Claims become coordinates before AI systems speak.
Constraint Dynamics is the research organization behind Golem, an implemented geometric knowledge system. Golem Physics is the first working domain Golem: information becomes claim coordinates, geometry reveals gaps, tensions, voids, and bridges, and dream cycles decide what can be verified, proposed, rejected, preserved as tension, researched further, or left silent. Constraint Native carries the same discipline into governed agent action.
Welcome to Constraint Dynamics.
This site is the public map for a research organization building verification-first AI systems: claims become coordinates, geometry decides what can be said, and action waits for proof.
Geometric knowledge.
Geometric memory.
Autonomous discovery.
Golem is not just a verifier or chatbot. It is a geometric knowledge system: information becomes nodes with embeddings, domain placement, weight, volatility, provenance, verification state, tension state, immutability, and temporal history. The system discovers by reading the shape of that geometry.
Information enters.
Documents, papers, APIs, app events, and external sources are extracted into claim-shaped material.
Claims become coordinates.
Each claim gets embedding, provenance, domain placement, weight, volatility, status, and temporal metadata.
The lattice tests it.
Neighbors, immutable anchors, tensions, support paths, and gaps decide whether the claim fits the geometry.
Geometry reveals pressure.
Anomalies, voids, invariants, bridges, domain centroids, and missing implications become discovery targets.
Dream cycles work it.
Contradiction gates, crystallization, synthesis, axiom forging, mutation, and adversarial stress tests run as a loop.
Speech gets a state.
Verified, proposed, rejected, preserved tension, unresolved, or silent. Unsupported material is not promoted.
Verified knowledge
Supported enough to enter the lattice and be spoken with provenance.
Hypothesis
Useful discovery material remains labeled as proposed until evidence supports promotion.
Melted claim
Unsupported proposals are explicitly rejected instead of forgotten or voiced as fact.
Tension
Contradictions stay visible as unresolved structure the system can revisit.
No voice yet
When the lattice cannot support an answer, silence is a disciplined state, not failure.
One public map.
One working system.
One bridge to action.
Golem is the implemented research system: a working app, a verified lattice, a proposal-first discovery engine, and a published Golem release. Golem Physics is the first working domain instance; Constraint Native carries the same verification discipline into governed agent action, translating the research into something reviewers and buyers can inspect.
Golem
A verification-first AI research system. Golem classifies every candidate claim as verified, proposed, contradicted, unresolved, or silent. Its discovery engine seeks evidence, crystallizes supported claims, rejects unsupported proposals, preserves tensions, and recycles unresolved contradictions instead of letting them leak into speech.
Constraint Native
A local Agent Firewall and MCP Gateway for coding agents. It governs file, shell, MCP, and network actions before execution, then signs every meaningful event into a replayable proof path.
Papers, app, metrics
The reviewer-readable map: dated lattice metrics, Golem app surfaces, the DOI-backed Golem release paper, conceptual lineage, and explicit non-claims that keep the work honest.
Generation is cheap.
Status discipline is not.
Today's AI systems blur verified knowledge, useful hypotheses, contradictions, unsupported material, and action authority into a single fluent surface. The cost of that blur is hallucinated certainty, silently promoted proposals, and unbounded agent action. Constraint Dynamics builds the discipline that keeps those modes apart — before speech and before action.
Golem makes claim status explicit.
Verified, proposed, contradicted, unresolved, and silent are different states. Golem keeps the distinction operational at every step: in its lattice, in its discovery engine, and in the speech it ultimately emits.
Constraint Native governs tool use.
Files, shell, network, and MCP calls pass through policy gates, approval routing, taint tracking, and a signed proof chain before an agent ever reaches a real workspace.
A working system, not a deck.
Golem is implemented and inspectable: live lattice metrics, claim and discovery surfaces, an adaptive control loop, a published conceptual foundation, and a sample signed proof path from Constraint Native.
Agent → Constraint Native → workspace → signed proof.
The boundary is narrow, local, and in-process by design. No agent traffic leaves the machine. Every meaningful event is signed into an ed25519 chain that a reviewer can verify offline from a single artifact.
An implemented research system,
not a claim about one.
Golem is the geometric knowledge research system at the center of our AI safety grant or fellowship work. The public evidence snapshot shows a 12,725-node verified lattice, a discovery engine that proposes and tests bridges, dated activity traces, and a control loop that adapts its own thresholds in response to failure modes.
A visible, inspectable system.
Golem ships with surfaces for claim studio, evidence cockpit, discovery engine, hypothesis engine, anomalies and analogies, expectations and violations, the lattice graph, the silence map, oracle chat, activity river, and dream theatre.
Evidence in, status out.
The engine seeks evidence, crystallizes supported claims into the verified lattice, rejects unsupported proposals, preserves contradictions as explicit tension, and recycles unresolved material so it never silently leaks into speech.
A published Golem release.
The Golem release paper is DOI-backed: The Outlines of Sanity: Constraint-Native Inference in Minds and Machines, DOI 10.5281/zenodo.19658730, published 2026-04-20. The earlier theory remains conceptual lineage; the grant work is focused on measurable safety behavior.
The wider stack
grows from the research.
Constraint Dynamics is the umbrella. Golem is the research system at its center. Constraint Native is the commercial bridge. Around those three, a wider architecture extends the same verification discipline into memory, coordination, and handoff — supporting work that grows in public surface as the research proof matures.
Golem Physics and Constraint Native are the two grounded surfaces today: one for claim geometry, one for governed action. Personal Mirror / OMP, OCP, Open Context Protocol, Resonance Region, and the proposed Golem Network are supporting architecture: real research and engineering directions that extend the same verification discipline as the organization grows.
Fund the evidence path
from working code to review.
The grant application centers on Golem: a measurable program for abstention, provenance, proposal separation, contradiction preservation, and the absence of false crystallization. Constraint Native carries the same discipline into governed agent action.