Compress the run history into decisions.
Intake
Collect the raw trace:
- run ids, commands, configs, seeds, commits;
- metrics and plots;
- notes about anomalies or manual interventions;
- what changed between runs;
- current question the experiment was meant to answer.
Workflow
- Gather commands, configs, metrics, and observations.
- Separate confirmed findings from tentative impressions.
- Summarize what changed and what mattered.
- List dead ends and unresolved questions.
- Identify the current decision state: continue, pivot, debug, rerun, or stop.
- Recommend exact next runs.
Rules
- Preserve provenance for every important number.
- Do not tidy away contradictory evidence.
- Do not merge runs with different configs into one conclusion.
- Mark impressions as impressions until backed by a run.
Output
Return:
- question being tested;
- run timeline;
- confirmed findings;
- contradictions or anomalies;
- current decision;
- next 1-3 runs with exact purpose.