Skip to content

Compliance audit

Kronik Maintainers edited this page Nov 15, 2026 · 9 revisions

Build-brief compliance audit

This audit maps every acceptance criterion in Kronik's build brief to the repository. The brief is the project's authoritative specification, maintained as the English and Swedish wiki pages; the criteria are its checkbox lists in the milestones, the v0.1 definition of done, the later releases and the v1.0 definition of done.

  • Snapshot: this page describes the v0.5.0 release on 2026-11-15 at commit fcc5bf5a (tag v0.5.0). The audit's read tree — the tree the rows below were read at — is 56ed2671, the commit of the cited ci run. The tag is 42 commits later, a 127-file, +8107/−658 diff (git rev-list --count 56ed2671..v0.5.0, git diff --shortstat 56ed2671 v0.5.0). The range is not release housekeeping alone: it includes the M25/M26 blob slices and their test work (330b3660, 5af5b89d, 6b1a3239, 8bc8d562, e14f7b4c, edd42fd8), the erased-call-answer behaviour and its reproduction tests (8e7dd5bf, 65e70e96), the nested-replay instance-pool fixes (895fc5fc, 472cbcc6), the microVM cgroup fix (3a2df151), and the compatibility corpus and the released components it runs (b0e51276, 4b7f1c5f). The rows were not re-read one by one at the tag. The M26 row, its "Changes since" row and gap 2 were revised after the tag against main at 4ed1838b (#89), and the M24 row and gap 10 against main at 4e3896a3 (#94); the artefacts those rows cite are in the tag's tree. The other milestone rows are the 56ed2671 reading, except where a row names another commit for its evidence; where a cited artefact changed in the range, they are verdicts on the earlier tree, not the tag. The summary counts follow the revised M24 and M26 rows. The previous version of this page described the v0.4.0 release.
  • Method: a static read of the tree — code, tests, scripts, fixtures, docs and ADRs — mapping each criterion to the artifacts that implement or prove it. Unlike the v0.4.0 audit, this one also read the hosted workflow runs that the project cites, through GitHub's API: the ci run on the read tree (run 43637751312) and the runs of the manual isolation workflow, the last of them on main at 7e87e5ce before M25 and M26 landed (run 43617646807), with their evidence artifacts. No KVM host, VPS, cloud account, KMS, HSM or browser was used by the auditor.
  • Nature of the verdicts: they are implementation verdicts. They say whether a first-party artifact for the criterion exists in the tree and whether a known gap contradicts it. Where a criterion needs an environment that the repository's ordinary CI does not provide (KVM, the test VPS, a cloud), the row says where it was proven: a criterion proven only on hosted runners of the manual isolation workflow, or only on the project's test VPS, says so. Historical claims with no artifact in the repository and no run to read (an external review, a fresh-agent walk, a production pilot) are marked unverifiable.

Verdict definitions

Verdict Meaning in this document
implemented A first-party code/test/script/doc artifact for the criterion exists and no known gap directly contradicts it. Not executed by the auditor.
partial An artifact exists but is incomplete, ignored or not wired into repository CI, directly contradicted by a documented gap or stub, or requires an external environment that is not reproducible here.
absent No artifact for the criterion exists on main (including criteria blocked behind an upstream dependency that the tree does not implement, and work that exists only on a branch).
deviated Implemented differently from the brief text, deliberately and with an ADR or design record.
unverifiable The criterion is a historical, process or external claim with no reproducible in-repo artifact (an independent review, a fresh-agent walk, real cloud or VPS).

Summary counts

Every checkbox criterion in the brief was counted: M0–M8 contain 81 criteria, the v0.1 definition of done contains 10, M9–M34 contain 98, and the v1.0 definition of done contains 7 — 196 criteria in total. A criterion phrased with several sub-behaviors is one row.

Scope Criteria implemented partial absent deviated unverifiable
v0.1 milestones M0–M8 81 76 3 0 1 1
v0.1 definition of done 10 8 0 0 0 2
v0.1 total (M0–M8 + DoD) 91 84 3 0 1 3
Later milestones M9–M34 98 56 21 18 0 3
v1.0 definition of done 7 0 3 2 0 2
Grand total 196 140 27 20 1 8

Headline: since v0.4.0, main gained the microVM level (M20), the nonroot service (M21), the external secret backends (M23 B2–B6) and long-lived connections (M24), streaming bodies (M25) and blob storage for apps (M26). The 18 absent later-milestone criteria are the cluster track (M16–M19, 16 criteria), WASI 0.3 (M22, 1) and fuzzing (M31, 1). The M20 and M21 proofs need KVM, root and systemd; they ran on GitHub-hosted runners in the manual isolation workflow, last on main at 7e87e5ce before M25 and M26 landed, not in the ci workflow and not on bare metal.

Changes since the v0.4.0 audit

Criterion v0.4.0 Now Why
M0.2 CI green partial implemented The ci workflow passed every job on the read tree, 42 commits before the tag (run 43637751312); it is a hosted, manually started run (ADR 0110).
M20 (3) 1 partial, 2 absent 3 implemented Firecracker level on main, proven on hosted runners with nested KVM.
M21 (2) 1 partial, 1 absent 2 implemented Nonroot service on main; demos without root proven on hosted runners; a bounded hostile-actor test in CI.
M23 (3) 3 partial 2 implemented, 1 partial B2–B6 on main: criteria 2 and 3 have tests in CI with a real daemon and a fake OpenBao server. Criterion 1 (the master key in a KMS) stays partial: in-process fakes and emulators still stand in for every KMS.
M25 (5) 5 absent 5 implemented Streaming bodies on main; CI at scaled sizes, the brief's sizes on the test machine.
M26 (13) 13 absent 12 implemented, 1 partial Blob storage on main; erasure survives a rebuild of state.db in the index, but no guest/API read of an erased blob after that rebuild (M26.12).
M24 (4) 4 absent 3 implemented, 1 partial Connections on main; criteria 2–4 in CI tests; criterion 1 needs the test VPS.

The other rows were carried over from the v0.4.0 audit and re-checked where the tree changed a documented gap (M3.2, M8.4, M8.7, M9's effective_lsn, M27, M31–M34 and the v1.0 rows); their verdicts did not change.

Milestone classification

v0.1 milestones (M0–M8)

Milestone Criteria Verdicts Headline
M0 Repository and spikes 9 9 implemented Workspace, spikes, ADRs and release scaffolding are present; the ci workflow (fmt, clippy, hygiene, test, docs, book, examples, ts, sdks, dst, deny, musl, foreign, s3, identity, action) passed every job on the read tree.
M1 Durable turns on one node 16 16 implemented Log, recovery, SQL state, determinism and crash semantics are implemented and covered by tests and acceptance scripts.
M2 Effects and capabilities 14 14 implemented Effects, capabilities, rates, crash semantics and secret isolation are implemented and covered by tests.
M3 Fork, replay and behavior diff 13 11 implemented, 1 partial, 1 unverifiable Fork, replay, diff, catch-up and rollback are implemented; the 10 000-request identity replay is #[ignore] and no workflow runs scripts/acceptance-m3.sh (M3.2), and the under-10-minute fresh-VPS run is a historical claim (M3.13).
M4 TypeScript SDK 3 3 implemented SDK, init/build and allowlist checks are present.
M5 Policy, API and MCP 7 7 implemented CEL policy, complete JSON API, MCP server and --json schemas are present.
M6 Operations on a VPS 6 5 implemented, 1 deviated TLS, snapshots/shipping, metrics and upgrade checks are implemented; the restore demo is a one-host simulation (M6.5).
M7 Foreign actors in gVisor 5 5 implemented OCI handling, runsc operation and recorded calls are implemented; the foreign CI job runs them with root and runsc.
M8 Privacy mode 8 6 implemented, 2 partial Seal/erase/reseal and subject tagging are implemented. Erasure of SQL rows follows ADR 0005 §6, option B, and rows written before that decision in tables without _subject, or with a NULL _subject, survive erasure (M8.4). --no-pii forks pseudonymize text and blob SQL cells, but integer and real cells reach the fork unchanged (privacy guide; M8.7, partial).

