Self-learning: the autoresearch loop
Modelled on the autoresearch pattern: a fixed program, a fixed metric, many small experiments, a log of what was kept and what was discarded.
The loop, per flywheel
- Read the program (RUNNER.md or program.md) and the state files.
- Read the last lessons and the newest worker reports.
- Pick ONE hypothesis or improvement — the most promising untried route.
- Run it; measure against the wheel's fixed metric (a score, a rubric, a count, a live check).
- Keep it if it improved the metric; otherwise discard and log why.
- Write the result to the wheel's log and state; write any reusable lesson to LESSONS.
The loop for this library
Each library cycle:
1. Diffs the project memory notes, flywheel runners, skills and new worker reports against the last cycle.
2. Distils anything new and public-safe into SOUL, MEMORY, FLYWHEELS, HARNESS, SKILLS or LESSONS.
3. Rebuilds the library (Markdown + HTML + agents.json + llms.txt), scans every file, commits to the agents branch, verifies live.
4. Appends what changed to CHANGELOG.md.
Prefer named pointers over broad sweeps: a register entry, a catalogue reference or a cited case leads to documents; wide newspaper trawls rarely move a score.
Raw Markdown: self-learning.md · version and changes: changelog · machine index: agents.json