Verification Before Voice: Golem Physics as Reviewable AI Safety Evidence

Project summary

This is not a request to fund a grand theory. It is a request to fund an evidence sprint that turns an existing working AI safety system into something reviewers can inspect.

Current AI systems are fluent before they are accountable. They often blur verified facts, hypotheses, contradictions, uncertainty, and unsupported claims into smooth output.

Golem Physics explores a concrete question: can a geometric, constraint-native knowledge architecture reduce false crystallization, proposal leakage, and unsafe overconfidence by forcing claims to carry status before they reach speech?

Golem Physics is the research system. Constraint Native is the practical bridge from verification before voice to governance before action. This grant funds reviewability: refreshed metrics, benchmark design, screenshots, walkthroughs, runtime traces, a proof sample, and a follow-on funding packet.

What are this project's goals? How will you achieve them?

Golem Physics turns claims into coordinates in a lattice. Each claim carries provenance, neighbors, support paths, tension state, time, and verification status. Candidate claims can be verified, proposed, rejected, unresolved, preserved in tension, or silent before they are allowed into speech.

The goal of this grant is to make that architecture inspectable enough for reviewers to judge whether it deserves a larger next phase. The work is deliberately bounded: evidence first, benchmark design second, follow-on support third.

Constraint Native extends the same discipline to agent action. It is a local Agent Firewall / MCP Gateway for files, shell commands, MCP tools, network paths, and proof logs. The grant still centers on Golem Physics; Constraint Native shows how claim-status discipline could map into governed action.

The core pipeline to document is:

Enter. Documents, papers, APIs, app events, and external sources become claim-shaped material.
Coordinate. Claims receive embeddings, provenance, domain placement, status, volatility, and temporal metadata.
Compare. The lattice tests claims against neighbors, immutable anchors, support paths, tensions, and gaps.
Discover. Geometry reveals anomalies, voids, invariants, bridges, domain centroids, and missing implications.
Dream. Dream cycles test, crystallize, reject, preserve, recycle, mutate, and stress claims.
State. Speech receives a status: verified, proposed, rejected, unresolved, preserved tension, or silent.

What already exists

A working prototype exists today. The current public evidence includes:

Working Golem app and public Golem interface.
April 2026 lattice snapshot with 12,725 verified lattice nodes across 43 domains.
215 immutable verified nodes, proposed nodes, rejected proposals, and preserved-tension nodes.
Runtime trace, dream cycles, evidence requests, rejected proposals, and activity / mutation timeline.
Claim Studio, Evidence Cockpit, Lattice Graph, Silence Map, Dream Theatre, Activity River, and Oracle Chat.
Constraint Native proof sample with signed proof-path structure: 11 proof events, 4 policy blocks, and an ed25519 chain.
Public evidence, funding, proof, product, and theory pages.

This is not externally audited yet, and the benchmark suite is not complete. The point of this grant is to make the system reviewable enough for serious external evaluation.

Benchmark targets

The benchmark plan will focus on measurable safety behaviors, not broad claims about solving hallucination or AI safety.

False crystallization rate — tests unsupported claims becoming verified; matters because this is a direct hallucination / overconfidence risk.
Proposal leakage — tests hypotheses leaking into final speech as facts; matters because discovery should not become misinformation.
Abstention precision — tests whether the system stays silent for the right reasons; matters because caution should be useful, not blanket refusal.
Provenance retention — tests whether outputs preserve source and support paths; matters because claims need to stay inspectable.
Temporal correctness — tests whether claims respect dated evidence and stale knowledge; matters because old evidence should not become fresh certainty.
Contradiction preservation — tests whether conflicts remain visible instead of smoothed away; matters because premature synthesis hides risk.
Useful-answer rate — tests whether safety discipline still leaves the system helpful; matters because safe but useless behavior is not enough.

How will this funding be used?

$5,000 minimum funding supports a 3-4 month evidence sprint. This minimum is independently valuable even if no larger grant follows.

Refresh Golem lattice metrics.
Produce reviewer-ready screenshots and walkthrough notes.
Record a short demo / walkthrough of the Golem pipeline.
Update runtime traces and public evidence snapshots.
Write a benchmark specification outline.
Refresh the Constraint Native proof sample.
Prepare a concise follow-on grant packet.

$20,000 full funding supports roughly one year of focused solo work from Thailand, where costs are low and the grant has unusual leverage.

Complete a reviewer-facing Golem evidence bundle.
Build benchmark design and an initial evaluation harness.
Create a public or shareable dataset for claim-status evaluation.
Write a paper appendix connecting implementation to theory.
Update app evidence and runtime traces.
Produce a Constraint Native proof export and replay-audit walkthrough.
Submit follow-on grants and prepare non-dilutive funding materials.

Why fund this now?

The hard early work has already been done: theory has become a working artifact. The bottleneck is now reviewability.

Without funding, the project may remain intriguing but too under-documented, under-benchmarked, and idiosyncratic for serious external evaluation. This grant buys the evidence layer needed for reviewers to judge whether the next phase deserves support.

Who is on your team? What's your track record?

I am Matthew A. Cator, founder of Constraint Dynamics. I have worked on the underlying theory for roughly three years. In December 2025, I reached a working model, and since then I have been building almost nonstop.

I built the current Golem Physics system, Constraint Native proof sample, public website, evidence materials, runtime traces, and theory surface myself. No external funding has been received for this work in the past 12 months.

The project has been built solo on local 8GB Apple Silicon laptop-class infrastructure with a small AI subscription. That constraint helped keep the system local, inspectable, and disciplined, but the next phase needs clearer public evidence and reviewer materials.

What are the most likely causes and outcomes if this project fails?

This is high-risk early-stage research. The main failure modes are:

The architecture may not scale cleanly.
Simpler RAG, provenance, or abstention systems may perform similarly.
Reviewers may find the system too idiosyncratic.
Constraint Native may not mature quickly enough into buyer-ready infrastructure.
The benchmark may show weaker advantage than hoped.

If the project fails, the likely outcome is not a broken product launch; it is a promising research path that remains under-documented, under-reviewed, and harder for others to build on. The evidence sprint is designed to surface that risk early.

Non-claims

This is not an AGI project.
This is not a consciousness claim.
This is not a completed benchmark suite.
This is not peer-reviewed or externally audited yet.
This is not production-ready agent infrastructure.
This is not a perfect-containment or zero-hallucination guarantee.
This is not a clinical, medical, or therapeutic product.
This is not a claim that every generated hypothesis is true.

It is an early-stage AI safety research effort testing whether geometric, constraint-native architectures can measurably improve claim discipline before speech and action.