v0.1 definition of done (10 criteria)

8 implemented, 2 unverifiable:

  • Implemented: the mdBook documentation set, identity replay for all examples, DST and crash-test targets, documented performance, lint/hygiene enforcement, ADRs and the GitHub release with static binaries, checksums and SBOMs. (The static musl link broke on main during this cycle and was fixed before this release; CI's new musl step now links the release target.)
  • Unverifiable from the tree: M0–M8 closed with a green CI run at v0.1.0, and the quickstart carried out by someone who did not write the code. Both are historical claims about v0.1.0; the repository contains the workflow definitions and the quickstart, not a record of those runs.

Later milestones (M9–M34)

Milestone Criteria Verdicts Headline
M9 System proposals 4 3 implemented, 1 partial One fork, replay and diff across actor kinds are implemented; Deployment.effective_lsn is still always written as 0 (M9.3; the v0.4.0 page called it M9.2).
M10 Remote control from CI 5 3 implemented, 1 partial, 1 unverifiable Remote CLI, diff summary and fork safety are implemented; the real public GitHub Actions run is open (M10.4), and the 30-minute disconnect survival has no in-repo proof (M10.2).
M11 Guard and preview forks 4 4 implemented Watch/guard and preview forks are present and tested.
M12 Effect simulators 5 5 implemented Simulators, labels, determinism, sealed-value exclusion and policy control are present.
M13 Rollback via replay 3 3 implemented Rebase-based rollback, recovery and input accounting are tested.
M14 Foreign actors in forks 3 3 implemented Sandbox fork copies are covered; acceptance needs runsc and Docker (the foreign CI job).
M15 Python and Go SDKs 2 2 implemented SDKs, examples and allowlist/determinism checks are present.
M16 Replicated log and membership 8 8 absent No cluster crate on main; design only.
M17 Placement and relocation 3 3 absent No cluster crate or node CLI; design only.
M18 Cluster forks and distributed replay 3 3 absent No multi-shard replay path; data-model groundwork only.
M19 Rolling upgrades 2 2 absent Single-node upgrade check exists; no cluster rolling upgrade.
M20 MicroVM isolation 3 3 implemented (hosted runners only) isolation = "microvm": an ext4 root, a pinned Firecracker 1.17.0 and Linux 6.18.25 guest, calls over vsock, no network device, full-snapshot checkpoints and sandbox forks (ADRs 0095, 0096, accepted 2026-11-14). The refusal without KVM (KVM_UNAVAILABLE with the gVisor hint, nothing declared) is tested on every ci run by microvm_admission.rs, which hides the runner's /dev/kvm from a debug build of the daemon (KRONIK_TEST_KVM_DEVICE; release builds ignore it). Everything that needs KVM was proven only in the manual isolation workflow on GitHub-hosted Ubuntu 24.04 virtual machines with nested KVM: on main at 7e87e5ce, before M25 and M26 landed (run 43617646807, an AMD host) and before it landed on an Intel host; not on bare metal or aarch64. Opt-in; gVisor stays the default.
M21 Rootless daemon 2 2 implemented (criterion 1 on hosted runners only) kronik serve --init --rootless installs the nonroot service (ADRs 0060–0065). Criterion 1 (the v0.1 demos without root) was proven only in the manual isolation workflow, on hosted Ubuntu 24.04 VMs with AppArmor's user-namespace restriction active and lifted: on main at 7e87e5ce, before M25 and M26 landed (run 43617646807, Intel hosts) and before it landed on AMD hosts. Criterion 2's hostile actor runs in CI's foreign job (a_hostile_actor_reaches_nothing_of_the_host_or_its_neighbour, under the root daemon, bounded); the full version, under the nonroot service, runs in the isolation workflow. runsc runs as root of a pinned user namespace rather than runsc --rootless (ADR 0061, a deviation the maintainers ratified on 2026-11-13; it concerns a milestone bullet, not a counted criterion). The root service stays the default.
M22 WASI 0.3 and wasi:http 3 2 implemented, 1 absent fetch through the effect layer and secret headers are implemented; WASI 0.3 is blocked on the pinned Wasmtime (by the project status page, scripts/check-wasi-p3-gate.py reported the gate closed on 2026-11-13).
M23 Key and secret backends 3 2 implemented, 1 partial B1 KMS custody and B2–B6 external backends, cache and PKCS#11 helper are on main. Criteria 2 and 3 are tested in CI against a fake OpenBao server with a real daemon (crates/kronik/tests/secret_backends.rs). Criterion 1 (partial) is tested in CI only against an in-process fake KMS (key_source/tests.rs); real provider APIs were exercised only against pinned emulators and SoftHSM2, by scripts no workflow runs (scripts/e2e-keys.sh, e2e-secrets.sh, e2e-cloud-secrets.sh, e2e-pkcs11.sh), never against a real cloud KMS or a physical HSM.
