fallow-rs Panel review loop
The real-world corpus is intentionally untracked. Before its first use, follow
the benchmark setup and run
npm --prefix benchmarks run download-fixtures.
- Define the user-visible surface and success criteria.
- Select representative projects from
benchmarks/fixtures/real-world/. - Capture actual output for each project.
- Run
panel-reviewon the evidence, not on an intended answer. - Implement the smallest coherent improvement that addresses consensus.
- Re-run the same corpus and compare behavior.
- Stop when the panel has no blocks and further changes do not materially improve the surface.
Keep the corpus stable across iterations. Preserve output contracts unless the plan explicitly approves a versioned change.