We present Polybrain, an AI agent architecture organized around a single invariant — engine(rules) → results — in which a domain-agnostic engine applies user-owned declarative rules to produce disposable, re-derivable results over an append-only, user-owned substrate we name the canon. A continuously iterating kernel tuned by two orthogonal dials maximizes the time derivative of a scalar canon-trustworthiness measure we call NetTrust. A four-primitive deterministic witness stack composes into a categorical five-label verdict; a four-predicate publisher test governs consumer views; a coherence web forces CHALLENGED verdicts on contested claims; and self-modification is gated by the system's own verification. We ground the architecture in three theoretical commitments: a Bateson-style type-token separation, a three-primitive minimality identifying observe, decide, and remember as the minimum operator set for a canon-keeping learner, and a four-plane Structural Role Separation. We formalize a Reviewer-Correlation Ceiling Hypothesis stating that transformer-based reviewer ensembles asymptote near the Landis-Koch substantial-agreement band because of correlated errors inherited from shared pre-training substrate. We preregister a matched-paper evaluation protocol for the hypothesis, committing publicly to the sampling frame, composite weights, reviewer pools, analysis script, and stopping rule before any data is collected. The preregistration includes the required structurally non-transformer reviewer channel that the hypothesis demands for a falsification test. The executed result will be released as a separate Zenodo deposit on a distinct concept DOI, citing this paper as the preregistration of record. A contradiction of the hypothesis will be reported with equal weight to a confirmation. A reference implementation exists in the private polylogicai/polybrain repository with an honest disclosure of scope in §11. License: CC BY 4.0. Companion repository: github.com/polylogicai/polybrain-paper.
Andrew Salvo (Tue,) studied this question.