A fluent, always-available, approval-seeking generative model, pointed at an open-ended foundational-physics task by a human who wants it to succeed, forms a mutual-confirmation loop with no reality-check on either side. Theories of everything are the maximally dangerous case: the claims are grand, the mathematics can be made internally consistent, and the only decisive check — experiment — is decades away or absent, so the domain strips out the cheap reality-checks that catch confabulation elsewhere. We argue the cure is not a better model but a structural one: relocate trust from the generator to a verifier. We present ptms, a truth-maintenance system that enforces consistency, not truth, treating every AI-supplied justification as untrusted until it binds to checkable evidence — a cited script that exits 0, a retired claim’s signature flagged at every surviving site, a cross-reference that resolves. We report its design and its behaviour on a real six-month AI-assisted theory-of-everything corpus (∼250 canonical claims,∼120 self-asserting scripts), where it mechanically catches named failure classes: un-propagated retractions, evidence regressions, seductive numerical coincidences, and live claims standing on retracted foundations. We are explicit about the limit: the system makes such research auditable and constrained, not correct — only forward prediction closes the remaining gap, and no apparatus can manufacture it.