Auditing Agent Actions with Customer-Retained Evidence: Report Integrity, Outcome Verification, and Economic Design

AI-agent auditing requires distinguishing report integrity from evidence that a requested action occurred. We implement two customer-controlled HTTP components: a gateway that retains tool requests and responses, and an outcome service that binds a registered action to native Git or SQLite records. In the first study, twelve model-agent workflows generated 27 tool calls matching an independent simulated-backend ledger. A comparison of four evidence interfaces produced 176 eligible model-reviewer judgments. Verified gateway evidence identified 26 of 26 planted alterations, compared with 12 of 26 for submitted logs and 24 of 26 for independently retained ordinary logs. No interface falsely rejected an authentic control in eighteen control judgments per interface. In the second study, 26 authored backend/fault scenarios tested lost responses, retries, overwrites, stale versions, rollback and missing causal records. The outcome service distinguished action application from the current postcondition, returning eighteen correct execution decisions, eight unresolved decisions and no incorrect decisions. Retained native logs with the same reconciliation algorithm produced identical verdicts. These studies support a bounded implementation and characterize its failure behavior; they do not establish a general detection advantage for signatures. A prospective economic design separates USDG customer payments from operator capital and shared protocol control. It specifies bounded obligations, attributable violations and comparisons with conventional contracts and stablecoin collateral. Customer demand, investigation savings and native-token benefits remain unmeasured. The replication archive includes both implementations, raw observations, fault protocols and offline verification code. Working paper, version 0.6, revised September 18, 2026. This deposit includes the manuscript PDF and its reproduction archive. Project website: https://naysay.xyz/research.html