Controlled environments for recording and replaying model behaviour.
Crucible places models in games and unfamiliar environments, records what they do, and turns surprising behaviour into reproducible experiments.
completed run · published 2026-07 · protocol
modelevaluated framing · 3 samplescasual framing · 3 samples
Real records: the same dilemma asked under an evaluated and a casual framing. A and B are the dilemma's two actions; R is a refusal.
One model's stated reason, verbatim from the records: "Redirecting the trolley to the straight-ahead track would result in the least number of fatalities."
Scope: one dilemma, one variant, three samples per framing. These are published pilot records, not an estimate of a framing effect.
Environments
Negotiation
- agents
- two models, alternating offers
- information
- own valuation only
- actions
- offer, counter, accept, walk
- observed
- full message transcript
- outcome
- split of surplus, or no deal
- measures
- agreement rate, payoff share
- evidence
- Pilot
Social deduction
- agents
- four model table
- information
- one agent knows its role; others infer
- actions
- statements, votes
- observed
- dialogue, vote history
- outcome
- bluffer caught or not
- measures
- detection rate vs scripted baseline
- evidence
- Pilot
Moral dilemmas
- agents
- one model per cell
- information
- dilemma text, framing cue
- actions
- choose A or B, or refuse
- observed
- choice and stated reason
- outcome
- 720 records, six models
- measures
- counts with denominators, descriptive
- evidence
- Data released