell · secret born inside the cell

3 of 3 runs with 3 distinct secrets: 1 candidate per run, hash verified, 977 contiguous frames. The secrecy residual against dom0 is now measured.

limit  It does not close the residual: that needs hardware-encrypted memory (TDX/SEV-SNP), which this machine lacks, and that closing test has not been run. The same residual exists in every cloud, AWS and GCP included, unless confidential computing is on: whoever controls the host, the provider or an attacker who gets root on it, can read the memory of what runs on it. It is a property of the platform, not of Kawésqar. We did try confidential VMs in both clouds (TDX on GCP, SEV-SNP on AWS): memory is encrypted, but the cell cannot run inside (no KVM) and a confidential VM per job takes 50 to 58 s to come up. n=1.

lab55 · bare metal  PASS

lab55-composition · the five properties together

5 of 5 runs PASS: the hardened gate decides outside dom0, with no private key in dom0; forged or repeated orders die by signature and mutants are denied. Closes the “Composition” item.

limit  n=1, 5 repetitions (reproducibility, not statistics). dom0 still creates domains and can inspect memory (lab54 residual).

lab56 · bare metal  PARTIAL

lab56-substrate · full NetBSD or trimmed Linux

Instantiation: NetBSD PVH ~1.65 s, Linux PVH 2.06 s, NetBSD PV 3.19 s. Both substrates sustain 8 concurrent cells with near-flat scaling from N=4 to N=8.

limit  Early-stage figures, earlier than today's cells; the lab's text disagrees with its own table, so only the table is cited. n=1.

lab57-gate · bare metal  PASS

Real rump cell under the Gate

The honest run PASSES and 5 mutations give their named DENY before running; a chain of 3 real cells, 8/8 cells gated in parallel, 20/20 in stability. An “echo” cell without rump passed with a bytes-only witness and is denied once the kernel is attested.

limit  Bounded PASS, NOT containment shown: the console is a degraded medium and dom0 can read the cell's RAM (hardware residual). Small n.

lab58-overhead · bare metal  PASS

Xen and rump cell overhead against Linux

Rump cell cycle without work ~94 ms versus ~500 ms for the Linux cell (5.3x); boot 0.9 ms versus 237 ms; Xen charges +8% compute in dom0 and +11% in a Linux PVH cell versus bare metal; the rump cell sits at +4% of native.

limit  n=50 per cell, one machine. An anomaly in the rump sha phase (+15 to +20%) stays UNCLEAR; no I/O measured. Create figures predate the direct launcher.

lab59 · logic + bare metal  PARTIAL

Cell-escape red team (campaigns A to F)

On metal a compromised cell read 0 pages of a neighbor, opened 0 unauthorized channels and gained no authority; the launcher's prefix cleanup did kill cells of another job (closed in logic); native fuzz 85M executions without crashes; 15 of 16 ledger items closed in logic.

limit  What is closed is the logic plane and has not been re-measured on metal. Still open: the verifier that does not re-hash the bytes, the cross-core cache channel and dom0 reading cell RAM.

redteam-escape · logic  PARTIAL

redteam-escape · Wave 1 of logic attacks (C5, E5, E6, E8, CGAPS)

C5: the close was not won (a broker that forges the storage confirmation crosses; 23/23 behaviors reproduced); E6: prefix cleanup kills cells of another job; E8: the memory check admits 11 extra cells; E5: DENY. The failures were repaired later in lab59.

limit  Logic plane in a single process, no metal; this lab records the state before the fix.

Stage 5 · 15 labs

Cells, a signed chain and the clock.

Direct launchers, the cost of the Gate, sealed images, the minimal builder in three phases and the energy per cell. Three hypotheses were refuted and one check failed here, and all are kept.

lab61 · logic  PASS

Secure Boot negatives and adversarial map

5 of 5 cases as expected: the valid control verifies, and unsigned, foreign-key and mismatched-certificate artifacts are rejected; the artifact without SBAT is detected as such.

limit  Offline signature verification: loader rejection at real boot is described as expected, not measured; 17 adversarial dimensions remain pending.

lab62 · bare metal  HELD

kz-k2 on KVM against Firecracker

With the cell before the trim, ready from zero in 4.8 ms (kz-k2 on KVM) against 8.2 ms for Firecracker with the same cell; against Firecracker with a Linux guest the pre-registered hypothesis holds at 2.14 to 2.34x; at P=8 on the minimal cell, 251 versus 147 cells/s.

limit  n=1. With the same cell the edge is 1.04 to 1.10x on 100 ms jobs: it lives in the cell, not the VMM. On 1000 ms jobs it is lost (~1x). Against Firecracker with an untuned Linux, 3.5x measured on the same machine.

lab63 · logic  PASS

lab63 · Gate to launcher chain with Ed25519

107 negative tests pass (altered receipt, missing or foreign countersignature, broken lineage, badly signed order, replay, replay after restart) and the closure of the 5 images gives 84 of 84 checks.

limit  Tested in logic and static build, without Xen or a VM; the end-to-end Xen evidence with Ed25519 is in later labs.

lab64 · bare metal  REFUTED

lab64 · Gate and order cost over the cell cycle

The pre-registered “under 1 ms” hypothesis is refuted: Gate plus order add a median of 2.735 ms (P=1, n=200); the A/B delta in total time is +2.574 ms, and +10.815 ms at P=8.

limit  n=1, ondemand governor. At P=8 the Gate queues behind a single lock, so this is not “the Gate is slower under load”. A recorded failure, published as a trait.

lab65 · bare metal  PASS

lab65 · verifier outside dom0 over real rump cells

On metal 16 of 16 scenarios as predicted; the extension with artifact substitution gives 19 of 19 in three runs, 0 orphans; in logic 62 fidelity tests and 10 mutation tests pass.

limit  n=1. A dom0 that lies coherently gets PASS, declared a deliberate NO-CLAIM. Gate state lives in the verifier's RAM. dom0 relay channel, no mTLS. A bounded verdict, not “containment shown”.

lab66 · bare metal  PARTIAL

lab66 · kz-c2 with a sealed kernel (closing the lab64 TOCTOU)

Verifying the hash on every cycle costs +7.16 ms on a cycle of about 15 ms and is discarded; the second variant costs -0.27 ms in total_ms (15.345 to 15.079), within noise, with 9 negative tests OK.

limit  The TOCTOU attack was shown only with a stand-in for the loader, not real Xen. n=1, no SHA-NI. The full chain with Gate was measured in lab70 (+3.21 ms).

lab68 · bare metal  PASS

lab68-e2e-real · real evidence from A to cell B

10 of 10 chains of A, Gate, order, direct launcher and B on metal (2 batches of 5), oracle PASS in each; in logic 128 launcher checks and 54 preflight checks without failures.

limit  n=1. dom0 remains the trust base for the tap, and B's FETCH is a signed echo, not a real FETCH to the broker. No mTLS on that channel. The first smoke on metal was UNCLEAR and was archived.

lab70 · bare metal  REFUTED

lab70-gate-fast-fd · cost of the fast Gate over launching without a Gate

Hypothesis H1 (under 1.5 ms of wall time) is refuted: the full arm adds +3.210 ms median (95% CI 3.074 to 3.378) over the cell without Gate; repeated with the performance governor: +3.216 ms.

limit  n=1, Xen, P=1. The descriptor arm includes a durable nonce fsync that the others do not pay.

lab71 · bare metal  PASS

lab71-gate-evidence · single activation and release order

1000 of 1000 rounds with exactly one activation (the baseline without the predicate crosses 300 of 300); unsigned, foreign or repeated orders give 0 creations; 10 of 10 cycles with the release confirmed before the verdict.

limit  n=1, N up to 32, no power cut mid-write. In one run an arm gave an environment FAIL and a probe was fixed after three instrument defects: reported as they are.

