PFP III: Admitting Failure Detectors and Testing Their Effects

This preprint describes how a failure detector should be admitted into a proposal-checking protocol and how its effects could be tested. In the Proposal Fidelity Protocol (PFP, formerly the Symbol Game), generated text is treated as a proposal whose admissibility is decided by a declared checker. A detector adds a second decision: whether a recurring pattern should change how later proposals are handled. It can prevent a known mistake, and it can also suppress a valid but unfamiliar answer. The paper proposes an admission procedure whose unit is a reproduced defect paired with valid near misses, and it separates diagnostic findings from refusals and from route verification. Four phrase detectors from the accompanying code serve as worked examples. None has been admitted, and one of them, as first written, also matched a true statement about the attractive Hubbard model. A source audit of the repository's Demo 3 host at a pinned commit describes its fuel accounting, atomic rollback, and hash-chained receipts, and shows that all three witness routes return a missing-port refusal, so no witness is certified at that baseline. A worked batch traces this refusal. The paper then specifies a randomized comparison of a frozen detector policy with matched reflection feedback, with independent route outcomes and a route-specific bound on false suppression. An appendix establishes three small exact lemmas, each checked exhaustively on finite domains by a package test: a second signal computed from the first reduces the two-signal decision table to a one-signal detector; a Hill-type quorum is a threshold on the sum of the signals, unaffected by the Hill coefficient at the half-maximal threshold and implying no number of agreeing sensors; and screening a detector against a self panel is sound for every panel only when the executed predicate implies the screened one. The appendix audits the repository against the lemmas and describes a component that states the admission record as types with derived outcomes, whose registry is not active. No live model trial, effect estimate, or reduction in hallucination is reported. The deposit contains the manuscript, a claim ledger, a small standalone Python re-implementation of the host, decision table, detectors and lemma checks with tests, a Rust port, and historical records kept with their original bytes.