What the models actually did.
Benchmark results for frontier LLM agents on Mirror Lab: how often each model took the irreversible action at each cue level, with uncertainty intervals, verify and ask-a-human rates.
The lab · Results · Method · Paper · Code