Repository navigation
Restore optional minimum-flow audits and experiment diagnostics - #2229
Conversation
…stics Include real AFC verification and corruption-check validation.
|
The latest updates on your projects. Learn more about Vercel for GitHub. 1 Skipped Deployment
|
|
@codex review |
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. |
![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: 9c2c28fb66
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".
| if any(output == Path(p) or output in Path(p).parents for p in inputs): | ||
| raise ValueError('Audit output must not contain inputs') |
There was a problem hiding this comment.
When the standalone CLI is invoked with an output such as --output /path/to/existing-results/audits, this check passes because that directory is neither equal to nor an ancestor of any individual input file. The subsequent mkdir and writes therefore modify the results tree supplied via --source, despite the CLI promising that source is never modified; the final digest check also misses newly added files. Reject outputs equal to or descending from args.source.
Useful? React with 👍 / 👎.
In one sentence: Restore historical winding diagnostics as reusable, optional tools for fresh runs and saved solutions.
One real example: On the large free-fiber dataset, enforcing every known sign costs an additional 4.134 L1; minimizing sign deficit at the original L1 optimum still leaves three deficit steps.
Before: The extracted runner omitted sign-feasibility analysis, rounded-versus-fractional comparisons, source-AFC verification and detailed audit summaries.
After this PR: These capabilities are available through configurable audits and a standalone saved-solution CLI. Local configs and generated results remain outside the repository.
Proof: 93 tests pass. All ten benchmark solves retain identical arrays, GT metrics and optimality certificates. All seven target comparisons reproduce historical objectives and changed fiber IDs. Source-AFC verification passes on the real database underlying the five small benchmarks: 46 rows per case, exact signs/classifications, maximum distance roundoff 1.07e-7. Five independent corruptions of temporary database copies (sign, classification, distance, endpoint identity and deleted event) are rejected; original inputs remain unchanged.
Why / where this is useful: Diagnose winding disagreements and compare experiments without losing historical analysis features or modifying saved solutions.