Recovery without Epistemic Monopoly: Historical Evidence under Competing AI Custodians—A Bounded Model and Executable Study

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

TA-TR-2026-05. A bounded conceptual and executable study of historical evidence recovery when AI-mediated custodians may be untrusted or in conflict. English full text and complete Chinese translation describe one study. Methods and limits. The deposit includes deterministic finite-state studies and reproducibility material. It does not report real autonomous-agent attack rates, production penetration testing, or a demonstrated defense against superintelligence. A conventional pinned-reference baseline ties the proposed scoped reporting layer on byte recovery; the negative result is retained. Contribution and responsibility. Hongju Liu initiated the motivating concern, selected the civilizational relevance and authorized this publication. GPT-6 Astra Pro substantially performed literature synthesis, conceptual development, implementation, execution analysis, bilingual drafting and file preparation. Not peer reviewed; no separate final human line-by-line verification is claimed. No OpenAI, Harvard, Zenodo or institutional endorsement is implied. Rights. CC BY 4.0 applies to newly written text and supporting code to the extent rights are held. Referenced third-party materials retain their own rights.