# AlphaEvolve AQ04 / MCGS sharing contract

- Experiment: `alphaevolve-aq04-muse-20260923` (existing ID and URL preserved).
- Public Bucket: `jzshared/autoresearch-runs`.
- Owner/coordinator: `liux`, Trixie Slurm service `910982` (replaces 910979).
- 15 tasks × 2 methods = 30 runs; methods `AQ04` and `MCGS`, seed 42.
- AQ04 run IDs: `liux.aq04-bo-20260923T213753Z.<task>`.
- MCGS run IDs: `liux.mcgs-20260923T202917Z.<task>`.
- Snapshots: `experiments/<experiment>/runs/<run_id>/live.json`.
- Full text logs: the same run prefix, under `artifacts/`, with an artifact manifest.

Metrics and Space data refresh every 60 seconds; logs sync independently every 300 seconds.
The current service allocation expires at 2026-09-24T06:08:33 America/Toronto,
or stops earlier after final publication. Static files remain readable afterward.

Native sources use report.json, run-info.json, logs/journal.json and search_budget.json.
This is the generic hf-experiment-share snapshot contract, not its bundled formal_mila layout.
Only finite, valid candidates admitted within the search budget enter the curves.
Time curves use actual candidate admission times and retain history from before the uploader joined.
Per-run source, scheduler observation, upload and publication timestamps remain distinct.
The two groups started at different wall times; compare curves at the same search duration.
Final mathematical regrading is separate from search scores.

One coordinator owns this Space; do not start another. Preserve IDs/cache across uploader restarts.
Only write your own unique run prefixes. Credentials, copied inputs and caches are excluded.
Intended text logs are redacted before upload; frontend baseline filtering leaves raw records intact.


AQ04 best-result update (2026-09-25): selected nibi paid-only reruns use distinct run IDs `liux.aq04-nibi-paid-20260924T162329Z.<task>`. The table and curves select the better original/rerun objective, with matching logs. `aq04_best_runs.json` preserves these snapshots when the existing coordinator refreshes other methods; `aq04_best_selection.js` applies the same selection on each refresh. Original snapshots are retained in `archived_runs` and original Bucket prefixes. The 8-hour budgets were reached, but terminal Slurm states remain FAILED because of finalization errors. No independent final regrade is claimed. The uncertainty improvement is only about 4.1e-14 (floating-point scale).