| name | research-synthesis |
| description | Compress an AI research experiment batch or branch into lessons, updated claims/questions, and a continue/stop/pivot decision. |
Research Synthesis
Use after about five serious experiments, at branch end, or when the user asks
what the research has learned.
Inputs
Read only what is needed:
RESEARCH.md- the program card of the branch being synthesized
experiments/index.yml- relevant experiment run records
- affected
CLAIMS.md, OPEN_QUESTIONS.md, DECISIONS.md
- artifact summaries if claims depend on exact metrics
Synthesis Questions
Answer:
- What did this branch test?
- What facts are now established?
- Which hypotheses failed?
- Which claim changed most?
- What should stop?
- What should continue?
- What is the next sharper question?
- What should be compressed before more experiments run?
Also glance at repo health while here: stale running experiments, a dashboard
turning into a diary, or a failed mechanism reappearing under a new name.
Required Output
Update or create:
- a
research-log/YYYY-MM-DD-branch-synthesis.md note;
- the program card (evidence snapshot, primary claims, status, next move);
CLAIMS.md;OPEN_QUESTIONS.md;DECISIONS.md with one continue/stop/pivot decision if behavior changes;RESEARCH.md dashboard.- scorecard
Status: lines if the next experiment or priority changed.
Discipline
Prefer one strong conclusion over many loose observations. Mark stale claims or
questions superseded instead of leaving them quietly misleading.
If the synthesis mostly adds chronology to RESEARCH.md or CLAIMS.md, stop
and compress instead. The dashboard should point to synthesis notes; it should
not become the synthesis note.
