Can an AI-Assisted Research Repository Preserve a No?

This descriptive case report asks whether an evolving, human-ratified AI-assisted research repository can preserve inspectable negative decisions rather than allowing internally coherent claims to inflate unchecked. The primary audit covers 32.3 days and a frozen denominator of 238 numbered run records, 11 top-level claims, zero promoted top-level claims, zero killed top-level claims, all three registered path kills, and seven proposed self-evaluation metrics. None of the seven metrics passed the program's combined execution and adversarial-validation bar: three failed independence requirements, two failed for construct or algorithmic reasons, and two were not computable from retained data. A separate post hoc, single-coder linkage analysis classified one downstream case as a same-family search constraint, one as human-input confounded, and one as preservation-only; deterministic checks verified source integrity and literal references, not semantic judgments, and no comparative method effect was identified.The contribution is a denominator-bounded archival record, an explicit map of machine-mediated checks and terminal human authority, and an operationally checkable disclosure checklist for scientific-agent reports. The paper does not claim method novelty, efficacy, autonomous model refusal, independent replication, or correctness of every recorded negative decision. The deposit includes the publication PDF, editable source, and a paper-minimal evidence archive containing the complete run denominator, exact cited source subsets, human-readable tables, deterministic verifiers, provenance, licensing, and checksums.