◐ mirror/lab

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