VerifierShift Restricted Evidence for Changed Agent Workflows

Downstream agent verifiers can retain complete tool responses or deliberately restrict themselves to typed predicate facts. The latter reduces raw-field disclosure but can require another observation when a policy changes. We study this tradeoff through a frozen replay of public tau2 retail tool methods. Of 40 provider-test tasks, 31 contain an executable first order-write trace; nine controlled edits yield 279 repeated task/edit pairs and 1,395 method decisions. Always-fresh, dependency-aware, sufficient-statistic, semantic-entailment and compiled-capsule methods all resolve 217 supported cases correctly and review 62 unsupported cases. Dependency-aware reuse, Z3 and compiled capsules each require 93 verification reads; retaining sufficient statistics requires 62, versus 217 for always-fresh verification. Mean retention is 505 bytes for capsules, 414 for projection and 1,394 for raw responses, including common provenance metadata. Projection therefore wins reads and bytes. Capsules retain four predicate bits without raw source scalars, but still reveal eligibility information. These results establish a restricted interface and explicit disclosure-versus-refresh tradeoff, not a superior caching algorithm, formal privacy, full workflow correctness or model-agent performance. Future work will determine whether realistic downstream disclosure constraints justify the extra refresh cost.