The proposed TH-RP-01 study repeats one fixed/adaptive procedure pair across five independently seeded histories, with three cycles per history. Both arms share a frozen model, task stream and tools; P(t) denotes a possible procedure version, not an observed improvement or an accepted edit. Development-only admission requires a valid artifact, no added critical error, compliant resources and at least 0.01 macro-edge-F1 gain on the admission partition. Rejection retains the current procedure and records the candidate, failure and resources. Equal limits per arm and history are 60 model calls, 120,000 input-plus-output tokens, 120 execution minutes and up to three agent roles. All attempts, rejected candidates and human assistance count; unknown spend halts. Both arms retain every cycle's task artifacts. After all five histories end, both arms freeze and the independent custodian evaluates the stored artifacts once on protected assessment. No protected cases, labels or final scores return to search. Five pairs support exploratory analysis only, and no experimental outcomes are shown.