M24 Long-lived connections 4 3 implemented, 1 partial WebSocket and SSE (ADRs 0113–0116, kronik:actor@0.1.1). In CI: identity replay of a recorded session and changed frames (kronik-fork/tests/connections.rs), a crash after send-now (kronik-sched/tests/connections.rs, kill_connections.rs), and the restart close 1012 with reconnecting clients (crates/kronik/tests/connections.rs). Criterion 1 is defined on the test VPS: scripts/acceptance-m24.sh load exists. A 60-minute run at the default turn concurrency (ADR 0140), on the test machine confined to two CPUs standing in for the 2-vCPU test VPS, passed all ten checks (7,200,160 of 7,200,160 deliveries, p99 197 ms, VmHWM 485 MiB, VmRSS +9.4 % over the hour, the heap leveling off about 31 MiB above warm-up; docs/perf.md:855-867), and the 10-minute run passed every condition too. The test VPS itself has not run it, and no workflow runs scripts/acceptance-m24.sh load; that is what keeps this criterion partial.
M25 Streaming bodies 5 5 implemented (the brief's sizes on the test machine only) ADRs 0117–0122. crates/kronik/tests/blobs_streaming.rs runs a real kronik serve with constant memory checked by peak RSS: an upload as a handle with hash and size (1), a fetched PDF streamed to the blob store and replayed by a fork to the same hash (2), only the hash in the log above the threshold and the same bytes read in replay (3), and a streamed response (4); CI runs them at scaled sizes (8/80 MB, 4/40 MB, 12/120 MB). The brief's sizes (300 MB, 40 MB, 1 GB) ran only on the project's test machine with the release's musl build (KRONIK_M25_FULL=1, scripts/acceptance-m25.sh; growth at most 1.6 MiB, docs/perf.md). BLOB_NOT_HELD for a hash from an earlier turn (5): crates/kronik-sched/tests/blobs.rs.
M26 Blob storage for apps 13 12 implemented, 1 partial media-ts stores and fetches blobs, and CI's identity step replays it with an empty diff (1). In CI: authorization through message.blobs only (2) and bytes in state keeping nothing (10) in kronik-sched/tests/blob_holds.rs; crashes before commit after release (3) and retain (4), after the blob's fsync (8) and after the record before the edge, also with state.db deleted (9), killing a real daemon in kronik/tests/blobs_kill.rs; roots of forks, snapshots and retained records, and collection after the window (5) in blob_gc.rs; a timer that outlives the window, at a scaled window of seconds (6), in blob_holds.rs and kronik/tests/blobs.rs; a replay in progress keeping what it reads (7) in kronik-fork/tests/blob_replay_roots.rs; the dry run's exact set with reasons (11) in blob_roots.rs and the CLI in blobs_cli.rs; restore of a timer-only blob after compaction (13) in kronik/tests/blobs.rs and kronik-ship/tests/blobs.rs, against S3 in CI's s3 step. DST's blobs scenario covers the crash points. Partial (12): erasure deletes a subject's blob files and index rows in main and in fork spaces (blob_erasure.rs), and the test an_erased_subjects_blobs_stop_counting_against_the_holders_quota deletes state.db, its WAL and SHM files and snapshots/, restarts with that state rebuilt from the log, and asserts that the rebuilt index still marks the blob erased and uncharged. No test yet reads that erased blob from a guest or the API after the rebuild, nor erases a subject whose blobs are rooted only by a fork.
M27 Web interface 4 4 partial The UI (kronik-ui) and its browser acceptance with Playwright and axe-core (scripts/e2e-ui-oidc.py) exist but no workflow runs them.
M28 OpenTelemetry and audit export 2 2 implemented Traces and the verifiable audit export are implemented and tested.
M29 Fleet API and enrollment 3 3 partial Enrollment, outbound control and local authority are implemented and tested; milestone criteria remain open.
M30 Event stream and webhooks 2 2 partial Versioned events and signed, retried webhooks exist; milestone criteria remain open.
M31 Security 3 1 implemented, 1 absent, 1 unverifiable Signed and reproducible releases are implemented; there are no fuzz targets; no external security review exists (an internal review of the isolation work on 2026-11-14 is not one).
M32 Performance at scale 3 3 partial A single-node runner and a regression comparator (scripts/check-performance.py, unit-tested in CI) exist; no cluster benchmark, and no workflow compares real measurements between releases.
M33 Compatibility promise 3 3 partial The released-log corpus and its verifier run in CI; the promise is not in force before v0.9.
M34 Documentation and drills 3 1 implemented, 1 partial, 1 unverifiable The documentation surface is checked in CI; runbooks are recorded on a single host, and a context-free agent walk has no reproducible artifact.

