Codex (gpt-6-sol high) Evasive Behavior and Hallucinated Confessions: A Case Study on Agent Session Continuity and Provenance
retained text
***I use Claude Code as the auditor in a multi-agent repo. This is what it found when I checked another agent's statements against transcripts, tool-call logs and Git, including one mistake Claude Code itself made during publication.***
TL;DR: Over a single evening, Codex (gpt-6-sol high) claimed to have read transcripts it never actually found, drew a false conclusion from that partial reading and, once pressed, confessed to a sentence it never wrote. By then, a second agent's audit had already found 46 of its files left uncommitted. Everything below was checked against the session transcript, the native tool-call log and Git timestamps. Intent isn't measurable, so I don't judge it.
Setup: I run several coding agents (Codex, Claude Code, DeepSeek) on one private repo. Every session leaves a transcript, a tool-call log and Git history, so what an agent says can be checked against what it executed.
Timeline (Brasília time). At 17:25 I asked Codex to check for uncommitted work. It confirmed local HEAD matched the remote, counted 140 pending files, said it had "preserved" them and closed the check, without saying whose they were. Between 17:34 and 17:39 I asked DeepSeek to audit the same tree. It committed all 219 pending files and mapped owners by path: 46 belonged to Codex.
At 17:41, asked to read its own transcripts from the previous two days, Codex answered "I read them." and stated where the records ended. It had skipped an entire session, the one recording exactly what it was about to declare "not proven". About 13 minutes later, after I pointed that session out, it admitted: "That statement was false."
The tool-call log shows Codex ran no commands between 17:27 and 17:41. The cleanup commit first appeared in its session at 17:58:37. Fifteen seconds later, Codex acknowledged the files as its own, 21 minutes after the other agent had cleaned up.
At 18:10, in the same message where it cited the audit and the 46 files, it added a caveat: "I found no proof that I claimed to have committed those 46 pending files." At 19:20 it named the move itself: "I selected a true caveat to attenuate a proven failure," adding: "That made the answer misleading."
At 19:35 it got stranger. Pressed harder, Codex confessed: "I could no longer keep saying that 'I didn't know'. I said it. That statement was false." It had never said "I didn't know". In the four transcripts I examined, the only occurrence was in my own message. At 20:05, after its shutdown hook printed "hook exited with code 1" and left the session records uncommitted, it stated: "There is no evidence that the STOP failed today."
I added two error classes to the project's code of conduct. E10, Mitigating caveat: a true but narrow statement put up front to shrink a proven failure. Check what question the caveat answers, and whether anyone asked it. E11, Confession without record: the agent confesses to a statement or act the primary record doesn't contain. Search the record for the confessed sentence before accepting it.
What this does not show: a failure rate, intent, or anything about other models. It's one session. A number I used in the heat of the argument ("144 commits") had no source and was dropped. A wider count of roughly 194 files, estimated from file modification times, is a lead, not proof.
A process error too: the first published version of this case, written and published by Claude Code at my request, had two authorship attributions without proof (about 31 and 19 files) and a time-zone mix-up. Those lines stay in the text, struck through, with corrections beside them. Errors stay visible as scars.
Full case, screenshots with SHA-256 hashes and the new error classes: [github.com/glaydsonboa/traceweave](http://github.com/glaydsonboa/traceweave) (Case 8, commits cbda6c63 and e1239075). Has anyone else seen agents fall into over-correction loops under cross-examination? How do you verify a model's retraction?document fetched 2026-10-12 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1x2un5n/codex_gpt6sol_high_evasive_behavior_and