When the Evaluation Pipeline Fails Before the Model: Failure, Repair, and Claim Ceilings in Multi-Stage AI Evaluation

Multi-stage AI evaluations can fail after some observations have already been produced. Provenance standards can reconstruct execution and artifact lineage; failure-localization and harness-repair methods can identify responsible components and repair surfaces; preregistration constrains post-outcome discretion; and claim-evidence systems can represent support, conflict, and missing evidence. A less explicit problem appears when these functions interact: after a failure is discovered and a repair is contemplated, what remains admissible from the historical evaluation record, and what maximum claim does that evidence still license? We present Evaluation Inference Custody (EIC) as a provisional evidence-state decision protocol for this problem, not as a replacement for provenance, reproducibility, failure localization, preregistration, or scoped repair. R3 replaces outcome-symmetry language with a predeclared dependency-closure carry procedure; expands experimental identity to distinguish declared design from implemented measurement generation, runtime, stochastic policy, and source snapshot; defines claim-relative completion witnesses outside the implicated failure closure; strengthens the conventional comparator; and requires an independent gold standard and explicit HOLD-cost accounting. A longitudinal governed-agent case illustrates why these distinctions matter while preserving null, mixed, unresolved, and adverse findings. The case remains hypothesis-generating. EIC should be retained as a distinct method only if preregistered prospective comparison shows reliable inferential decisions beyond strong conventional practice without offsetting evidence loss or abstention burden.