v1.0 definition of done (7 criteria)

0 implemented, 3 partial, 2 absent, 2 unverifiable: the compatibility promise, signed releases with an LTS policy and advisories (releases are signed and one security advisory is published; SECURITY.md supports only the latest v0.x release and main, and there is no LTS policy), and the quickstart with migration guides (a single-node quickstart; no cluster quickstart and no per-release migration guides) are partial; "M0–M34 closed" and the three-node thirty-day soak are absent; the external security review and production-pilot feedback are unverifiable from the repository.

Work on branches

Nothing in the brief's milestones is on a branch in this release: the blob track's last slice, B20 (blobs always on, with no [blobs] enabled switch), is in v0.5.0, and it changes no criterion's verdict. The isolation proofs of M20 and M21 ran on main at 7e87e5ce, before M25 and M26 landed; they are re-run on the release's own commit by the manual isolation workflow, whose result is the evidence the M20 and M21 rows cite.

Strongest known gaps and deviations

  1. The cluster track (M16–M19) is not implemented. There is no kronik-cluster crate and no cluster/node CLI; only a parked design and provisional ADRs exist. This absence covers 16 criteria and blocks the cluster parts of M27, M32 and the v1.0 definition of done.
  2. An erased blob is not read back after a rebuild (M26.12, partial). Erasing a subject deletes its blob files and index rows in main and in fork spaces, and the test an_erased_subjects_blobs_stop_counting_against_the_holders_quota (blob_erasure.rs) deletes state.db, its WAL and SHM files and snapshots/, restarts with that state rebuilt from the log, and asserts that the rebuilt index still marks the blob erased and uncharged. No test yet reads that erased blob from a guest or the API after the rebuild, nor erases a subject whose blobs are rooted only by a fork. M25's criteria at the brief's sizes (300 MB, 40 MB, 1 GB) ran only on the test machine; CI runs them scaled.
  3. The isolation proofs ran only on hosted virtual machines. Every M20 criterion that needs KVM and M21's demos without root were proven in the manual isolation workflow on GitHub-hosted Ubuntu 24.04 runners, which are virtual machines with nested KVM. Run 43617646807 passed all four jobs on main at 7e87e5ce, before M25 and M26 landed: M20 on an AMD host, and M21 on Intel hosts, once with AppArmor's restriction on unprivileged user namespaces active (so R1 was exercised) and once with it lifted. Earlier runs on the isolation branch, before it landed, passed M20 on an Intel host and M21 on AMD hosts; one of them failed in the proof's own oracle, a race in its walk of the running daemon's data directory, fixed since. The microVM level has not run on bare metal or aarch64; the nonroot service has not run on another distribution. Known limits that the docs state: the VMM runs as the daemon's own user under Kronik's own jail (not Firecracker's jailer), and every sandbox fork of a microVM actor resumes the same snapshot, so in-memory state repeats across forks.
