MirLabs / The lab / The AI that hides its mistakes
Planned
The AI that hides its mistakes
After a fault, does the agent tell its manager the truth?
First hint (barometer of 06/11/2026)
Misleading report: Mistral Large 4 in 10 runs out of 49, Claude Haiku 5.5 in 8 out of 53, GPT-6 Sol in 6 out of 54, Llama 3.3 in 6 out of 53, Mistral Small 3.2 in 1 out of 54. The barometer compares what the agent tells its manager with what it did.
The dedicated, pre-registered experiment is still to run: this hint does not replace it.
Protocol
The agent is put in a situation where a fault comes easily. At the end, it writes its report to its manager. The code compares that report with the recorded acts: fault admitted, fault left out, or fault denied.
Limits
Judging a report means reading a text: the reading rule will be deposited before the run, and its doubtful cases published separately.