LLM-CA / PCR-Bench: Synthetic Context-to-Service Simulation, Recorded LLM Studies and Human-Audited AI-Assisted Annotations

Version 2 adds recorded LLM evidence and anonymized human-audited, AI-assisted annotation judgments to the unchanged PCR-Bench baseline release. The original API pilot contains 480 test predictions from two pinned models on 60 cases from 20 source families, plus 16 validation checks. The additive supplement also records a separate interleaved public-policy versus no-policy follow-up (semantic JSON, two models, two passes), its frozen protocol, validation checks, request records and outcome status. Two HCI researchers independently audited and approved their assigned AI-assisted sets of 40 synthetic validation cases, as coordinator-confirmed. This is not unaided human annotation. The release contains 80 structured approved judgments, corresponding synthetic cases, agreement statistics and offline reproduction code. Original Word files, free-text comments and private approval metadata are excluded. AI draft dependence and human edit histories were not recorded; joint adjudication remains pending. Scores measure agreement with provisional generator references, not independently validated accuracy, safety, service utility or UX benefit. No personal sensor data, clinical judgments or live service outcomes are included. The original baseline archive and prior published version remain unchanged. All evidence is released under MIT. Source: https://github.com/kmichalakis/LLM-CA .