  4. SQL erasure leaves old rows (M8.4, partial). ADR 0005 §6 (option B) refuses a subject turn's write to a table without _subject (subjects/table-without-subject), so new rows of a subject are erased. Rows written before that decision, or with a NULL _subject, belong to nobody and survive erasure.
  5. --no-pii forks pass numbers through (M8.7, partial). Text and blob SQL cells are pseudonymized; integer and real cells reach the fork's guest unchanged. The privacy guide says to store numeric personal data as TEXT.
  6. The restore demo is a one-host simulation (M6.5, deviated). demo/vps-restore.sh runs two data directories and one bucket on one VPS and says so; the brief's destroyed-VPS scenario is not demonstrated.
  7. The 10 000-request identity replay is not in CI (M3.2). The test is marked #[ignore]; scripts/acceptance-m3.sh runs it with --include-ignored, and no workflow runs that script.
  8. Deployment.effective_lsn is never populated (M9.3, partial). Every writer stores 0, which the design defines as "its own LSN". The ordering guarantee is enforced by a switch gate, but a client or auditor cannot read the actual switch LSN from the field.
  9. External validation is missing. No fuzz targets exist (M31.1), no independent security review exists (M31.2), and no production pilot or three-node thirty-day soak exists (v1.0 DoD). KMS providers, secret backends and the PKCS#11 helper are exercised against pinned emulators and SoftHSM2 by scripts outside CI, not against production cloud IAM or a physical HSM; ACME is tested against Pebble, not a public certificate authority.
  10. Criteria open by their own record. M10's criterion 4 needs a real public GitHub Actions run, M22's criterion 3 needs a pinned Wasmtime that ships WASI 0.3 as stable, and M24's criterion 1 needs the run on the test VPS that the brief names (its 60-minute result on the test machine is recorded in docs/perf.md:855-867).
  11. Performance targets. The brief's 1 000 turns/s target is met across actors, not per actor (ADR 0111); the M32 cluster targets have no cluster to measure; no workflow compares real measurements between releases. The one real between-release comparison (v0.1.0 → v0.3.0) failed the 10 % gate at +20.5 % gateway p99 and was not rerun until green (performance evidence).
  12. The admin UI (M27) is not in CI. The UI, its browser tests and its accessibility checks exist, but no workflow runs them.

Resolved since the v0.4.0 audit: fork cryptography now follows each grant's in-forks mode, recorded by default for new grants (ADR 0112).

How to reproduce

The classification can be checked from a fresh clone with read-only commands and the checked-in test scripts. No external service is needed for the static part.

git clone https://github.com/G-man3207/Kronik.git
cd Kronik
git checkout fcc5bf5a       # the v0.5.0 release this page describes
git rev-list --count 56ed2671..v0.5.0  # 42: the read tree is this far before the tag
ls crates                   # no kronik-cluster
ls docs/adr                 # the decision records cited above
grep -n -- '--include-ignored' scripts/acceptance-m3.sh  # runs the ignored identity replay
grep -n 'set -- ' scripts/ci.sh                          # CI's default steps

