Evaluation Records
This directory contains curated, immutable milestone reports rather than every benchmark run. Each committed report contains the human-readable evidence needed to understand the result:
- original answer-bearing skeletons;
- rendered problems;
- complete solver outputs and parsed answers;
- scores, run metadata, and qualitative anomalies.
Raw JSONL from ordinary runs belongs in the Gitea Actions artifact, not Git. Promote a run into this directory only when it establishes a baseline, changes the benchmark protocol, exposes an important failure mode, or is explicitly selected for long-term comparison. This keeps repository growth proportional to meaningful experimental milestones instead of CI volume.
Runs
| Date | Run | Task | Cases | Accuracy | Commit |
|---|---|---|---|---|---|
| 2026-09-22 | Gitea Actions #90 | short-smoke |
5 | 100% | fab07f8 |
Model identifiers were not emitted by run #90 and therefore are intentionally recorded as unknown rather than inferred. Future artifacts should capture the renderer and solver model IDs as non-secret metadata.