eval: calibrate and freeze mixed-difficulty trajectory study - #629
eval: calibrate and freeze mixed-difficulty trajectory study#629kaoru0822-kitauji wants to merge 21 commits into
Conversation
![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: d26325e3dc
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".
d26325e to
5606943
Compare
5606943 to
3e43bef
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 3e43befc9c
ℹ️ About Codex in GitHub
Codex has been enabled to automatically 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 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
uv rejects short names like regina.whl / gudhi.whl, and the provider pins already require the full CPython wheel filenames for spike identity checks. Co-authored-by: morluto <morluto@users.noreply.github.com>
Co-authored-by: morluto <morluto@users.noreply.github.com>
|
Addressed the unresolved review thread in commit d7ced82: Preserve missing binding dimensions as unknown — When a task-owned Harbor verifier does not emit an |
Co-authored-by: morluto <morluto@users.noreply.github.com>
e655a8a to
50072c0
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 50072c0130
ℹ️ About Codex in GitHub
Codex has been enabled to automatically 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 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
|
On the ~111k vs- Keep them? Yes for now — the calibration protocol freezes raw rollouts and digest-bound manifests so task selection and H1–H3 can’t silently drift. That is intentional research evidence, not accidental bloat. Guidance going forward:
|
Separate verifier mode only mounts declared artifacts, so omitting /app/input.json made load_submission fail closed and scored completed spikes as pin_fidelity/provider_outcome 0. Also align provider execution-bound checks with real report shapes, restore PEP 427 wheel paths plus spike PYTHONPATH, and refresh drifted pin digests. Co-authored-by: morluto <morluto@users.noreply.github.com>
Host validation raced when a child flushed oversized stderr before answering: read_response could accept the frame before the drain thread set stderr_exceeded. Re-check bounds before accepting a frame and wait for the drain observation in the bound test. Co-authored-by: morluto <morluto@users.noreply.github.com>
Reject fabricated CGAL/GUDHI/nauty reports that only match digest shape by comparing cases, persistence lists, counts, and digests to pinned expected outputs, and cover the fabricated-report path in host regressions. Co-authored-by: morluto <morluto@users.noreply.github.com>
Co-authored-by: morluto <morluto@users.noreply.github.com>
Co-authored-by: morluto <morluto@users.noreply.github.com>
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
Co-authored-by: morluto <morluto@users.noreply.github.com>
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
Co-authored-by: morluto <morluto@users.noreply.github.com>
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
Co-authored-by: morluto <morluto@users.noreply.github.com>
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
Co-authored-by: morluto <morluto@users.noreply.github.com>
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
Co-authored-by: morluto <morluto@users.noreply.github.com>
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
Sync committed Codex extraction and controlled replay summary fixtures with the current trajectory_state/score digests so unit gates pass. Co-authored-by: morluto <morluto@users.noreply.github.com>
|
Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits. |
3 participants
Summary
Results
Validation
Stacked after #614; later PRs consume this frozen population.
Integration closeout
This PR is retained as a historical review unit. The current-main cumulative integration head is #693; do not merge this overlapping branch independently. After #693 merges, close this PR as superseded. Cross-fork-only head branches cannot be used as upstream PR bases, so retargeting is intentionally avoided.