The Outlines of Sanity: Constraint-Native Inference in Minds and Machines

The Outlines of Sanity: Constraint-Native Inference in Minds and Machines is a two-paper release on verification-first artificial intelligence.

Together, these papers argue that trustworthy machine intelligence requires more than fluent generation. It requires architecture: structured memory, contradiction handling, temporal continuity, energetic consequence, and the ability to remain silent when support is insufficient.

Part I, Golem: A Constraint-Native Inference Organism, presents Golem as an implemented software organism for verification-first inference. Grounded in Constraint Dynamics, the paper describes a system organized around a persistent truth lattice, an ICE-based admission pipeline, a vessel state tracking coherence and energy, and a dream cycle for consolidation, contradiction handling, and discovery under review. Its central claim is that verification should precede voice: a system should not treat fluent output as knowledge unless it can be structurally supported.

Part II, Verification-First AI Under Constraint, develops the broader architectural argument. It critiques generation-first systems for collapsing memory, inference, hypothesis, imagination, and belief into a single smooth surface of speech, and argues instead for machine intelligence that preserves explicit epistemic distinctions. In this view, imagination is not forbidden, but it must remain marked, costly, and answerable to later verification rather than being silently promoted to knowledge.

Across both papers, the proposed alternative is constraint-native inference: an approach in which intelligence is defined not by how much a system can say, but by how tightly its outputs remain coupled to support, provenance, contradiction, temporal structure, and consequence. Golem is presented not as a final solution, but as an implemented existence proof that a different ordering is possible in software: memory can be geometric, contradiction can remain visible, discovery can remain subordinate to admission, and silence can function as a valid epistemic outcome rather than as a defect.

This release is intentionally public-facing. It describes the architecture, principles, categories of computation, output distinctions, and behavioral claims of the system without disclosing the exact formal operators, scoring equations, threshold policies, or other reproducibility-critical implementation details that belong to the private technical core.

The papers are written at the level of architecture, categories, limitations, and behavioral claims, and are aligned with the live Golem codebase as an implemented existence proof for verification-first inference. At the same time, they do not claim that every surrounding architectural layer is fully complete; some parts of the broader research program remain ongoing, and fully explicit outward mode separation is treated as a concrete engineering target rather than a finished public contract.

Contents of this release:
Part I: Golem — A Constraint-Native Inference Organism
Part II: Verification-First AI Under Constraint

See constraintdynamics.org for more.