LLM-CA / PCR-Bench: Recorded LLM Studies and Independent Human Annotation

This version provides the recorded LLM studies and independent human annotation for LLM-CA / PCR-Bench, with an author-confirmed annotation-origin correction. The submitting author explicitly confirms that two HCI researchers independently produced all 40 final judgments in their respective sets from scratch, without using or viewing AI-generated suggestions. The original forms' AI-origin declarations and the preceding version's AI-assisted description were erroneous according to that confirmation. The public provenance correction records this discrepancy and its author-confirmed basis; original forms and previously published versions are preserved unchanged. This is a provenance correction, not a new annotation round, and it changes no judgment or numerical result. The anonymized evidence includes 80 structured human judgments on 40 validation cases, their synthetic inputs, agreement statistics and offline reproduction code. Joint adjudication remains pending. Original Word files, private approval metadata, personal annotator identifiers and free-text comments are excluded. The original API pilot contains 480 test predictions and 16 validation checks. The paired public-policy/no-policy control contains 480 test predictions and 32 validation checks. Exact payloads, responses, ledgers, full analyses and reproduction source are included. The unchanged baseline archive contains the original synthetic corpus and recorded baseline results. Scores measure agreement with provisional generator references, not independently validated test accuracy, safety, service utility or UX benefit. No personal sensor data, clinical judgment or live service outcome is included. License: MIT. Source: https://github.com/kmichalakis/LLM-CA .