Engineering a local profile-aware validator for structured AI-agent operation records: a 19-fixture boundary study

Author preprint, version 1. This paper reports a local engineering boundary study of agent-evidence v0.6.0 over 19 synthetic, author-labelled JSON fixtures. It compares JSON Schema-only validation, the EEOAP profile-aware reference validator, and a same-author, same-repository boundary checker. The three paths accept 5/5 labelled-valid fixtures and reject respectively 2/14, 10/14, and 14/14 labelled-invalid fixtures. These counts describe declared coverage differences only; they are not estimates of accuracy, independent validation, real-world robustness, compliance, certification, or external event truth. The deposit includes the author preprint and a reproducibility supplement with fixture hashes, result files, scripts, known-gap probes, an artifact matrix, and a threat-model boundary. POP integration remains unimplemented. This preprint has not been peer reviewed.