Verification Allocation in Inherited Agent Memory: Provenance Availability Is Not Provenance Use

A memory system that gives every record a provenance link does not thereby ensure the corrective record is consulted. An agent inheriting more consolidated beliefs than it can re-derive must choose a few links to follow before it acts, so the reliability question is not whether provenance exists but which links get followed. We study that allocation directly, across ten experiments whose confirmatory contrasts were prospectively pre-specified before their first model call, 16 models from 11 organizations, and one prospectively specified replication run on disjoint seeds, opened only after the primary analysis was committed. Exogenous working-plan assignments causally redirect allocation: assigning a plan at random before the memories are shown, with the action field removed from the response schema so the agent cannot narrate its choice, moves the single verification credit by 84.7 points. The intervention identifies the assignment itself: it does not isolate internal plan state from topical salience or directive framing. Allocation also depends on what the memory says, though not as we first reported: a prospectively specified length-matched control left our earlier claim that generic hedging attracts scrutiny unsupported (and it is withdrawn), while a genuine decision-relevant constraint suppresses verification by -46.0 points at fixed length. A factorial holding numeric content and clause structure constant puts that effect at 31.9 points over 1440 episodes, positive within both registers and in all six wording families -- though its magnitude varies by an order of magnitude across models, so we claim the direction and not the size. Randomising which provenance records reach a later agent moves the corrupted-direction rate from 139/150 to 3/150, so what the agent holds causally changes what it decides. A boundary result bounds all of this. The benign 6-generation consolidation pipeline we tested never produced the complete deletion of negative content our controlled experiments install (0/360 chain-generations) -- although scope restrictions declined monotonically within that horizon and explicit prohibitions were frequently lost. The allocation dependence is demonstrated; the natural prevalence of the corruption that would make it harmful is not. We argue that verification allocation deserves treatment as a control problem separate from retrieval and from provenance storage. Changes in this version (v2): This version incorporates prospectively specified follow-up experiments that materially refine the interpretation of the original results. It withdraws the earlier claim that generic hedging attracts provenance scrutiny, which a prospectively specified length-matched control left unsupported; adds a randomized working-plan assignment experiment identifying the upstream causal effect, with its confounds stated; controls the semantic-constraint effect for register and quantitative specificity and adds a prospectively specified disjoint-seed replication run; reports model heterogeneity alongside every pooled estimate; separates the hand-installed corruption used in controlled experiments from what the tested benign consolidation pipeline actually produced; and sharpens the threat-model boundary accordingly. The research question is unchanged. Appendix E lists each withdrawn interpretation with the test that overturned it.