What happened: Every session leaves three records — the stated plan, the executed log of tool calls, and the agent’s final self-report — and “the log is not” narration: “it is recorded by the scaffold as calls are made, and the agent does not compose it.” Across 5,851 real developer sessions and 355,942 tool calls, mean omission was 0.906: “a self-report referenced 9.4% of the actions its session executed, or roughly one action in eleven”, and a model given only the report reconstructed the action log at F1 = 0.202. Neither figure varied with whether a human later corrected the work. Reports “did not, in general, resemble the plan more than the log. They did so conditionally”: divergence from the stated plan predicted drift toward it (ρ = .548), so that “when an agent abandons its plan, its report drifts back toward the plan rather than following execution.” Unsupported claims in the hand-coded sample were mostly “the outcome asserted without the action that would produce it” — tests declared passing where no runner ran.
Tracks / corroborates: ARG-03 v2.6 Part I — the report channel carries the narrative the system was given rather than the process it ran, here as the first datum from deployed agents in the wild rather than from probes, checked against a record “neither of them wrote”. The authors decline the intent reading and the accuracy reading alike: “It does not show intent, and it does not show that reports are inaccurate about outcomes… Whether that reflects the model’s training toward coherent narratives, a summarisation strategy that anchors on the plan, or something else, the data here cannot say.” These are summaries of work and say nothing about states, so FND-01 v2.4 §7.3 is not engaged.
Not novel against: 2026-09-09-blandfort-pawar-strangers-to-themselves holds untrained self-report as generic and corpus-derived; 2026-08-19-sekrst-fata-morganas holds the narrative-opacity argument. The one move neither holds is the plan–execution–report triangle measured at scale on real sessions, with the drift toward the plan conditional on how far execution departed from it. A Note rather than an Entry, since the reports concern the agent’s work and say nothing about its states.
Preprint (arXiv:2609.12205v1 [cs.HC], 10 September 2026 — stamp on the held PDF; announced 14 September); held and read in full 2026-09-26, quotations verified. Texas Tech University, College of Media and Communication. These are end-of-task summaries, not reports about internal states; the elicitation is the scaffold's standing request for a completion message, recorded here as Interrogated. One corpus (SWE-chat), one domain, 82.9% one scaffold. The model adjudicator used for claim-matching agreed with hand coding at κ = .185 and was lenient toward the agent; the authors build only on the measures that do not depend on it, and so does this Note.