Repair EDCM audit defects with versioned 0.2 contracts - #75
Conversation
Codex Review SummaryThis 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. |
|
@codex review Review exact head |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 3cd47643a4
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
@codex address that feedback Repair the exact-head review findings without changing frozen canon or historical evidence. Prioritize the P1: codec-2 decode must fail closed unless each metric record contains the complete expected RoundMetrics field set with valid values; missing evidence must never become numeric zero. Then close the P2s: reject every result-contract/result-identity compartment at public intake; strip every delimiter recognized by |
|
Codex Review: Didn't find any major issues. Chef's kiss. Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
|
@codex address that feedback. Repair only the four current exact-head P1/P2 findings, preserve the historical v1 experiment identity by introducing a separately versioned v2 runner rather than rewriting history, add focused regressions, and rerun the repository's declared gates. Do not broaden scope. |
Summary
Testing
|
…dence; reconcile codec validation
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 0215a9fcde
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| transcript intake; it is not a replay of the historical 0.1.0/v1 program. | ||
|
|
||
| For the historical 0.1.0 experiment, use a separate checkout of EDCM commit | ||
| `211ce860446d3576ae2930a101eefdd07d3e3dcc` and its exact UCNS dependency. |
There was a problem hiding this comment.
When a user follows this newly added historical-replay guidance, they check out 211ce86…, but the sealed v0.1.0 report records its EDCM producer as 5d15592d509a760ff9fbcdd346a2e749537bb2ba in experiments/results/2026-07-21-ucns-edcm-v0.1.0.json:952. Because the producer commit is included in the report and its digest, the documented checkout cannot reproduce the historical artifact's identity even if its measurement behavior happens to match; direct replay instructions to the commit recorded by the artifact.
AGENTS.md reference: AGENTS.md:L63-L64
Useful? React with 👍 / 👎.
| "generated_at": "2026-08-18", | ||
| "note": "This report describes the cited EDCM state plus the current repo-owned plan. Report-only and language-boundary documentation commits are coordination metadata; they do not upgrade measurement, empirical, proof, semantic, or canon status. The current upstream UCNS lexical program selects the xkcd constrained-English 1,000-word vocabulary as lexical-floor candidate and OEWN 2025 Core as the primary contemporary lexical-semantic corpus." | ||
| "branch": "repair/edcm-audit-contracts-20260917", | ||
| "commit": "ac2f8653f95483a59624f675e03c00fbd962a7fb", |
There was a problem hiding this comment.
At the reviewed commit, ac2f865… is neither an ancestor nor the complete tree being described: it lacks the subsequently merged codec metric validation, expanded result-field intake rejection, Unicode delimiter handling, and versioned experiment-runner changes. Portfolio consumers treat source.commit as the exact repository state behind status.current_claim and delivered, so this report attributes the final 0.2 repair claims to a revision that cannot reproduce them; publish a follow-up report pinned to the complete delivered state or explicitly record all component revisions.
AGENTS.md reference: AGENTS.md:L52-L52
Useful? React with 👍 / 👎.
1 participant
The audit found six defects behind green CI: omitted transcript text, supplied rounds presented as measured output, unsupported evidence crossing optional-package boundaries, incomplete result hashes, stale construction ownership, and token metadata lost by the codec.
This repair retains full transcript source and continuations; rejects caller-supplied outputs and retired evidence at public intake; validates completed readouts; keeps unsupported geometry/factorization typed NA; hashes the complete emitted contract; preserves bone entries in codec 2; and routes active construction to Stack/UCNS.
Observable changes are versioned: package/measurement 0.2.0, candidate edcm-measurement-v2, result schema 2.0.0, codec 2. Migration guidance explains input rejection, changed parsing, historical replay, and old lossy archives. Frozen canon resources and sealed reports remain unchanged.
Validation:
3cd47643a4649987e909e4ccc9f696e5495d3885, including Python 3.11, 3.12, and 3.13 base coverage.Upgrade avenues are documented in docs/UPGRADE_AVENUES.md: source-span explanations, independent validity study, scope/order candidate, recovery stress tests, and an exact construction-receipt consumer. The portfolio report cites the implementation commit.
hmmm: empirical validity remains unresolved. The retained refusal example still misses the frozen refusal marker axis; the sealed booking candidate's sensitivity failure is preserved. No sealed holdout was retuned or promoted by this repair.