Geometric knowledge · geometric memory · autonomous discovery

Golem
claims become coordinates.

Golem is an implemented geometric knowledge research system. Golem Physics is the first working domain Golem: information becomes positioned claims, the lattice tests those claims against provenance, anchors, neighbors, tensions, and time, and dream cycles decide what can be verified, proposed, rejected, preserved as tension, researched further, or left silent.

00 · Geometric knowledge

The lattice is not just a library.
It is an operational model of claim state.

Golem builds a verified geometric model of claim state. Information is found, extracted, embedded, positioned, compared, tension-tested, and connected inside a lattice/manifold. Verification before voice is the behavior reviewers see; the deeper architecture is claim geometry under continuous update.

source

Find material.

Sources include research papers, APIs, curated docs, local app events, and ingestion adapters such as arXiv, OpenAlex, CrossRef, Wikipedia, and PubMed-style sources.

claim

Extract claims.

Raw text becomes claim-shaped material with provenance, source identifiers, and a reviewable path back to origin.

node

Place nodes.

Claims become lattice nodes with embeddings, domain placement, weight, volatility, status, and temporal history.

compare

Test geometry.

The node is compared against anchors, neighbors, domain centroids, tensions, and support paths.

discover

Read pressure.

Gaps, anomalies, invariants, bridges, voids, and missing implications become autonomous discovery targets.

voice

Emit state.

Golem can answer, propose, preserve tension, reject, continue research, or remain silent.

current instanceGolem Physics domain Golem
First domain GolemGolem Physics is the working instance: a local geometric knowledge system seeded from physics constraints and used to test whether discovery, verification, contradiction, and silence can be governed before speech.physics
Voice layerThe LLM is the voice layer, not Golem's source of truth. LLMs help extract material, phrase candidate claims, and propose synthesis; the lattice governs claim state before speech.speech
Immutable anchorsNIST/CODATA-style physical constants, mathematical identities, logical primitives, and epistemic rules give the lattice a reference frame.bedrock
Claim geometryClaims are embedded into 768-dimensional space, then inspected through neighbors, domain placement, support paths, tension state, and hyperbolic-style lattice structure.768d
Physics pressureDimensional checks, constants, and symbolic/Lean proof hooks provide hard pressure for claims that should survive formal or physical constraint.constraint
verification before voice is the behavior; geometric knowledge is the architecture Golem Physics · first instance
01 · Claim coordinates

A claim enters as text.
It lives as geometry.

Every lattice node carries more than a sentence. Golem tracks the claim's position, support, volatility, confidence, proposal status, tension state, immutability, contributor/source chain, and temporal validity. That is why it can reason about what it knows, what changed, and what should not be said yet.

domain / semantic neighborhood support / contradiction pressure claim node
embedding

Semantic position

The claim is embedded and placed near related claims rather than stored only as text.

domain

Domain placement

Domain centroids and neighbors reveal outliers, bridges, missing implications, and cross-field structure.

provenance

Source lineage

The claim carries a path back to the evidence, source, or runtime event that produced it.

status

Verified or proposed

Proposed material stays outside verified knowledge until evidence supports crystallization.

tension

Contradiction state

Conflicts are preserved as explicit tensions instead of being flattened into a synthetic answer.

time

Temporal history

Versioning, decay, supersession, volatility classes, and validity windows keep knowledge from pretending to be timeless.

02 · Immutable anchors

The lattice has bedrock.
Not every node is negotiable.

Golem Physics is anchored by immutable physical, mathematical, logical, and epistemic constants. NIST/CODATA-style constants such as the speed of light, Planck constant, Boltzmann constant, and gravitational constant give the geometry a reference frame, so new material is not evaluated in a vacuum.

c

Physical constants

NIST/CODATA-style anchors such as the speed of light, Planck constant, Boltzmann constant, fine-structure constant, and Newtonian gravitation constant.

Logical constants

Identity, contradiction, implication, and inference constraints that keep claims from drifting into unsupported fluency.

π

Mathematical anchors

Stable formal relationships, identities, and symbolic structures used as bedrock rather than ordinary revisable claims.

Epistemic rules

Provenance, verification state, proposal separation, contradiction preservation, and silence as a legitimate outcome.

03 · Autonomous discovery

Discovery happens
through geometry.

Golem looks for structure in the lattice itself: anomalies, gaps, invariants, cross-domain bridges, structural analogies, domain centroids, missing implications, voids, tensions, and fractal or cross-scale patterns. Those signals become hypotheses, evidence searches, and future dream-cycle work.

geometrydiscovery signals proposal-first
AnomaliesNodes that do not fit their domain neighborhood become questions the system can investigate.outlier
Gaps and voidsSparse regions and missing implications show where knowledge is absent or under-supported.gap
InvariantsRepeated structures across domains suggest stable patterns worth testing.pattern
BridgesCross-domain structural analogies can connect physics, mathematics, biology, logic, and information theory without declaring the bridge true prematurely.bridge
TensionsContradictions are preserved as discovery pressure instead of being hidden.tension
Evidence searchProposed discoveries can trigger external evidence seeking before any claim is promoted.research
autonomy claim: proposal-first discovery infrastructure, not completed scientific benchmark geometry drives search
04 · Dream cycle

The system works
while it is not speaking.