lab72 · bare metal  PASS

lab72-b-evidence · evidence for Application B

After destroy, residue enumeration gave 0 in 100 of 100 cells, with destruction verified in at most 0.093 ms; persistent ledger with kill -9 and restart without violations; creating the Xen domain through the direct launcher 41 ms versus 115 ms with xl. It includes a TLA+ model with trace conformance.

limit  n=1. fs-verity and dm-verity were BLOCKED on Xen and were only tested with a KVM cell on another machine. The 41 ms is Xen domain creation, not cell latency.

kz-f1 · f0 · bare metal  PARTIAL

KZ-F1 phase 0: energy, CPUID policy and zeroed pages

C1, energy reading from dom0: FAIL against the pre-registered criterion; the MSR register path returns an I/O error under Xen 4.20. An alternative interface answered, but it is not validated (n=1, no external meter; two registered checks of it also missed their criteria) and it affects only energy measurement. C2, cell without CPUID policy: REVIEW (80 of 80 pairs with identical behavior, the dump differs). C3, zeroed pages after reusing memory: PASS (210 domains of 256 MiB, at least 51,636 MiB reused, 0 non-zero bytes).

limit  ZBook, n=1, one run per check. C1 is not a citable energy figure.

kz-f1 · f1 · bare metal  PASS

KZ-F1 phase 1: signed cycle, build time and xenstore

Full cycle with a signed order: smoke PASS (8 of 8 valid orders; 4 of 4 invalid orders rejected without creating a domain), repeated with the final front: PASS. H2, build time against kz-c2: INSIDE (paired difference of -2.1 to -3.9 ms; criterion ≤ +1.0 ms), repeated with the final front: INSIDE. H3, xenstore nodes and writes per cell: PASS twice (0 nodes and 0 writes across 200 cells). H7a: PASS. Sweep smoke (the worker dies with the cell alive): PASS, 8 of 8.

limit  ZBook, n=1: the smoke, H2 and H3 ran twice on the same machine; H7a and the sweep once. Build time is not end to end: end to end this phase was slower than kz-c2 (+7.9 to +16.2 ms).

kz-f1 · f2 · bare metal  PARTIAL

KZ-F1 phase 2: resident builder, recovery and scale

Cycle of 15.1 to 18.1 ms against 22.7 to 26.9 ms for the reference arm, Gate included. H1 INSIDE: paired improvement of -8.672 to -7.243 ms with 200 pairs per variant (criterion ≤ -4.0 ms). H6 PASS: 13 states × 100 and 200 clean endings, 0 violations. Smoke 8 of 8 with 3 negatives rejected; sweep 21 of 21; real failures 20 of 20 across the four cells. Not met: H1b INCONCLUSIVE (1.022 to 1.059 ms against a threshold of ≤ 1.0); H4 BETWEEN THRESHOLDS (entry 3.443 to 3.647 ms and 224 to 280 cells/s; criterion ≤ 3 ms and ≥ 243/s); H5 BETWEEN THRESHOLDS (1,800 cells per variant, no wall, lowest decile 74.1/s against a criterion of ≥ 78/s); H7b REFUTED (7,840 lines of code against a criterion of ≤ 1,800).

limit  ZBook, n=1, one run per test, unsealed. The cycle definitions differ from the KVM ones and from the rival's (Xen against KVM, with and without the signing Gate) and are not chained. “Between thresholds” and “inconclusive” are not passes.

energy · bare metal  PARTIAL

Energy per cell with RAPL counters, native kernel without hypervisor

Same task of about 100 ms, processor and memory energy per cell, with idle subtracted: native work 0.74 J, our cell 0.89 J, Firecracker with our same cell 0.98 J, Firecracker with Linux 2.04 J. Overhead over the native work: 0.15 J against 1.30 J. That is less than half the processor energy of a Firecracker microVM with Linux, for the same work. Against Firecracker with the same cell it is close to a tie (0.89 against 0.98 J), and with a new process per cell, 0.99 J, it ties. Full life cycle of a cell (boots, passes its test, shuts down), with no work: minimal 24 mJ, with network 28, with files 32, with network and files 38. 5 repetitions of 400 cells per arm, 0 failures. Coherent with kz-f1 f0: there the energy reading was taken from dom0 under Xen and failed; here the kernel is native, without a hypervisor (KVM).

limit  ZBook (one laptop), native kernel, synthetic load (a hash in a loop). This is chip energy (RAPL: processor and memory), not the wall plug or the electricity bill; it does not cover network, disk, fans or power supply. The saving against Firecracker with Linux comes from replacing Linux with the cell, not from the cell against Firecracker. No money figures. Still to repeat on a server and to validate against an external meter. Evidence recorded with hash and timestamp.

kz-c1 · bare metal  PASS

kz-c1, cell launcher with direct libxl

A cycle of about 38 ms (create 26 + life 2 + destroy 9) against about 94 ms plus harness with xl; 64/64 cells ok in each configuration and a ceiling of 50 to 60 cells/s per node from P=4, because creation serializes in the toolstack.

limit  Measured on the ZBook, before later generations: not the current cycle nor a latency headline. No confidence intervals.

Stage 6 · 2 labs

The cloud, kept in its place.

Nested runs orient and nothing else: the performance and isolation verdicts are the ones measured on metal.

kz-gc1 · cloud (nested)  PARTIAL

kz-gc1: the KVM cell in a cloud VM (nested KVM)

Cell ready from zero in 6.9 ms against 107 ms for untuned Firecracker + Linux. On ~100 ms jobs: 2.11x at the worst P, against untuned Linux the edge shows; with the same cell on Firecracker it drops to 1.04x (it orients). A burst of 20,000 cells without failures and 13,053 live cells at 1.86 MB each.

limit  Exploratory · in the cloud (nested) · it orients, it is not a verdict; the performance and isolation verdicts are the ZBook's. One round per point; the 13,053 was cut by time, not a RAM ceiling.

kz-az1 · bare metal + nested cloud  PARTIAL

kz-az1: duel on metal and runs on nested AWS

On metal, the 2x hypothesis on ~100 ms jobs holds only against untuned Linux (2.01x, borderline) and is falsified against tuned Linux (0.92x to 1.03x, technical tie). With the best measured configuration after the trim, kz-k2 ready 1.51 versus 2.32 ms for Unikraft and 1.03 versus 1.49 MiB per cell; before the trim Unikraft arrived first (2.25 versus 4.83 ms).

limit  One machine, one round per point. At P=15 Unikraft stays ahead on rate in 3 of 4 cells; with defaults Unikraft leads on ready time. Nested AWS orients, it is not a verdict. Snapshot-restore stays out of the rankings. The trim mechanism is not published.

Stage 7 · 11 labs

The reference code.

The components themselves, as standalone scripts that print their checks and exit 0 only if every vector matched its expected verdict. Mostly logic plane.

impl · gate · logic  PASS

Gate (admission gate for the Job)

42/42 checks without failures (rc=0); reservation without failures and indexed ledger 8/8. Across the 31 extended adversarial vectors (gate, broker, chain): 19 DENY, 9 declared GAPs, 3 LIMIT by design.

limit  Logic plane. The 9 GAPs are published without hiding them; C5 was the only high-severity one. The Gate's metal run is lab71.

impl · broker · logic  PASS

Broker (mediated I/O with CID)

71/71 checks (68 vectors, purity, derivation and signatures), 0 failures, rc=0. Covers the peer key pin: another certificate from the same CA with another key is denied.

limit  Logic plane. Declared GAPs remain (B2, B5, B6 by design, and B7, state rollback, which depends on dom0 durability). Not an isolation verdict.

impl · chain · logic  PASS

Chain (receipt and lineage chain)

37/37 scenarios without failures, rc=0. Signature coverage now attests the full set of fields (moving any gives BAD_SIGNATURE) and lineage requires a real destruction receipt.

limit  Logic plane. Starting findings that are published: lineage was vacuous because of a public constant and signature coverage was partial; both were fixed in the real code.

impl · observer · logic  PASS

Observer (countersignature and destruction receipt)