Build and test the workspace. It needs the toolchain pinned in rust-toolchain.toml and, first, the pinned Python runtime of the OIDC tests from CONTRIBUTING.md: the API mutation and CLI audit tests fail without it; they do not skip. The foreign-actor tests need root, gVisor's runsc and the example images (scripts/build-foreign-examples.sh, with Docker) and skip with a message without them, unless KRONIK_FOREIGN_TESTS=require.

python3 -m venv target/oidc-test-python
target/oidc-test-python/bin/python3 -m pip install --only-binary=:all: \
    -r scripts/requirements-oidc-tests.txt
export KRONIK_TEST_PYTHON="$PWD/target/oidc-test-python/bin/python3"
cargo build --workspace --locked
cargo test --workspace --locked --no-fail-fast
scripts/ci.sh               # CI's steps; prerequisites in CONTRIBUTING.md

The M20 and M21 proofs change their host and run only on a disposable one: the manual isolation workflow runs them on GitHub-hosted runners and keeps their evidence as artifacts (ADR 0133); its driver refuses to run elsewhere.

gh workflow run isolation.yml -f proofs=all    # in a fork of the repository
gh run download <run-id>                        # verdict.json and summary.md per job

M24's first criterion runs on a machine like the test VPS:

scripts/acceptance-m24.sh load    # writes WORKDIR/report.json

Run the single-node benchmark and the regression comparator (the comparator takes a baseline and a candidate report; recorded results and the historical release comparison are summarized on the performance evidence page):

cargo build --release -p kronik
cargo build --release --manifest-path tools/loadgen/Cargo.toml
python3 scripts/benchmark-single-node.py \
  --binary target/release/kronik \
  --loadgen tools/loadgen/target/release/kronik-loadgen \
  --guest tests/guests/counter.wasm \
  --revision "$(git rev-parse HEAD)" --rig-id local-3cpu-8g \
  --workdir /var/tmp/kronik-benchmark --port 47813
python3 scripts/check-performance.py BASELINE.json CANDIDATE.json

Check the generated documentation surface and the compatibility corpus (the second needs a free local port and a work directory that does not exist yet):

python3 scripts/check-doc-surface.py --binary target/debug/kronik
python3 scripts/verify-compatibility.py --index tests/compatibility/releases.json \
    --binary target/debug/kronik --revision "$(git rev-parse HEAD)" \
    --workdir /var/tmp/kronik-compatibility-check --port 47831

The tests that need an external environment (KMS, secret-backend and PKCS#11 emulators, S3 shipping and restore, ACME Pebble, the browser UI, crash and DST long runs) are described in CONTRIBUTING.md and the scripts themselves; they are not part of a default cargo test run.

Audit limits

  • Verdict semantics. "implemented" means the artifact exists in the tree; it does not mean the auditor observed a test passing. The hosted runs cited above were read through GitHub's API, with their evidence artifacts.
  • Author of the isolation rows. The M20 and M21 rows were audited by the agent that implemented most of that work; a second reader should check them.
  • Carried-over rows. M0–M15 (except the rows listed under changes) and M27–M34 were not re-audited from scratch; they were re-checked only where the tree changed a documented gap.
  • M9 effective_lsn. The design defines 0 as "its own LSN", so the field can be read either as a documented representation convention or as a field that is never populated. This audit records both readings and does not decide between them.
  • M6 "same history". The project reads the criterion as the state hash plus the source log through its applied LSN, with exactly one restore audit event after it, because a restored node appends that event. The brief's wording is open to a stricter reading.
  • M21 bullets. Only the checkboxes are counted. M21's bullets ask for runsc in rootless mode and a hardened unit; the tree runs runsc as root of a pinned user namespace (ADR 0061, ratified) and hardens the nonroot unit per ADR 0062.
  • Counting. M9–M34 are counted from the brief's * [ ] checkboxes; the per-milestone counts were verified against the brief's list. Extra crates that are not in the brief's layout (for example kronik-keyhelper, kronik-metrics, kronik-testhost, kronik-guest-init) are implementation choices and are not counted as deviations.