Practice
House rules for how experiments are planned, scored, and filed
- 1Filed before it runs. Hypotheses and decision rules go into the research log ahead of any compute.
- 2Separate scorer. The model behind a headline figure never chooses outputs or trains anything.
- 3Stop, don't guess. Analyzers halt on missing data; corrections are new log entries.
- 4Bounds on every figure. Paired differences carry bootstrap intervals and exact tests.
- 5A complete record. Code heads, data hashes, environments, seeds, and GPU-hours are kept for each phase.