PAI-Bench Measures Persistent Identity in Deployed AI Agents

Summary

The paper introduces PAI-Bench, a provider-neutral benchmark for evaluating whether deployed AI agents faithfully maintain a versioned identity contract. Rather than treating identity as simple fact recall, it separately measures recall, composition, behavioral enactment, resistance, persistence, lineage, and role-conditioned updates, with scoring oracles kept outside the target process. Two frozen evaluation campaigns used 16 synthetic profiles, 32 probes, and three independently initialized target configurations, producing 1,536 retained responses. In a literal audit, direct-parent identifiers appeared in all 48 atomic responses, but implicit self-portraits appeared in only one. On eight profiles, adding explicit field cues raised the joint presence of three identity identifiers from 0/8 to 7/8 under the same four-sentence instruction. A startup body-label substitution similarly increased full-designation presence from 1/8 to 7/8, while parent identifiers remained absent. These results indicate prompt-dependent component selection and sensitivity to startup cues in the tested deployments. The study also found evaluator sensitivity: replaying identical factorial responses produced a Claude headline mean 12.5 percentage points below Astra’s. The authors note that each condition used a single target sample and that audits and follow-ups were post hoc. PAI-Bench is presented as a reproducible protocol for distinguishing factual availability, identity expression, and behavioral enactment.