PR #1074 turned a response-time version string into a persisted contract for simulation semantics. It is a useful quality-gate record because both mechanisms rejected apparently passing implementations for concrete, inspectable reasons.
The implementer and reviewer were both Codex-family agents running in separate contexts with different roles. The reviewer received the exact base/head diff, the committed spec, read-only repository access, and authoritative gate results.
Test catch
On test attempt 2, one test failed while 4,923 passed:
tests/test_assumptions_version.py::
test_current_manifest_is_well_formed_and_matches_governed_semanticsThe governed fingerprint had changed without the corresponding manifest update. Later attempts passed after the implementation was corrected. This is a narrow CI tripwire, but it did exactly what it was designed to do: prevent an unacknowledged change to governed semantics.
Review catches
The independent reviewer requested changes three times and reported four P1 findings:
- The semantic fingerprint covered a facade but omitted the actual simulation core and country, cost-of-living, and tax-policy modules.
- Deterministic replay did not preserve the source run’s assumptions-version provenance.
- MCP cancellation returned a null, unverified receipt for a newly versioned run.
- Currency normalization and then the broader server-side input transformation path were absent from the governed fingerprint.
The fifth review approved the corrected implementation. All four finding bodies, file locations, reviewed SHAs, and decisions are preserved in the artifact.
Run record
| Field | Value |
|---|---|
| Size | 456 changed lines across 13 files |
| Implementation and review | Codex implementer; separate Codex reviewer |
| Iterations | Four implementation attempts |
| Rejections | One failed test and three rejected reviews |
| Human intervention | None recorded |
| Outcome | Passed and merged |
Source material
- Complete committed spec
- Sanitized PR description and file ledger
- Focused semantic-contract and test diff
- Gate history
- All preserved review decisions and findings
- Exact final independent-review prompt
What it demonstrates
The aggregate “review or tests caught something” rate is not merely a count of formatting objections. In this run, the test guarded a machine-verifiable contract while the reviewer identified missing behavior across replay, cancellation, and transformation paths that the passing suite had not fully covered.
The run also shows the value of role and context separation within one model family. A fresh reviewer context found four material omissions that the implementation context and automated gates had not resolved.