Problem
PR bodies written by the autonomous implementation loop routinely carry gate-verdict: accepted @ <sha>
and Exact head: <sha> values that are not objects in this repository. The SHAs are not stale —
they never existed. They share a 8–9 hex-character prefix with the real head and then diverge, which
is the signature of a fabricated suffix rather than a superseded push.
Measured on 2026-08-18 over the 40 most recent open PRs:
Two more (#7102 af658fb5ca1338d…, #7104 7cf5a91d374499d…) were found and corrected in place
during PR resolution on 2026-08-18; #7102 additionally had the same fabricated SHA repeated across
four separate ll7 audit comments claiming exact-head validation against it.
Why it matters
scripts/dev/merge_queue_gate.py treats an exact-head gate-verdict: accepted @ <head> as the
trusted human/reviewer attestation. The gate itself is safe — it fails closed when the trailer SHA
does not equal the live head, so a fabricated SHA cannot admit a PR. The damage is upstream of the
gate:
- Reviewers and the maintainer read "validated at exact head
X" for an X that does not exist,
so the recorded validation evidence is unverifiable and, as written, false.
- It burns gate cycles: every such PR is
pending_gate_verdict and needs a fresh review comment.
- It contradicts the repository's hard rules in
docs/maintainer_values.md (be honest,
transparent, reproducible) at exactly the surface those rules exist to protect.
Suggested fix (fix-or-track)
Add a cheap fail-closed check so a bad SHA cannot be written or can be detected immediately:
- In
scripts/dev/gh_pr_body_rest.py, reject any gate-verdict: / base-policy: / Exact head:
40-hex value in the incoming body that does not resolve via git cat-file -t and does not
equal the PR's live headRefOid. This is the narrow, high-leverage spot — every body writer goes
through it.
- Optionally extend
.github/workflows/pr-body-contracts.yml to run the same assertion, so already
written bodies are caught.
- Backfill: the six PRs listed above still carry the bad trailer.
Claim boundary
Diagnostic-only tooling/provenance defect. No benchmark, planner, scientific, or paper-facing claim
is involved. Evidence tier: reproducible measurement over live GitHub state at 2026-08-18, script
inlined above (gh pr list --json number,headRefOid,body + git cat-file -t).
Problem
PR bodies written by the autonomous implementation loop routinely carry
gate-verdict: accepted @ <sha>and
Exact head: <sha>values that are not objects in this repository. The SHAs are not stale —they never existed. They share a 8–9 hex-character prefix with the real head and then diverge, which
is the signature of a fabricated suffix rather than a superseded push.
Measured on 2026-08-18 over the 40 most recent open PRs:
gate-verdict: accepted @ <40-hex>trailer in the PR bodyheadRefOidgit cat-file -tcannot resolve: fix(training): bind optimization smoke arms to matched order (#7254) #736474c9856d8ffc…, fix(workflow): harden stale-base observation provenance (#7261) #7336e35c236ea245…,docs: align evidence template vocabularies #7307
81718057c793…, fix(benchmark): require staged radius sweep checkpoints #726050c21bca33a7…, feat(benchmark): prepare legacy checkpoint cutover parity #7236c91587c2d9e0…, docs(research): record representative selector contract #7152c978918e2c2f…Two more (#7102
af658fb5ca1338d…, #71047cf5a91d374499d…) were found and corrected in placeduring PR resolution on 2026-08-18; #7102 additionally had the same fabricated SHA repeated across
four separate
ll7audit comments claiming exact-head validation against it.Why it matters
scripts/dev/merge_queue_gate.pytreats an exact-headgate-verdict: accepted @ <head>as thetrusted human/reviewer attestation. The gate itself is safe — it fails closed when the trailer SHA
does not equal the live head, so a fabricated SHA cannot admit a PR. The damage is upstream of the
gate:
X" for anXthat does not exist,so the recorded validation evidence is unverifiable and, as written, false.
pending_gate_verdictand needs a fresh review comment.docs/maintainer_values.md(be honest,transparent, reproducible) at exactly the surface those rules exist to protect.
Suggested fix (fix-or-track)
Add a cheap fail-closed check so a bad SHA cannot be written or can be detected immediately:
scripts/dev/gh_pr_body_rest.py, reject anygate-verdict:/base-policy:/Exact head:40-hex value in the incoming body that does not resolve via
git cat-file -tand does notequal the PR's live
headRefOid. This is the narrow, high-leverage spot — every body writer goesthrough it.
.github/workflows/pr-body-contracts.ymlto run the same assertion, so alreadywritten bodies are caught.
Claim boundary
Diagnostic-only tooling/provenance defect. No benchmark, planner, scientific, or paper-facing claim
is involved. Evidence tier: reproducible measurement over live GitHub state at 2026-08-18, script
inlined above (
gh pr list --json number,headRefOid,body+git cat-file -t).