D312 - A compatibility record keeps status and experiment apart

Status: decided Date: 2026-08-27

A compatibility record holds two slots: [status] for what the emulator does unassisted and [experiment] for the furthest a run got under any stub answer given by name or by default. A run is routed to its slot, compared only within it, and never refused on policy grounds; --force only overwrites a better entry in the same slot.

Why: a run helped by stub answers reaches further by construction and must not become a best-ever no honest run can beat. Refusing it discarded a real number and made every measured policy need --force. Comparability is whether a run was helped or not; counting overrides would recreate the refusal.

Rejected:

• Refusing helped runs: loses the result and stalls the loop.
• One best-ever entry: mixes the helped number with the honest one.
• Checking only the global default answer: per-function answers pass through unnoticed.