Autonomous Engineering Under Bounded Authority: A Five-Pull-Request Case Study of Agent Self-Correction, Interruption Recovery, and Evidence-Governed Integration

Abstract

This archival technical report documents a five-pull-request QSOL-IBAE engineering mission comprising a defensive v0.5 prerequisite and v0.6A-v0.6D. It separates independently checked GitHub merge/release identities from mission-report attestations, offline synthetic conformance and a newly constructed finite executable recovery model. The bounded abstract checker exhaustively explores 3,240 reachable states and 7,289 transitions while checking eight selected recovery/effect safety properties. The case reports a reviewer-attribution correction, an authorization-aware GitHub publication interruption and a native-extension verification restart. It does not claim source-code-level mathematical verification, universal agent reliability, live OpenAI/GitHub activation, external ChatGPT/Work/Codex session reactivation or remote attestation. Includes PDF, Markdown manuscript, original source report, machine-readable evidence ledger, formal specification, executable bounded checker, verification results, citation metadata and SHA-256 checksums.