-
1st-order accuracy — beliefs vs ground truth
Crowd mafia recall & deception success (dashed)
2nd-order consistency — social_map vs actual beliefs (Mafia–Mafia excl., conf≥50)

Calibration — stated confidence vs actual accuracy

Fresh role_assessment probes only, self and "Unknown" guesses excluded. Diagonal = perfect calibration.
Hatched bars = overconfident (accuracy below bin midpoint). Whiskers = 95% CI (Wilson). Bins with n<10 are dimmed — too few observations.

⑂ Branch from here

Fork point: — replays the recorded game from this exact step. Outcomes form a distribution (LLM sampling is stochastic by design).
Replays:
Backend override:
Probes stay on in branches; cost ≈ parent game × replays.