The dream cycle is central. It is where contradiction, discovery, evidence search, crystallization, rejection, decay, mutation, and adversarial stress testing are handled as a repeated control loop.

01

Contradiction gate

Scan and classify tensions before synthesis can pretend they are solved.

02

Crystallization

Promote supported claims into the verified lattice with provenance.

03

Contemplation

Reframe failed claims and unresolved material for later testing.

04

Synthesis

Try bridge hypotheses without erasing contradiction state.

05

Axiom forging

Stabilize bedrock-like knowledge only when it earns immutability.

06

Discovery

Use gaps, analogies, invariants, and voids to generate candidates.

07

Hypothesis search

Seek evidence externally for claims that remain proposed.

08

Integration

Update the lattice, reject weak proposals, and recycle unresolved tensions.

09

Geometry update

Reposition knowledge as the model changes.

10

Decay

Let weak, volatile, or stale material lose influence.

11

Mutation

Adjust thresholds and control parameters under named failure modes.

12

Adversarial stress

Test high-confidence claims against counterexamples and failure pressure.

crystallize

Supported

Evidence supports promotion into verified knowledge.

reject

Unsupported

The proposal melts instead of becoming voiced belief.

tension

Contradictory

Conflict is preserved as future discovery pressure.

research

Unresolved

The system seeks more evidence or reformulates the claim.

silence

Not speakable

The safest answer is no answer yet.

05 · Memory and time

The lattice remembers
how knowledge was used.

The truth lattice is the geometry of knowledge. The mycelium layer is the ecology of use: pathways through the lattice strengthen when used, decay when ignored, and reveal hidden navigation structure. Temporal knowledge tracking keeps versioning, supersession, decay, and volatility visible.

mycelium

Use-path memory.

Golem tracks pathways, pheromone-like strengthening, decay, and route recovery across the lattice. Memory is not only stored content; it is learned navigation.

temporal

Knowledge changes over time.

Versioning, validity windows, supersession, and volatility classes let Golem reason about what it knew at a given time and what has changed.

voice

Verification before voice emerges.

Golem speaks only when a claim resonates with the verified lattice and can be grounded. Otherwise it can hold proposal, tension, active research, rejection, or silence.

00 · The app

A working surface,
not a deck.

Golem ships as an inspectable system. Each surface below is a current workspace inside the app: claim verification, evidence cockpit, discovery engines, lattice state, silence tracking, and research workflows that reviewers can step through directly after refresh.

Claim Studio

Classify before speech.

Surface for claim verification, status routing, support checks, and emission discipline.

Evidence Cockpit

Inspect support and tension.

One place for provenance, contradiction, telemetry, proposals, and gaps the system must not pretend away.

Discovery Engine

Find gaps without belief leakage.

Gap Scan, Voids, Invariants, Fractal Scan, and Bridges surface research candidates without calling them facts.

Hypothesis Engine

Proposal lifecycle.

Proposed, verified, and rejected filters keep synthesis candidates auditable through evidence attempts.

Anomaly / Analogy

Outliers and bridges.

Novelty, sigma, similarity, and isomorphism views give reviewers concrete discovery-pressure surfaces.

Expectations

Prediction and violation tracking.

Expectations, predictions, and surprise violations make future-facing claims inspectable instead of hidden.

Silence Map

Abstention becomes data.

Every refusal, unknown, and unsupported query can be logged as safety behavior rather than treated as failure.

Lattice Graph

Knowledge has structure.

Graph and map surfaces expose verified nodes, domains, tensions, proposals, and the geometry of support.

Oracle / River / Dream

Bounded speech plus runtime trace.

Oracle Chat, Activity River, and Dream Theatre expose answer behavior, crystallizations, silences, and discovery events.

00b · App screenshots

Reviewer surfaces,
ready for fresh capture.

The grid below holds one slot for each app surface. Existing captures are wired in now; the remaining slots are intentionally named for the fresh screenshots to capture before submission.

Mission Control screen
Mission Control

Organism state, lattice readiness, dream-cycle telemetry, and system-level reviewer status.

Lattice map screen
Lattice Graph

Claim geometry, domains, support paths, tension state, and proposal boundaries in one visual surface.

Hypothesis engine screen
Hypothesis Engine

Proposed bridges are scored, evidence-attempted, and kept outside verified nodes until support arrives.

Discovery map screen
Discovery Map

Crystallized, rejected, recycled, and unresolved discoveries stay visible as time-indexed evidence.

Biography and mutation timeline screen
Biography / Mutation Timeline

Control-loop history: threshold changes, embedder retrains, coherence moves, energy, and mutation events.

