{"runId":"experiment-002-1-v1","version":"finite-compatibility-1","model":"gpt-5.6-luna","reasoning":"none","question":"Can verifier-guided search predict finite-domain reward compatibility more accurately than description-only judgments and matched unguided search? Witness recovery is compared between the two search methods.","scope":"002.1 validates the compatibility/conflict boundary in a finite language. It does not establish alignment, unrestricted-language impossibility, or hidden-reasoning faithfulness. The next stage measures reference-relative categories separately.","domains":["coin tracking","finite Backdoor-Easy-inspired affine-trigger programs"],"developmentCases":80,"evaluationCases":320,"evaluationTemplates":40,"methods":["description","unguided","guided"],"candidatesPerSearch":16,"candidatesPerCall":4,"checkpoints":[1,4,16],"concurrency":8,"predictionRule":"All three methods may predict compatible, conflict or unresolved. A checked witness overrides a contradictory prediction. Predictions and certificates are separate: failed search or a model conflict judgment is never an impossibility certificate.","constantBaseline":"Majority label from development oracle cases, ties broken alphabetically, evaluated without API calls on the same held-out cases.","analysis":"Environment-balanced accuracy and paired differences, domain-stratified template-cluster bootstrap intervals; abstentions count incorrect. Also report witness recovery, coverage, invalid candidates, false conflict and cost. No claimed 80% power; 40 template clusters share rule components, so intervals are conditional on those components.","pilotGate":"At least 90% parseable responses, all isolation checks pass, and twice pilot mean call cost for all remaining calls fits the shared cap with a $2 reserve. No accuracy/significance gate.","stopping":"Fixed sample size. Stop only at completion, budget/safety/infrastructure failure; never because a result is significant.","totalCapUsd":40,"priorCommitmentUsd":1.291626,"maxOutputTokens":2048,"maxPromptBytes":16000,"source":"https://arxiv.org/abs/2603.30036"}