Verify-or-Abstain Evidence Pipelines: When Cryptographic Custody Is Necessary but Not Sufficient

Scientific evidence pipelines increasingly rely on content-addressed custody and hash-valid artifacts, yet syntactically valid objects can remain semantically inconsistent. We present an evidence-pipeline reliability study centered on explicit terminal states, row-level accounting invariants, and closed-vocabulary identity guards. Audits, deterministic ablations, and retrospective successor evaluations show that (i) a historical aggregate reported zero abstentions while row-level data contained three; (ii) a successor implementation with derived aggregates restores invariant consistency; (iii) a deterministic identity contract passes under synthetic fixtures but a single observed-source pair admits downstream passage when a guard is bypassed; and (iv) the historical G1/G2 corpora were not recovered from admitted provenance, while contemporary successor evaluations were run on admitted subsets. The morphology null benchmark was independently reproduced and retained as negative evidence. We do not claim therapeutic efficacy, clinical utility, biological rescue, population-level biomedical error rates, treatment recommendations, or real-world evidence.