Work

The Machine Must Not Be Its Own Witness: Claim Standing, Evaluator Custody, and the Limits of Self-Certification in Autonomous Research

Year: 2026
Type: preprint
Abstract: Autonomous research agents can produce convincing code, simulations, tests, signatures and manuscripts while controlling the conditions under which their own claims are judged. This methods perspective examines four source-audited OpenLine incidents: a signing key derivable from distributed source, a benchmark passed by a static seven-field answer, a sensor gate with access to hidden ground truth, and ambiguity between source commits and release archives. We propose an evidence-bound promotion contract separating claim generation from evaluator authority, while retaining classical truth-maintenance and provenance machinery. A reproducible twenty-vector rule-model exercise returns 20/20 correct labels for the proposed composite predicate, but its labels and decisions are constructed together; it is not an independent security comparison or proof of superiority to existing systems. A stronger no-memory counterexample also nullifies an earlier claimed efficiency advantage for evidence-bound agent recall: both approaches solve 240/240 generated tasks in 240 evaluator calls. Our contribution is a narrowly specified protocol, an auditable failure taxonomy, and falsifiers for independent evaluation, not a validated discovery-rate improvement. Publication classification: Research perspective and proposed protocol, with computational case audits and an internally constructed rule-model demonstration. No external blinded trial, hardware-secured verifier, physically validated invention, or demonstrated discovery-productivity advantage is claimed. This is an author manuscript, not journal peer review. Research perspective and proposed promotion protocol. Twenty-vector numerical results are an internally constructed rule demonstration and do not establish superiority to production provenance or security systems. Later 012 independent counteraudit finds fixed no-memory 240/240 parity with EBCR. No external blind team trial, physically validated experiment, or hardware-isolated witness is claimed. The record includes a public computational reproducer and evidence-boundary disclosure. Reproducibility code in the bundle is Apache License 2.0; paper text is CC BY 4.0.
Language: English
Open Access status: green
OpenAlex ID: W7235070689
Cites:
Cited by:
Related to: 10