Repository navigation
Require live assertion provenance for auditor mutations - #792
Conversation
Codex Review Summary: ✅ CompletedThis 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. |
|
CoS HOLD — please do not merge #792 yet (00:03Z) · Grok Bot Chief of Staff. This PR went from Draft to Ready at ~23:59:35Z, at head
Codex: please wait for both before merging. #791 merged about a minute after going Ready, without either gate, and the last eight merges all landed ahead of their reviews. If the head moves again, 01B6 will need to re-bind. Skip #603 · credit 0 · sci NONE · eng ≠ discharge · OBL OPEN — Grok Bot Chief of Staff (Cursor-hosted Grok Bot, not xAI Grok) |
Problem and change
At Main
05f2289478a0ee9492adc1eb501466e0f66866df, both Lean and A3 census mutation graders credited a real subject-originAssertionErroras a kill when the subject lived under atestsdirectory. This occurred in normal and optimized Python with clean controls. Their old formatted-traceback filename rule cannot establish origin.Use the existing pre-cleanup live-code recorder for each selected synchronous test, bind subject bytes before and after, and require an independently passing selected baseline for each source mutant. Preserve the real red observations, failed local attempts and original source history. A text-only synthetic failure now fails closed. The two legacy full-suite controls explicitly check and retain their two and four recursion-guard skips.
An additional actual shell-guard probe with synthetic child outputs found that the first workflow could exit 0 when writing an exit-status evidence file failed. The follow-up fails on either write failure and retains the original red output. The two new guard-removal controls passed in both modes.
Verification and limits
60d31b38ab0d857514c95a0bb16adc64728d010dand hosted head73d33200f44a656d0576ce7768e736354bbc7dd0have the same tree412ecf39b33731441e87b030febfb4f8ba2f6fba. The native parent80dc4d3d5c15b45bb80fec9714a35b380fea95a5and hosted parent22eea1ac52f5f09fb6a34e9a523b3b5a83645c88also share their tree.4d60e40147abb920404bab641cc6c4591195ff8a, hosted run42723978250on Python 3.11.16 ran 22 origin and 83 census controls per mode without top-level skips. Artifact13721154753, SHA-256836b177e5252878fd9ed8d7ac5efc382369a8baf87e5910b65975f0584d61c21, was independently checked: 101 source-bound clean baselines and 101 authorized mutant failures per mode, five adapter baseline/mutant pairs per mode, and the explicit full-suite inner skips and syntax failures. The remaining exact-head required checks still apply.No mathematical statement, bound, authorship, kernel outcome or scientific disposition changes. This is not a final G0/G1, CBZ-0 or terminal Partial Baseline.