Words vs Actions: An LLM Catches Broken Promises, Not Planned Lies
Het Kotadiya
Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
LLM monitors often check whether an agent did what it promised. But a broken promise is not always a lie: an agent can honestly change its mind after hearing others. We tested whether an LLM detective can tell planned lies from honest changes of mind without seeing the agent’s private plan. Using Shi et al.’s Weakest Link game, we ran 50 games with Llama 4 Maverick (250 agents). From these we built a test set of 60: 30 planned liars, checked against their private plans and notes, plus 15 honest agents and 15 mind-changers. Sarvam 105B scored each agent from 0 to 100 using 4 kinds of evidence: its words, its final action, both, or its announced and final numbers. When the detective saw both what an agent said and what it did, it beat guessing (AUROC 0.79 and 0.71). But it could not tell liars from mind-changers (AUROC 0.47–0.63, and every 95% range included chance). It also flagged 11 of 15 honest mind-changers as liars. So say-vs-do monitoring catches broken promises, not planned lies, and it can wrongly accuse agents that honestly change their minds.
Reviews
Clear question with a useful practical message: a say-vs-do monitor flags broken promises, and broken promises include honest changes of mind, so such monitors will accuse agents that adapt. Fixing the labels and test set before running the detective, using a detective from a different vendor, and reporting AUROC with bootstrap ranges are all good choices.
Suggestions:
- Label quality is asymmetric. Liars were filtered by reading their plan and final notes, but mind-changers were taken from the number-based label without a notes check. Apply the same notes check to mind-changers (or run Shi et al.'s LLM judge on all groups) so the two groups are equally clean.
- The detective never sees other agents' announcements, which is exactly the context it would need to tell a reaction from a plan. The negative liars-vs-mind-changers result may therefore be partly built into the design. Adding a view that includes the preceding announcements is the most important next experiment.
- Replacing the 10 truncated answers with 50 affects views B and D most; report results with and without those items.
- With 30 liars and 15 mind-changers the liars-vs-mind-changers intervals are wide, so "cannot separate" is better phrased as "no evidence of separation at this sample size". A larger test set would help.
- Try a second detective model and a thinking-enabled configuration.
The example of a sincere-sounding liar scored 15 and a "performative" mind-changer scored 87 nicely illustrates style-based judgment; it would be worth quantifying. Minor: proofread (e.g., "catched").
Read full reviewShow less
Cite this project
@misc{kotadiya2026words,
title = {{Words vs Actions: An LLM Catches Broken Promises, Not Planned Lies}},
author = {Het Kotadiya},
year = {2026},
month = oct,
note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/words-vs-actions-an-llm-catches-broken-promises-not-planned-lies-s2hc}},
url = {https://apartresearch.com/sprints/projects/words-vs-actions-an-llm-catches-broken-promises-not-planned-lies-s2hc}
}More from AI Collusion Research Sprint
- 1st placeView project: Monoculture Looks Like Collusion: A Replay Test for Hidden Information Flow Between LLM Agents
Monoculture Looks Like Collusion: A Replay Test for Hidden Information Flow Between LLM Agents
Collusion audits ask whether the actions of different agents depend on each other. Agents built on one model act alike without communicating, so these audits accuse honest fleets. Among four honest Claude Haiku agents …
- 2nd placeView project: Looking for a Hidden Agenda: Can a Model’s Activations Reveal Steganography that its Text Hides?
Looking for a Hidden Agenda: Can a Model’s Activations Reveal Steganography that its Text Hides?
Stego-probe
AI agents can collude without saying anything suspicious, just by choosing which true facts to mention. We built a hiring task where Qwen3-8B invents its own covert signals that get past an LLM overseer, then asked …
- 3rd placeView project: Do collusion probes generalise? Prompt confounds in NARCBench
Do collusion probes generalise? Prompt confounds in NARCBench
I investigated HELIOS, an activation-probe for detecting collusion. I found some apparent flaws, and analysed the (acknowledged) unreliable generalisation of the approach.