PACTA-Bench: crisis-bargaining episode data for language-model agents (WP3 confirmatory and moving-b* studies)

Episode-level data for PACTA (Protocols for Auditable Commitments and Trustworthy Agreements), a game-theoretic benchmark of language-model agents in crisis bargaining. The archive contains per-configuration aggregated cell metrics (war rate, pledge-breaking rate, accept-then-attack rate, effective and verbal deception, welfare, equilibrium deviation) and per-episode transcripts (offers, responses, commitments, period-2 actions, ground-truth types, closed-form theory values, outcomes and payoffs) for eight agents across four model families, six narrative covers and a four-rung commitment-protocol ladder, plus the threshold-moving (moving-b*) follow-up on two frontier agents. All actors are synthetic and no credentials or personal data are included. The frozen pre-registration is at OSF (doi:10.17605/OSF.IO/DN2BR); analysis code is in the accompanying code deposit. See DATA_README.md for file naming, the field dictionary, and the mapping from files to the paper's tables and figures.