Repository navigation
ci(proofs): run the Rocq and Lean recheck legs in a digest-pinned image - #184
Conversation
dashboard-check has been red on main since b503488: TOPOLOGY.adoc and READINESS.adoc claimed 143 tests, cargo test --list measures 144. READINESS also still said T-1 is PWG-2ii; #181's receipt derives PWG-1. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GVPPsHYqVJSoPzdNYD9tRg
…e PWG-1 The follow-up #180 named: its legs for G-1, T-2, T-3 and E-1m, and the stricter recheck gate, made the receipt stale (the CI job warned). Main's run 43945986467 at e8deeaf checked the Agda cone (7 + 8, 0 postulates) and all 4 Rocq and 4 Lean legs (22 axiom reports in each prover, every one clean; 8 refusals as named). - proof-receipt.sh: source commit e8deeaf; current rows for that run; b503488's rows superseded; T-1-only recheck rows listed for T-1 only; every claim now has agda, rocq and lean legs. - The pinned reader derives PWG-1 for G-1, T-1, T-2, T-3 and E-1m; the §8.5.3 git checks hold; the receipt is fresh at HEAD. - PROOF-GRADES, PROOF-NEEDS, READINESS: census PWG-1 5 of 5. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GVPPsHYqVJSoPzdNYD9tRg
A2ML is retired; content may be deleted, never added. These [proof-grades]
fields duplicated the receipt and went stale with the previous commit:
- grade-3: per-claim :pwg in verification/receipts/januskey_praxis.deed
(derived by the pinned standards reader)
- receipt-commit: :source-commit and each (receipt :run :commit … :state)
in the same deed
- claimed-higher: the claims' :pwg and :foundation-diverse fields there
- open-obligations, first clause ("record the first green main run of the
recheck legs …"): done; the current rocq/lean receipts in the deed
No workflow, script or tool reads these keys (grep of crates/, scripts/,
.github/); dashboard-check still passes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GVPPsHYqVJSoPzdNYD9tRg
…erpolymath/januskey into claude/awesome-wright-va45xo
Codex Review Summary: ❌ FailedThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
Pull request was converted to draft
This reverts commit d1ca467.
…ws derive PWG-1" This reverts commit 87a0e49.
Owner ruling 2026-11-16 (PROOF-POLICY §8.2, "pinned, headless, isolated"): a fresh ubuntu-24.04 runner with apt and SHA-256 version pins is not enough for PWG-1; the checker must be isolated as the Agda job's is. The recheck job now runs in ubuntu:24.04@sha256:73b99939… and installs coq=8.18.0+dfsg-1build2 and Lean 4.35.0 (tarball, SHA-256) inside it. Prover versions and the gate (scripts/check-recheck.sh) are unchanged. Verified locally in that exact image: 4 Rocq + 4 Lean legs, 22 theorems each, no axioms; 8 fixtures refused as named; [check-recheck] OK. actionlint 1.7.7: clean. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GVPPsHYqVJSoPzdNYD9tRg
|
Warning Review limit reachedYou've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository. Next included review available in 29 minutes. View limit details
No actionable comments were generated in the recent review. Summary
Merge Risk:
|
…#1625) ## Summary This PR puts the owner's 2026-11-16 ruling on PROOF-POLICY §8.2 into the policy, the reader, the machine-readable PWG mirror and the shared receipt generator. §8.2 asks for a "pinned, headless, isolated" checker. The ruling says isolated means the checker ran in a **container image pinned by its digest**. A fresh CI runner with the prover pinned only by package version or checksum does not qualify, because the runner image underneath can change without notice. Before this change, the canon reader could not see isolation: it derived `PWG-1` for any receipt whose witnesses were all checked at the source commit. With this change, the reader and the generator enforce the ruling. ## Changes - **`1-formats/deed/tools/deed_lint.js` (the reader)** - A receipt may carry `:image "NAME[:TAG]@sha256:<64 hex>"`. The reader validates it against the distribution/reference grammar: - an optional registry host: a DNS name or a bracketed IPv6 address, with an optional port; - path components that are non-empty and lower-case; - a tag of up to 128 characters, which may be upper-case; - a name (host plus path) of at most 255 characters; - a lower-case `sha256` digest. - `PWG-1` is derived only when every counted witness's current receipt names an image. - A witness without an image still counts toward `PWG-3`, `PWG-2i` and `PWG-2ii`. - **`scripts/proof-receipt.sh` (the shared generator from #1620, merged in from `main`)** - A run-ledger row takes an optional 13th field: the image the checker ran in. - `--run-row … --image REF` writes that field. - A current receipt emits `:image`. - The generator derives `PWG-1` only when the Agda, Rocq and Lean runs all recorded an image, matching the reader. - The image is recorded per run, never stamped from a hook. An older run on a bare runner therefore cannot be relabelled as isolated. - **`3-practice/PROOF-POLICY.adoc`**: §8.2 adds `:image` to what a receipt records and defines *isolated*. §8.5.1 lists `:image`. - **`3-practice/proof-witness-grades.k9.ncl`**: the machine-readable mirror gains the isolation requirement and `"image"` in `PWG-1`'s `receipt_fields`. - **`1-formats/deed/tools/README.adoc`** and **`fixtures/proof-receipt/echo_praxis.deed`** are updated to match. ## Testing | Check | Result | |---|---| | `bun test 1-formats/deed/tools/` | 332 pass, 0 fail. Cases covered: <br>• a counted leg without an image derives `PWG-2ii`, and a recorded `PWG-1` is refused <br>• an image never raises a `PWG-3` claim <br>• malformed references are refused: a bare tag, an empty tag, an empty path component, a doubled separator, two tags, an upper-case path, a 256-character name, an unclosed IPv6 bracket <br>• valid references are accepted: a host with a port, an upper-case tag, no tag, a bracketed IPv6 host, a 255-character name | | `scripts/tests/proof-receipt-test.sh` | 73 passed, 0 failed. Before the generator change it was 50 passed and 19 failed, because the generator wrote `PWG-1` receipts that named no image. New cases: <br>• a run recorded without `--image` derives `PWG-2ii` and carries no `:image` <br>• a tag-only `--image` is refused <br>• each current leg names its image | | `check-proof-receipt-test.sh`; `proof-inputs-sha256-test.sh` | 20 passed; 14 passed | | `deed_lint.js --self-test`, lint of `echo_praxis.deed` | OK | | `k9-validate.sh --strict` on the PWG mirror, with Nickel 1.18.0 (the CI pin) | OK, 0 skipped. `nickel format` leaves the file unchanged | | `asciidoctor --failure-level WARN` on the changed docs | clean | | shellcheck on `proof-receipt.sh` | no new finding | ## Effect beyond this repository (after merge) - **hyperpolymath/januskey:** on `main`, T-1 is graded `PWG-1` from Rocq and Lean runs on a bare runner. hyperpolymath/januskey#184 moves the recheck job into `ubuntu:24.04@sha256:73b99939…`. A run recorded after that, with its images, restores `PWG-1`. - **hyperpolymath/secret-types:** its `PWG-1` receipt rests on bare-runner Rocq and Lean runs (tracked in hyperpolymath/secret-types#53). Under this reader it derives `PWG-2ii` until those legs run in a pinned image. - **mathesis** picks up the rule when it re-pins the reader. --- _Generated by [Claude Code](https://claude.ai/code/session_01GVPPsHYqVJSoPzdNYD9tRg)_ --------- Co-authored-by: Claude <noreply@anthropic.com>
|
…eipt branch Conflict in .github/workflows/agda-proofs.yml, job recheck, install step. Main runs the job in ubuntu:24.04 pinned by digest and installs coq by version. This branch downloads the three Rocq packages, checks their SHA-256 and installs exactly those files. Resolved to both: inside the pinned image, install the base tools, then the digest-checked download and install (no sudo, as the container runs as root). #184 carries the owner's ruling of 2026-11-16: a PWG-1 checker must run in a digest-pinned image; a fresh runner with version pins is not enough. The docs on this branch said the opposite, so they are corrected: PROOF-GRADES 'What the grade rests on', PROOF-NEEDS, and the D-2 and D-3 deviation rows, which now name the image. The superseded Rocq and Lean rows of run 43945986467 record that reason too. The reader still derives PWG-3 for the five rows. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013CuBToxr5h9catptGiTsFX
Records the merge that 41caee5 resolved by content: that commit carried the resolution of .github/workflows/agda-proofs.yml but was made without main as a parent, so GitHub still saw a conflict. The file is unchanged here. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013CuBToxr5h9catptGiTsFX



Summary
Owner ruling, 2026-11-16 (PROOF-POLICY §8.2). A
PWG-1checker must be "pinned, headless, isolated". A freshubuntu-24.04runner with apt and SHA-256 version pins is not enough: the checker must run in a digest-pinned image, as the Agda job already does.This PR does that for the Rocq and Lean recheck job, and nothing else.
This PR was first opened to record run 43945986467 as
PWG-1for all five theory rows. Under the ruling, that run's Rocq and Lean legs ran outside a pinned image, so it cannot earnPWG-1. Those commits are reverted here; they remain in this branch's history.Changes
.github/workflows/agda-proofs.yml, jobrecheck:container: ubuntu:24.04@sha256:73b999393d07f2bf65a40373778a44dc9a6d54ad7718d29247d8ff9413c4d6a8. The digest was resolved on 2026-11-16 and should be re-pinned deliberately.coq=8.18.0+dfsg-1build2, plus Lean 4.35.0 from the release tarball, checked by SHA-256. Both are the same versions as before.scripts/check-recheck.sh, is unchanged, and so are the prover versions.Testing
scripts/check-recheck.sh[check-recheck] OKactionlint1.7.7 on the workflowThis PR's own run of the recheck job is the CI test.
After this merges
agda-proofs.ymlrun onmainproduces Agda, Rocq and Lean receipts, all from pinned images. A follow-up records that run inverification/receipts/januskey_praxis.deed, and the reader then derives the grades.mainoverstates T-1 until then.main's receipt still derivesPWG-1for T-1 from run 43213340167, which ran outside a container, so T-1 is overstated under the ruling. The canon reader cannot see isolation: it reads commits and digests. The fix belongs at the source, in standards: §8.2 should state the bar, and the receipt should carry evidence that the reader checks. I'll draft that in standards.PWG-1receipt rests on non-container Rocq and Lean runs.🤖 Generated with Claude Code
https://claude.ai/code/session_01GVPPsHYqVJSoPzdNYD9tRg