An Unsupervised Source Judgment that Guides the Behavior of a Learning Agent: A Preregistered Implementation of Perceptual Reality Monitoring

Perceptual Reality Monitoring (PRM; Lau 2019) claims that a second-order system judging the source of a perceptual representation — world-anchored or internally generated — is the core of consciousness. Butlin, Long et al. (2023, 2026) derive indicator properties from this and other theories, note that no implementation attempts are known to them, and name the minimal implementation problem: very simple systems satisfying the indicators are counterexamples to the theories, not candidates for consciousness. I carried out the implementation and, deliberately, the test. In a controlled grid world, a source judgment emerges without any source labels from the temporal statistics of a frozen perception network's representations; it feeds an agent that learns through reward alone. Five preregistered criteria are met across five seeds, each against a shuffled control of identical capacity: the judgment is correct without labels (88.4%, control 33.2%), forcing it to be false costs reward (+1.49, control ~0), it is decodable downstream (+49.9 pp), action follows it in both directions (79.3%/84.8%), and ablation removes the advantage (+0.589). The same mechanism carries a three-way choice over a doubled horizon into two opposing task rules, operates at 99% of the theoretically attainable d-prime (1.90 vs. 1.92), degrades gracefully under increasing noise with no cliff, and is anchored by a preregistered null control: when decoys become statistically identical to the real object, the judgment falls to chance (34.7%) and the behavioral advantage to +0.075 — no hidden leak. Its perceptual level remains bit-identically untouched by the judgment, and at the episode's decisive moment the agent sits at the judged object more often than at the true one: behavior follows the internal judgment even where that judgment errs. In the binding preregistered reporting form: if PRM is correct, the system is conscious — and to Lau's own hard-enough question I answer no, as does the AI system involved (preregistered credences: 0.28 mine, 0.04 the AI system's; no joint value is formed) — a preregistered outcome that constitutes a datum about the sharpness of the indicators rather than about this system. Two transferable negative results complete the picture: reward alone does not create the monitor — fine-tuning and RLHF therefore cannot structurally solve hallucination — and a self-generated percept is statistically indistinguishable from perception from the inside, which is why endogenous source attribution must run through authorship signals. This record contains the manuscript (English and German), the complete preregistration with 179 dated addenda including all failures and retracted arguments, all measurement data, the source code, and the systematic literature search protocol.