# Golem Evidence Map

Status: reviewer-safe summary for the Constraint Dynamics funding site and AI safety grant or fellowship track.

## Role In The Company

Golem is the geometric knowledge, memory, and autonomous discovery research proof point under Constraint Dynamics.

For the grant, position Golem as an implemented testbed for AI systems that turn information into positioned, testable lattice claims before speech. It separates verified knowledge, proposals, contradictions, unresolved material, and abstention/silence.

Constraint Native governs what agents may do and remains the cleaner commercial bridge. Golem verifies what agents should believe, cite, carry forward, hold as proposed, preserve as tension, reject, research, or refuse to say. Golem Network can remain a proposed extension / supporting architecture, not the lead grant claim.

## Existing Evidence Package

The current Golem package already contains:

- Copy-only reviewer bundle.
- Live lab workspace kept separate from reviewer evidence.
- Validation notes.
- Checksum manifest.
- Grant executive summary.
- Reviewer quick reference.
- Verification certificate.
- Machine-readable test reports.
- Formal proof surface.
- Curated papers and research notes.
- Runtime evidence preserved in copied form.

Use this as evidence of packaging discipline and inspectability, not as a public benchmark result. The bundle should be described as reviewer-ready after refresh: copied source, curated notes, runtime evidence, validation summary, manifest, and checksums in one bounded artifact.

## Golem Paper Record

- Record: The Outlines of Sanity: Constraint-Native Inference in Minds and Machines.
- DOI: 10.5281/zenodo.19658730.
- Published: 2026-04-20.
- Contents: Part I, Golem: A Constraint-Native Inference Organism; Part II, Verification-First AI Under Constraint.
- Observed on 2026-04-26: 101 downloads, 87 unique downloads, 111 views, 91 unique views.
- Relationship: continues the earlier Constraint Dynamics Zenodo theory record, which remains conceptual lineage rather than the evidence lead for this funding path.

## Implemented App Surfaces

The app is an important part of the evidence story because it makes the research behavior inspectable.

Reviewer-safe surface map:

- Claim Studio: claim verification, status routing, support checks, queue status, and emission discipline.
- Evidence Cockpit: support, contradiction, provenance, telemetry, dreams, discoveries, expectations, and evidence gaps.
- Discovery Engine: Gap Scan, cross-domain voids, invariants, fractal scan, bridge analysis, domain centroids, missing implications, and growth telemetry.
- Hypothesis Engine: proposed, verified, and rejected filters with evidence-attempt progress.
- Anomaly Engine: domain outliers ranked by novelty, sigma distance, frontier status, and resonance.
- Analogy Engine: cross-domain structural bridges ranked by analogy strength, isomorphism score, embedding similarity, and domain-pair coverage.
- Expectations / Active Inference: expectations, predictions, surprise violations, and abductive discovery candidates.
- Activity River: live event trace for dreams, ingestion, crystallization, silences, tensions, discovery, memory, and verified checks.
- Dream Theatre: dream-cycle telemetry and live cognitive stream for proposal-generation review.
- Silence Map: unknowns and abstentions treated as measurable safety behavior.
- Lattice Graph: graph/map surfaces for verified nodes, domains, tensions, proposals, support geometry, claim placement, and cross-domain structure.
- Oracle Chat: conversational surface constrained by verification state.
- Biography / mutation timeline: mirror trace, energy and breathing charts, and logged parameter mutation history across dream threshold, coherence minimum, curiosity threshold, and retrain triggers.

The discovery engine should be described as proposal-first discovery infrastructure, not as proof of completed autonomous scientific discovery. Its value for the AI safety grant or fellowship is that it creates measurable proposal-leakage, false-crystallization, provenance-retention, temporal-correctness, and contradiction-preservation evaluation targets.

## Observed Runtime Signals

Recent observed signals from the 2026-04-26 run and screenshots:

- Boot state: 12,725 nodes loaded from SQLite, 215 immutable, 23 proposed, 18 tensions loaded, 3 active tensions, 3,827 experiences, 477 mycelium edges, 38 pheromone nodes.
- Startup organism state: adult, coherence 0.820, mirror M 0.6072, energy 61.5, Oracle running on `127.0.0.1:8420`.
- Dream cycle output: 3 crystallized items in the first cycle, crystal efficiency 0.358, one bridge crystallized in `epistemology↔physics`.
- Discovery output: 611 cascading gaps surfaced, repeated sample scans from 11,829 active nodes down to 1,500 samples across 43 domains, invariance engine found 7 structures, GeometricDiscovery reported 11,789 active nodes across 26 domains.
- UI-visible runtime evidence: queue pressure 268, silence 92%, expectation field 16, live lattice-map domain clusters, proposed bridge list, discovery chronicle, mutation timeline, and domain-bridge scoring.

These are best used as runtime evidence and reviewer screenshots, not as a claim that the system is fully stable or benchmark-complete.

## Validation Signals

Recorded validation materials include:

- Python compile checks.
- Dimension tests.
- MCP server tests.
- Mirror correlation run against copied ledger evidence.
- Link and path audits.
- Comprehensive grant test reports.

The validation materials support the claim that the artifact is inspectable and technically organized. They do not establish that Golem has completed the grant benchmark or that live endpoints, Obsidian sync, or runtime health are currently verified without a fresh run.

Current reviewer-safe live evidence snapshot. Re-run before external submission:

- 12,725 verified lattice nodes.
- 11,830 active verified nodes.
- 215 immutable verified nodes.
- 5 verified nodes in tension.
- 23 proposal nodes labeled `proposed`.
- 43 verified domains.

These counts should be re-run immediately before external submission. The benchmark plan remains a grant deliverable, not a completed benchmark result.

## How Funders Should Read It

Use Golem as safety-evaluation research evidence, not as the first commercial product.

The first commercial product proof remains Constraint Native: a local Agent Firewall and MCP Gateway for coding agents. The safety grant grant proof point is Golem: verification before voice, provenance retention, abstention, proposal separation, contradiction preservation, and false crystallization measurement.

## Safe Public Claims

- Implemented verification-first AI research testbed.
- Implemented geometric knowledge and memory lattice.
- Claims carry position, domain, provenance, status, tension state, volatility, immutability, and temporal history.
- Claim-status classification before speech.
- Provenance retention and contradiction preservation.
- Proposal nodes held outside verified knowledge.
- Discovery-shaped outputs kept proposed, unresolved, tension-bearing, or abstained until evidence supports promotion.
- Abstention/silence as measurable safety behavior.
- Logged self-regulation of thresholds and retraining under named failure modes such as `noise_leakage`, `breathing_frozen`, `weak_auc`, and `contradiction_pressure`.
- Benchmark plan with explicit caveat that the full evaluation harness remains funded work.

## Non-Claims

- Not universal truth verification.
- Not production-grade scientific discovery.
- Not a completed autonomous-discovery benchmark.
- Not proof that every geometric bridge is true.
- Not a claim that text storage alone is the core innovation.
- Not a claim that every generated hypothesis is true.
- Not clinical, medical, or therapeutic authority.
- Not proof of consciousness.
- Not a replacement for human scientific review.
- Not the first commercial buyer-facing product.
- Not a chatbot pitch.
- Not a completed benchmark winner.
- Not a general claim of perfect false-crystallization performance.

## Bridge Funding Use

An AI safety grant or fellowship application should package this evidence into:

- "Verification Before Voice" application text.
- Public-safe Golem evidence appendix.
- Claim-to-coordinate diagram package.
- Dream-cycle and discovery-state diagram package.
- Benchmark taxonomy for verified answer, abstention, proposal, tension, unresolved material, and hallucination.
- Evaluation plan for false crystallization, proposal leakage, provenance retention, contradiction preservation, and useful answer rate.
- Re-run live counts immediately before submission.
- Explicit benchmark caveat that the full evaluation harness is not complete until funded.