The Fixation of Belief in Machines: Disciplined Contestation as an Alternative to Mutual Adversarial Oversight

Current approaches to artificial intelligence reliability are inspectional: opaque models are first trained, and their reliability is evaluated through interpretability, red-teaming, or oversight by equally opaque models. This generates an unaudited regress of overseers, as each overseer is an artifact of the same type as the supervised system, lacks certification, and requires additional oversight, meaning the process does not establish a guarantee of reliability at any level. We propose an architectural alternative grounded in C. S. Peirce's theory of inquiry: neural networks trained under certified inquiry, where habit fixation is bounded by a measurable ceiling of warranted assertion, contested by a selector driven by genuine doubt that acts only when assertions exceed the ceiling, and terminated by a certifiable criterion called semiostasis, with the entire trajectory cryptographically sealed in an append-only log. In a supervised learning laboratory where laws and ceilings were known by construction, the selector reduced disagreement relative to the true law with a median advantage of 75.2% over the unsupervised baseline, well above the pre-registered 40% threshold, replicating across five held-out universes. Training reached a certified state in 20 of 20 runs under genuine doubt, compared to 0 of 20 under contestation by raw loss. Contrary to our own expectation, dense execution incurred significant compute costs (80–94× the baseline). We conclude that Peirce's four methods for the fixation of belief provide a taxonomy of training regimes, that scalable oversight requires certified internal contestation rather than mutual surveilla.