Skip to content

Require the promotion-gate result to report every named gate check (SYN) - #77

Merged
seathatflowsinourveins merged 5 commits into
mainfrom
claude/promotion-gate-completeness-20260922
Sep 23, 2026
Merged

seathatflowsinourveins merged 5 commits into
mainfrom
claude/promotion-gate-completeness-20260922

Conversation

@seathatflowsinourveins

Copy link
Copy Markdown
Collaborator

Summary

  • A blind Codex review of Follow-up: resolve the Codex cross-family findings on PR-3 and PR-4 #64 (relayed by peer session finalize-llm-harness-catalogs) found a fail-open in runner._check_promotion_gate. A gate result with checks=[{status:'pass'}], row_count=1, versions=null, checked_at=null and the matching hash of a header-only CSV passed the paper preflight.
  • The preflight now requires:
    • the result's checks to name every promotion_gate.CHECK_NAMES entry exactly once, with no missing, extra, unnamed or duplicate checks;
    • versions to be a non-empty object;
    • checked_at to be an ISO-8601 timestamp.
      Anything else raises promotion_gate_incomplete. The failed-check and empty-row checks are unchanged.
  • The test helper _full_gate_result now emits the gate's full check list.
  • test_runner_check_names_match_the_gate_module parses promotion_gate.py, so the runner's list cannot drift from the gate's.
  • The gate-side fixes in promotion_gate.py are the peer's separate branch.

Evidence (SYN)

  • 6 new tests. PromotionGatePreflight + MainCommandGateWiring + OvernightHoldPreflightScope: 40 OK.
  • Mutation check: with the name-set guard removed, 3 new tests fail. The blind reproduction is also caught by the null-versions guard.
  • The real 2026-09-22 SIP snapshot gate result (480 rows, 12/12 checks) and fixtures/good-gate-result.json are still accepted.
  • python3 -m unittest: 2043 OK (324 skipped). Engine interpreter: 2043 OK (7 skipped).
  • All validate.yml steps and trading_gates.py --check rc=0.
  • Guarded gitleaks ran before the commit: no leaks.
  • Receipt block promotion_gate_completeness added; source-hashes, manifest and explorer repinned.

🤖 Generated with Claude Code

Blind Codex review of #64: the paper preflight accepted a gate result with an
unnamed passing check, null versions and null checked_at. It now requires the
checks to name every promotion_gate.CHECK_NAMES entry exactly once, a
non-empty versions object and an ISO-8601 checked_at (else
promotion_gate_incomplete). A test parses promotion_gate.py so the two lists
cannot drift. The real 2026-09-22 gate result and the good fixture still pass.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
… row binding

Review follow-ups on eb2e9c8: a non-string check name is refused cleanly
instead of raising TypeError; checked_at must be a timezone-aware ISO-8601
date-time; version values must be non-empty strings; for a .csv snapshot
row_count must equal the file's data rows; an unmapped_failures entry from the
gate raises promotion_gate_unmapped_failures (peer suggestion). Mutations of
the name-set guard and the row binding each fail their tests.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
Second review of the gate check: the CSV row count re-read the snapshot
outside the OSError guard; hash and count now use the same bytes. Receipt
limits name the quoted multi-line CSV false-refusal case.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
…-completeness-20260922

# Conflicts:
#	docs/ecosystem/index.html
#	manifests/evidence.json
…-completeness-20260922

# Conflicts:
#	docs/ecosystem/index.html
#	manifests/evidence.json
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant