Before Deliberation Closes: The Pre-Deliberation Horizon and Invocation Problem for Reflective Restraint

An accessible argument can fail to matter because it never enters the relevant process, arrives after the relevant control opportunity closes, or enters in time without changing the protected outcome. This paper formalizes those distinctions for reflective inputs in distributed or agentic decision pipelines. It defines criterion-relative control windows, restricted AND/OR completion recurrences, causal noninterference conditions, trace-relative bypass defeat, and a bounded completion lemma for instrumented pipelines. The human-protection layer separates invocation-concurrent protection from switch-time policy-effect attribution and assignment-level input attribution. In the deterministic worked example, all four comparison laws—target policy, baseline policy, input-treated, and input-control—are derived from one regime evaluator rather than independently stipulated. Recorded execution evidence is aggregated with negative-preservation precedence, so an invalid trace reference cannot mask an established path failure. Active revision guards are anchored to their operative sections, predecessor guard sets remain monotone, and a whole-manuscript audit accounts for substantive deletions. These mechanisms protect internal consistency and revision continuity; they do not establish behavioral effectiveness or AI safety.