PFP VI: Retained Failure Evidence and Controlled Reuse This paper examines what it takes to keep rejected proposals as evidence in the Proposal Fidelity Protocol (PFP), a proposal-and-check protocol developed under the legacy name Symbol Game. A rejected proposal can explain a defect or supply a negative test, but the record must stay interpretable, its integrity must be checkable, and later reuse must not turn a rejection into authority. The paper audits the repository's refusal store at a fixed commit. The store writes JSON Lines records chained by SHA-256 predecessor digests under a header that holds the entry count and head digest, and it emits a count-and-head receipt of cached state, requiring successful disk verification before persisted-snapshot comparison. The paper states which modifications the reload checks detect, shows that a rewritten but internally consistent shorter file still passes local verification, and explains why an independently retained anchor is needed. A worked two-record example, replayed by a package test, separates local replay from snapshot comparison. The store is logically append-only through its interface but rewrites the whole file on each append. It offers no immutable media, concurrent-writer safety, authenticated provenance, or power-loss durability, and its sorted-key JSON encoding is not a full RFC 8785 implementation. The paper then specifies a bounded procedure for turning a retained failure into a proposed detector, using existing checks for benign examples, an inclusive trigger-based downgrade threshold, invalid patterns, and batch-invariant decay under fixed-parameter assumptions. A retained catalog of 61 harvested detector candidates is described as a data file only: 56 entries are marked candidate, 49 carry placeholder fixtures, and no audited code applies the fixtures. No improvement in model accuracy, cost, or discovery yield is claimed. The first-publication review distinguishes the historical audit from later source changes, adds replay checks for cached receipts and failed writes, and includes a digest-bound catalog snapshot for offline census. The teaching companion contains 24 graded problems, a worked study guide, a separate instructor key, and an exact standard-library worksheet. Returned maintenance records do not by themselves establish host enforcement.
No takes yet. Share an insight, caveat, or question.
JEREMY H. CARROLL (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: