Experiment Ledger
Keeps you honest: you state what you expect before you see the result. Over time this is where you find your calibration blind spots.
File
competitions/<slug>/experiments/EXPERIMENTS.md. Append-only. Newest last.
Create the file with a header if absent:
# Experiments — <slug>
Format per entry:
- **ID / date / branch-or-commit**
- **Hypothesis** — predict direction + magnitude
- **Setup** — what's different from baseline
- **Priors** — learning IDs cited
[Description truncada. Veja o README completo no GitHub.]