offspine-experiment-discipline

When to use. Any prototype/research work in a gated repo. Why: the experiments lane's value is that any session — or any successor — can re-run and extend it; an unrecorded conclusion or an unreproducible number is the failure mode.

Procedure.

1. Keep experiments off the requirement spine by contract (no registry rows; conclusions to the design doc, scripts as evidence, promotable cores named for later placement).
2. Pin inputs and seed randomness (cv2.setRNGSeed etc.); separate camera/hardware-dependent steps from headless ones and pin the regeneration commands for the former.
3. Scripts end with an explicit verdict line + matching exit code, including designed reds (INSUFFICIENT OVERLAP, "met, but saturated") — honest nonzero exits are recorded facts.
4. Record findings with numbers, negative results first-class; flag interpreted defaults as veto-able; when metrics change, keep the old number as a reference column.
5. Maintain the resume surface every session: where it stands, exact resume commands, the human's bench errands as copy-pasteable blocks with consequences stated.
6. Done when: a fresh clone can run the headless set and reproduce the cited numbers, and a cold session can resume from the surface without re-surveying the repo.

Knowledge: FIELD-KNOWLEDGE-GILBERT.md §G1 + §G4.