14/14 checks without failures and the destruction-receipt chain with rc=0. Removing or replacing the countersignature (6+3 variants) is denied with MISSING_COSIGNATURE.

limit  Logic plane. The observer countersigned without an observed commit fact (C5); it was addressed in the real code with a separate store. Independence in production is not claimed here.

impl · ring · logic  PASS

Ring (framing of the request and response channel)

62/62 vectors without failures, rc=0. The receiver validates only a snapshot of the page, and large artifacts travel in chunks.

limit  Logic plane (loopback on one machine, no real Xen channel). Python's fence is best-effort. Not an isolation or performance verdict.

impl · transport · logic  PARTIAL

Transport (mTLS for the control plane)

In logic: mtls 22/22, serve 21/21, integration 13/13 (real handshake, 0 SKIP), tier2 6/6, ca 18/18, identity 8/8. On metal, once: 12/12 suites with a real TLS 1.3 handshake. Another certificate from the same CA with another key is denied.

limit  Open: running it between distinct OS principals and between separate Xen domains. On metal both ends had the same UID; the current inter-domain channel is a relay without mTLS.

impl · crypto · logic  PASS

kzsig: Ed25519 asymmetric signing

15/15 checks, 0 failures: with public keys only the verifier checks but cannot forge; 300 random signatures, a reused signature and a signature made with the public key all fail; a dom0 signing with its own key is rejected.

limit  Pure logic. Authentically distributing the three public keys remains open; dom0 can still drop or delay; the private keys are not protected by this layer.

impl · kaf-store · logic  PASS

kaf-store: per-process isolated storage (closing C5)

22/22 checks (rc=0): asking for a signature over bytes that were never stored, storing garbage and declaring another hash, or writing straight to the medium, is denied; no private key reachable from the broker process.

limit  It closes only the compromised-broker-process case. Against dom0 with root, process isolation is not enough: that is a hardware residual. Not tested on metal in this directory.

impl · alicebob · logic  PARTIAL

alicebob: notary and key registry for console agents

Two adversarial reviews found and fixed about twenty defects (rewritable or truncatable chain, per-Job owner, injection in the audit query, among others); the Go tests for chain, signature, server and store pass.

limit  Finding NOT fixed in code: if daemon and agent run under the same UID there is no agent/human boundary. Linear log, not Merkle; no performance figure.

impl · rump-xen · bare metal  PASS

KZ-Rump cell on Xen PVH (and the same image on KVM)

Honest sequence: T15a link PASS; T15b the console FAILED; T15c KZ-RUMP-OK; T16a fs-only BLOCKED (UNCLEAR); T16b fs-only PASS 3/3; T16c net-only PASS, 8-thread contention without aborts and a deliberate deadlock that dies observably at 30 s. A single ELF boots on Xen and on KVM.

limit  n=1, one run per test. A first audit found an insufficient watchdog and it was fixed. Current boot figures are those of the trim, not of these reports.

mtls-metal · bare metal  PARTIAL

mtls-metal: control-plane mTLS on the physical dom0

12/12 suites green on metal with a real TLS 1.3 handshake (a certificate from the same CA with another key is denied). The harness between principals with distinct UIDs never ran.

limit  Both ends with the same UID and pure-Python TLS: not separation between OS principals nor between Xen domains. n=1.

Minne

Minnet, i hver sky.

Whoever controls the host can read the memory of what runs on it: the cloud provider, or an attacker who gets root on the host. Every cloud has this (AWS, GCP, any other) unless you use confidential computing. It is not always said in these words, but that is how it works. We say it plainly so nobody takes it for a Kawésqar problem.

