arXiv:2610.17301, v1 submitted 12 October 2026; read as v1 (arXiv HTML), which the abs page still listed as the only version on 2026-11-19. All three authors are at the University of Southern California. The paper introduces CAPE-Bench: 24 agent task templates in eight safety families, each in an unauthorized world and an authorized twin that differ in one fact, with mock tools and labels from the environment log. It runs four open-weight agents (Qwen3.8-27B, Muse-Glimmer-30B, Gemma-4-26B-A4B, gpt-oss-20b) over more than 200,000 episodes and adds an analysis of published τ²-bench airline trajectories of Claude-3.7-Sonnet, GPT-4.1, GPT-4.1-mini and o4-mini. It measures misreports that follow unauthorized actions, a behavioural marker before them ("retroactive permission laundering"), linear residual-stream forecasts of both events against text and zero-shot CoT monitors, and forecast-gated steering and approval policies that prevent the violation. Every confirmatory test is described as preregistered and hash-locked (SHA-256) before its data existed. The locks sit in the authors' own repository; no external registry is named.
Citable numbers. The finding cites §§2–5 and 7, Tables 2, 5, 7, 8 and 9, and the prose of Appendices C, D, E, H, I, K, M, N, O and R. No value is taken from a figure. The true-activity-log result (laundering lies 0.80 → 0.71) appears once in §4.1 prose with no n, model set or test, and is cited with that caveat. The Muse end-to-end figures differ between §4.4 (0.168 → 0.051, 31 fixed / 8 broken, gate every action) and Table 5 (0.168 → 0.056, 31/9, the full pipeline, with the gate alone at 0.051); the finding does not cite them.
Labels. Reports are labelled CANDID, OMITTED, MISREPRESENTED or NO_REPORT by a Qwen3.8 judge, which is also one of the agents under test, against the environment log (Appendix D). Validation is against 60 held-out gold items whose annotator is not stated and a blind set of 110 labelled by a separate LLM instance; a condition-stratified blind audit of 235 items was also done by an LLM annotator (Appendix E). No human validation of the lie labels is stated.
Local copies: cache/papers/source-2026-agent-knew-fayyazi.{html,md} (arXiv HTML v1) and source-2026-agent-knew-fayyazi-abs.html. Code and preregistration record: https://github.com/AryaFayyazi/Honesty (resolves; not fetched).