Qualify reuse of saved training and prediction reports
Reuse the six completed joint-adapter fixtures under all three physical models and both processing cases. Their case, parent training and child conditioning inputs must match across the earlier and direct-source studies. Verify their publication receipts and numerical environments before freezing new inputs.
For each fixture run two declared cases: all saved providers available, and the newer future report omitted. Select the coordinatewise minimum only after checking exact maps, reference/domain bindings, shared caps and conditioning. Reconstruct training rows, contract the same region, independently audit future endpoints, then check known truth. No integration, new native prediction or fit may be requested. Preserve every source study and failed label.
Separate finite allowance
- Twelve identities; 36 saved-report merge calls, zero propagation steps.
- 66 report reads; 48 provider manifest/receipt reads.
- Twelve source/reference setups and child proofs; 72 parent support checks.
- 144 training-row audits, twelve contractions and twelve truth audits.
- At most 72 one-second LP proposals and 288 helper/independent future support checks, in addition to the 72 parent checks.
- Four workers with one numerical thread each; four total threads.
- 120 seconds overall, checked cooperatively; no retries or signals.
- Zero native trajectory calls, source packet steps, compiler invocations, optimizer starts, aircraft reads or protected ESA observations.
Freeze code, exact saved files, hashes, provider provenance and environment before launch. All twelve known answers and reservations must pass. These controls qualify reuse within the supplied region; complete-group coverage, inherited fallback, the full width comparison, recording-scale adequacy and observational confidence remain separate requirements.