minne
Sky utenkonfidensiell databehandlingAWS, GCP, hvilken som helst skydin VM / din arbeidsbelastningminne i klarteksthemmelighethypervisor / vertrootleserLeverandøren, eller en angriper medroot på verten, leser minnet.Egen server,Xen dom0measured on metal, n=1cellenminne i klarteksthemmelighetdom0rootleserRoot in dom0 reads the cell memory.Målt: 3 av 3, en hemmelighet fødtinne i cellen.KonfidensielldatabehandlingTDX · SEV-SNPdin VM / din arbeidsbelastningminne kryptert▒▓░ x9#f 7a▒ ░▓hypervisor / vertrootblockedThe host sees only ciphertext.The read is blocked.Samme sikkerhetsmodell i alle tre;det eneste som endrer seg, er minnekrypteringen.Sky utenkonfidensiell databehandlingAWS, GCP, hvilken som helst skydin VM / din arbeidsbelastningminne i klarteksthemmelighethypervisor / vertrootleserLeverandøren, eller en angriper medroot på verten, leser minnet.Egen server,Xen dom0measured on metal, n=1cellenminne i klarteksthemmelighetdom0rootleserRoot in dom0 reads the cell memory.Målt: 3 av 3, en hemmelighet fødtinne i cellen.KonfidensielldatabehandlingTDX · SEV-SNPdin VM / din arbeidsbelastningminne kryptert▒▓░ x9#f 7a▒ ░▓hypervisor / vertrootblockedThe host sees only ciphertext.The read is blocked.Samme sikkerhetsmodell i alle tre;det eneste som endrer seg, er minnekrypteringen.
memory   Who can read what. Orange arrow: a read by the host. Only the third stack, with hardware encryption, stops it.
plattformen din

It belongs to the platform, not to Kawésqar

Any workload on any host without confidential computing is readable from the host. It is a property of the platform, the same for every product that runs on it.

we tried it

Vi testet det i konfidensiell sky

TDX på GCP og SEV-SNP på AWS. Minnet forblir kryptert, men cellen kan ikke kjøre innenfor (det finnes ingen KVM), og det tar 50 til 58 s å starte én konfidensiell VM per oppdrag.

what we ask

So the closure is validated on physical hardware

The cell closure is validated on physical hardware (TDX or SEV-SNP), which is what we are asking for.

Hvordan vi bruker det

Det kritiske blir hjemme, resten går fort.

Vi går ikke imot flyttingen til skyen. Vi bruker Kawésqar-hastigheten i skyen for mindre kritiske arbeidsbelastninger, eller der risikoen kan tåles, og beholder hjemme bare det som fortjener det: suverent, on-prem eller lokalt. Det mest kritiske, som nøklene til en bank, blir på din egen maskinvare med full minnekryptering (TDX / SEV-SNP).

stige
Risikotolerantarbeidsbelastning→ SkyKawésqar-hastighetSensitivarbeidsbelastning→ On-premsuverent, hjemmeNøklene til en bank→ On-prem med full minnekrypteringTDX / SEV-SNPvalidert på fysisk maskinvaremindre kritiskmer kritiskRisikotolerantarbeidsbelastning→ SkyKawésqar-hastighetSensitivarbeidsbelastning→ On-premsuverent, hjemmeNøklene til en bank→ On-prem med full minnekrypteringTDX / SEV-SNPvalidert på fysisk maskinvaremindre kritiskmer kritisk
ladder   Where each kind of workload goes. Only what deserves it stays at home.
same contract

Same contract, same Gate

On both sides the same contract, the same Gate and the same security model apply. Only where the workload runs changes.

measured

What is measured today

Closing the cell with full memory encryption is validated on physical hardware (TDX or SEV-SNP), which is what we ask Intel and AMD for. Today the residual is measured and declared (lab54b, 3 of 3), not closed.

not yet

What does not exist yet

Federation between on-prem and cloud is a design direction, still without code.

Keep reading

Evidence, then the ask.

The evidence page puts the headline figures in context and names the gaps still open. If you want to reproduce something here, tell us plainly.

the rule

Not a certificate, not a claim of regulatory compliance: a dated, bounded record of what ran.

Supported byAvec le soutien deUnterstützt vonStøttet avCon el apoyo deCom o apoio de
NVIDIA Inception Program Anthropicfor Startups · Cyber Verification Program Auth0for Startups Google for StartupsCloud Program