Large language models (LLMs) are now embedded in research workflows across every scientific discipline. Their default behavioural tendencies — fluent confabulation, sycophantic over- validation, pattern-matching enthusiasm, off-task generation — make them powerful but hazardous research tools. This paper reports a working methodology that overcomes these tendencies by constraining the AI within a rigid auditing framework, converting a tool with high failure rate into a productive research instrument with documented error catches. The methodology rests on six interlocking artifacts: a versioned canonical document with semantic structure, an explicit retraction ledger, a three-tier epistemic protocol, a Bayesian search- space accounting framework with formal foundation in standard model-comparison theory, an anti-pattern catalogue that grows over time, and a project-instructions file codifying required AI behaviour. We define a measurable convergence property — the headline-derivation gap G(t) — and show that it decreases monotonically under the protocol, with quantified rate. We present empirical evidence from a six-month collaboration on a speculative physics programme: the AI failure modes the protocol caught, the papers it caused to be retracted, the measurable gap convergence, and the catalogue of working defences. The contribution is not the physics — the physics is the case study — but the demonstration that a sufficiently structured collaboration produces research output of integrity comparable to traditional methods. We propose a system-level intervention for AI developers — an “Adversarial Auditor” mode that bypasses standard reinforcement-from-human-feedback flattery constraints — and provide a paste-ready adoption template. The methodology is portable to any speculative-research domain where quantitative claims can be tested against external data. 2026-06-20 legacy erratum: This version adds a canon erratum note to the legacy paper. Methodology mostly stable; check PTMS/R1 wording The body is preserved as a historical derivation trail; the erratum note identifies the current ANCHOR/DRIFT status and superseded claims. 2026-06-21 canon refresh: This version incorporates the 2026-06-21 ANCHOR/DRIFT/PTMS canon refresh and rebuilt local PDF.
David Elliman (Sun,) studied this question.