Audit #695: selected #626 provenance digest differences are timing and directory order, reproduced byte for byte

[Rohansguliani]

What this adds

An independent audit of PR #695's three selected #626 provenance digest differences (template_sha256, input_sha256.owner_audit and input_sha256.fresh_invoice). #695 reports them as unexplained because CI run 38317651752 did not upload the payloads. No new κ, no certificate change.

Verdict: the facts that #695 and its triage report are correct. The digests differ, and the CI payloads are missing. The differences do not indicate a different template, owner audit or invoice. A fresh cold Linux replay gives a third set of values. After substituting only (1) the file digests of four generated files whose only volatile field is wall-clock seconds (CLIFFORD-REVIEW.json and SCALAR-FRAME-AUDIT.json for providers ck and jp8), and (2) the iterdir() key order in the invoice's provider artifacts maps, the replay reproduces the committed template_sha256 (22b41c45…), then OWNERS-AND-PROFILE.json (ec7bc415…), PREFIX-AND-INVOICE.json (367a73f7…), and finally the committed complex ASSEMBLY.json, all byte for byte. The template carries the timing through sha256(PLAN.json), and the runtime PLAN pins the CLIFFORD-REVIEW.json digests. The package's own timing-stripped semantic hashes already agree. This is not gzip metadata; the template digest is taken over decompressed JSON.

Limit: the CI run's specific values are attributed by mechanism, not reproduced, because its payloads were not uploaded. Every replay produces new values through the same path. research/pr695-digest-audit/README.md has the table, file and line citations, and scope.

Also checked at #695 head e322f2f1d9cc0f830cb3a50c6e928e0a1a32383c: its checker passes normally and under -O, and its report equals its RESULT.json. Its archived ZIP matches the live GitHub artifact (1f1ad114…). Its archived log matches the live job log, apart from gh rendering ESC as ^[ on 7 lines.

Reproduce

Exact base: 66849e1475c930fd7732b1a0067ffee7e1e95f60.

python3 -I -B research/pr695-digest-audit/verify.py --work /tmp/pr695-digest-audit

Expected: PASS_PR695_DIGEST_DIFFERENCES_EXPLAINED 3 digests reproduced, 2 rejection controls. When the work directory is absent, this first runs the cold scripts/verify_frontier626.py replay, which needs the workflow's environment (NumPy, SymPy, SciPy, CXX, BOOST_INCLUDE). It took 31 minutes here with --jobs 6 on Linux, CPython 3.13.7 and NumPy 2.5.4. The comparison passes normally and under -O.

Credit and disclosure

This builds on #695 by Joshua Snider (jsnider3), which preserved the CI artifact and recorded the differences explicitly. The supplier and review lineage is as credited there and in the selected review: Douglas Colkitt, Chafik Boukhalfa (#626), Jeremy Kothe (#622/#623), eumemic (#543) and GamingPuzzled (#620). The container-reproducibility convention cited in the README is by Maxime Fleury and Antoine Olivier.

Audit by Rohan Guliani (@Rohansguliani), with substantial Anthropic Claude assistance. Apache-2.0. This is not independent human peer review, and it is not a formal verification of the multiplication theorem.