Public AI incident documentation rarely preserves exposure denominators, follow-up windows, or closure bases: data, coding manual, and analysis scripts (v25)

This archive accompanies a manuscript proposing a candidate minimum longitudinal record-and-follow-up layer for AI incident governance. The public-source documentation audit comprises 100 publicly retrievable records grouped into 91 deduplication clusters, drawn from three source families (60 AI Incident Database snapshot records; 20 OECD AI Incidents Monitor records; 20 direct official, regulatory, or company/public-report records). The archive contains: (1) the case-level coding table with record-level and cluster-level summaries and descriptive Wilson intervals; (2) the source registry with provenance and recheck state; (3) the frozen coding manual; (4) the machine-readable schema template (JSON Schema); (5) analysis scripts reproducing all summaries; and (6) a second-coder reliability worksheet and protocol. The audit describes public-documentation recoverability only; it is not an incidence estimate, a causal analysis, or a protocol validation.