Reusing Evaluator Evidence under Mutable State: Transactional Boundaries and Two Controlled Cost Studies

Reusing computational evidence does not renew permission to commit an effect. We study this distinction in TGE-R1, a local SQLite prototype, through two separately versioned controlled studies. Study I matched the frozen expected outcomes in 96 scenario combinations and 20 receipt combinations through the measured request path, then completed 9,216 requests. Its reuse strategy was faster in all four within-session pairs in 12 cells with reuse windows; per-request changes produced mixed directions. However, re-evaluation also retained more evidence for subsequent validation, so these timings compare bundled strategies rather than isolated computation savings. A separate matched-retention follow-up therefore shares pure support content while preserving distinct, truthful acquisition and action histories. Its own qualification preceded one 2,048-request M2 measurement. Across 5,136 paired audit windows, common object roles and traversal order matched. For 65,536-element inputs with stable state, reuse was faster in all four pairs, with a median ratio of re-evaluation time to reuse time of 1.429 within the instrumented-stream boundary. The other three cells each split two pairs in either direction. The follow-up shows that unequal pure-support retention is not necessary for the benefit observed in that comparison. It does not isolate evaluator CPU cost or permit subtraction of results across versions. Both studies use synthetic workloads and one measurement session per version, with no external replication. The contribution is a version-bound methodological case study linking qualification, truthful provenance, comparison design, and limited cost observations, not a new caching or authorization primitive.