Engine, Rules, Canon: An Architecture Statement for Continuously Iterating, User-Owned AI Agents
Year: 2026
Type: article
Abstract: SUPERSEDED BY THE FULL EMPIRICAL PAPER. This v1 architecture statement is being superseded by a full empirical paper on a separate fresh concept DOI chain, which adds a theoretical foundation (Bateson-style type-token separation, three-primitive minimality, Structural Role Separation), a formalized Reviewer-Correlation Ceiling Hypothesis, and a preregistered empirical evaluation on PolybrainBench. The successor paper's concept DOI will be listed here when deposited. The author affiliation has been updated in this metadata to reflect the academic identity used on the successor papers. This note is an architecture statement (a position paper, not an empirical paper) placing the load-bearing vocabulary and eleven numbered contributions of the Polybrain architecture into the public record for priority claiming on 2026-04-14. We introduce 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 drives the architecture, tuned by two orthogonal autonomous dials (per-item effort and background iteration rate) to maximize the time derivative of a scalar canon-trustworthiness measure we call NetTrust. We argue that fused-weight large language models are a special case of this invariant in which engine and rules have been welded together into a single opaque weight matrix owned by the model vendor, and that separating them yields a strictly more general and more trustworthy class of AI system — one that can be continuously self-improving, structurally auditable, and owned by the user on whose machine it runs. The eleven numbered contributions claimed for priority: (1) engine(rules) → results as a strict three-way type separation re-applied to LLM-driven agents, under which fused-weight LLMs are a special boundary case; (2) the canon as a first-class, append-only, user-owned substrate distinct from vector stores, memory buffers, and knowledge graphs; (3) NetTrust scalar and its time derivative as the kernel's primary objective function; (4) the continuously iterating kernel with a gas pedal and per-item effort dial as two orthogonal autonomous self-tuning axes; (5) the four-predicate publisher test (TOTAL, BOUNDED, NON-RANKING, WRITE-CAPTURING); (6) the four-primitive witness stack composed into a five-label verdict; (7) the coherence web with negation-polarity contradiction detection; (8) self-modification gated by the system's own verification stack; (9) the Reviewer-Correlation Ceiling Hypothesis; (10) the ownership axis as a structural architectural commitment orthogonal to the feature axis; and (11) Polybrain as a reference implementation of the above. This is not an empirical paper. Its only empirical commitment is the existence of the reference implementation in the private polylogicai/polybrain repository. The full empirical paper — Off-Model Verification: The Reviewer-Correlation Ceiling and a Deterministic Canon Substrate for LLM Agents — is in preparation and will be released separately with a reproducibility package. Companion repository: github.com/polylogicai/polybrain-architecture-statement, release tag v1-architecture-statement. License: CC BY 4.0.
Author: Salvo, Andy
Institution: Pennsylvania State University
Language: English
Subfield: Safety Research
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7154221959
FWCI:
Cites:
Cited by:
Related to: