Retrospective Access to Execution Events in a Black-Box Language Model: Self–Observer Dissociations, External Ground Truth, and the Limits of Replay

This study investigates whether a black-box large language model can retrospectively discriminate an externally observed execution event associated with its own immediately preceding response. The central paradigm compares live SELF reports against external Observer baselines while using an interface-visible COMMAND / NO COMMAND event as trial-level ground truth. Across the main PAT conditions, live SELF reports showed strong correspondence with externally observed execution state, including balanced-state performance and resistance to deliberately contradictory false cues. External observers, given the same visible task outcome, did not reproduce the same level of discrimination. A series of negative and diagnostic controls was conducted to test alternative explanations. Transcript-ablation, decoy-marker, and observer-replay experiments were used to examine leakage, visible-record dependence, response bias, and replay-based confounds. These controls showed that explicit transcript-readable information can account for performance in some paradigms, while synthetic replay failed to reproduce the live SELF condition and was therefore classified as nondiagnostic rather than as evidence against the primary effect. The results support a reproducible SELF–Observer dissociation for retrospective reporting of externally grounded execution events. However, they do not establish introspection as the underlying mechanism. Privileged self-access remains confounded with possible ephemeral session, backend, routing, or continuity-dependent state. The paper therefore treats the effect as a behavioral dissociation and defines the platform-level experiments required to distinguish competing mechanistic explanations.