The scripts/ folder is now organized by function instead of keeping every entrypoint flat at the top level.
scripts/run/: experiment launchers and shell orchestrationscripts/plot/: figure generationscripts/analysis/: summary and paper-facing analysisscripts/precompute/: direction extraction and preprocessingscripts/eval/: judge and evaluation entrypointsscripts/merge/: worker/shard merge helpersscripts/scoring/: rescoring utilitiesscripts/infra/: Modal/Lambda/remote orchestration helpersscripts/data/: dataset build/prep scripts
Flat historical entrypoints such as scripts/run_phase0_multimodel.sh and scripts/plot_exp7_tier0.py are kept as symlinks where practical. New commands should prefer the grouped locations, for example:
bash scripts/run/run_phase0_multimodel.shuv run python scripts/plot/plot_validation_tier0.pyuv run python scripts/precompute/precompute_directions_multimodel.pyuv run python scripts/infra/repo_doctor.py
For a lightweight project-scoped validation pass, run:
uv run python scripts/infra/repo_doctor.py
Optional:
uv run python scripts/infra/repo_doctor.py --pytest
This checks:
- Python syntax under
src/andscripts/ - shell syntax for launchers
- broken script and result symlinks
- a few canonical
--helpsmoke entrypoints
- Experiment map:
docs/EXPERIMENT_REGISTRY.md - Top-level overview:
README.md