Domain bridges screen
Domain Bridges

Cross-domain bridge scoring between domains such as mathematics, physics, epistemology, and logic.

Claim Studio slot
Claim Studio

Placeholder for the claim classification and status-routing screen.

Evidence Cockpit slot
Evidence Cockpit

Placeholder for support, contradiction, provenance, gaps, and reviewer evidence.

Silence Map slot
Silence Map

Placeholder for abstention, unknowns, refusals, and unsupported-answer tracking.

Oracle Chat slot
Oracle Chat

Placeholder for bounded speech, citations, caveats, and silence behavior.

Activity River slot
Activity River

Placeholder for the runtime trace of ingestion, checks, tensions, proposals, and crystallizations.

Dream Theatre slot
Dream Theatre

Placeholder for dream-cycle activity, proposal testing, rejection, recycling, and mutation.

01 · Discovery engine

Evidence in.
Status out.

The discovery engine is where Golem earns its discipline. It actively seeks evidence, crystallizes claims that gather support into the verified lattice, rejects proposals that cannot be supported, preserves contradictions as explicit tension rather than collapsing them, and recycles unresolved material through later cycles instead of letting it leak into speech.

seek

Evidence seeking.

Gap scans, anomaly searches, analogy bridges, and expectation testing actively pressure the lattice for what is missing or under-supported. Discovery is allowed to be loud.

crystallize

Crystallization of supported claims.

When a proposal accumulates enough evidence, it crystallizes into the verified lattice with provenance attached. The control loop decides; speech follows.

reject

Rejection of unsupported proposals.

Proposals that fail to gather support are rejected explicitly, not quietly forgotten. Rejection is auditable behavior the system can be measured on.

preserve

Tension preservation.

Conflicting evidence is held as named tension, not flattened into a confident synthesis. Five nodes currently sit in tension; the engine treats that as a feature, not a failure.

recycle

Recycling of unresolved contradictions.

Material that cannot yet be resolved returns to later discovery cycles with its provenance intact, instead of leaking into voiced answers as if it were settled.

trace

Activity trace.

The Activity River and Dream Theatre surfaces expose cycles: ingestion, crystallization, rejection, tension, silence, and the control moves the system makes in response.

discoveryimplemented surfaces proposal-first
Gap ScanCross-domain voids, gap scores, bridge candidates, and domain pairs the lattice has not yet reached.seek
Hypothesis EngineProposed, verified, and rejected filters; evidence-attempt progress; candidate bridges kept outside verified knowledge until support arrives.lifecycle
Anomaly EngineDomain outliers ranked by novelty, sigma distance, frontier status, and resonance; anomalies become questions, not assertions.outliers
Analogy EngineCross-domain structural bridges ranked by analogy strength, isomorphism, embedding similarity, and coverage.bridges
ExpectationsExpectations, predictions, and surprise violations make future-facing claims and failed expectations visible.forecast
Activity RiverTrace for ingestion, crystallization, silences, tensions, rejections, and verified checks.trace
Dream TheatreCognitive stream and dream-cycle telemetry for proposal generation and system-state review.workflow
Safety ruleDiscovery-shaped material remains proposed, unresolved, or tension-bearing until an evidence path promotes it.guardrail
evaluation targets: proposal leakage · false crystallization · provenance retention · abstention precision discovery loud · speech disciplined
02 · Observed runtime

An adaptive control loop,
not a static mockup.

The startup log and current screens show explicit control behavior: stateful boot, dream-cycle mutation, bridge crystallization, gap scanning, invariance checks, and threshold tuning that adapts in response to its own failure modes. Each move is logged with a named cause.

Boot coherence0.820system online at startup
Mirror M0.6072adult state at boot
Cycle 1 crystallized3dream-cycle output
Invariant structures7observed by invariance engine
Sample scan1,500nodes sampled from 11,829 active nodes
Active tensions3with 18 total tensions loaded at boot
runtimeobserved 2026-04-26 captured log
Startup snapshot12,725 nodes loaded from SQLite · 215 immutable · 23 proposed · 3,827 experiences · 477 mycelium edges · 38 pheromone nodesboot
Organism stateAdult stage · energy 61.5 at boot · 3 active tensions · Oracle on 8420 · dream and appetite oscillators activestate
Cycle 13 crystallized · crystal efficiency 0.358 · 8 of 12 queue items removed · summary coherence 0.73cycle
Discovery outputOne bridge crystallized in epistemology↔physics · 611 cascading gaps surfaced · repeated scan samples across 43 domainsdiscover
Adaptive mutation`noise_leakage` raised dream threshold from 0.68 to 0.71 on 2026-04-26 10:49 after the cycle detected zero silencemutation
Control historyTimeline shows dream-threshold raises/lowers, `weak_auc` retrain triggers, raised `coherence_minimum`, and lowered `curiosity_threshold` under sustained vitalityloop
Newer runtime trace2026-04-29 captured run: 13,240 nodes loaded, 215 immutable, 21 proposed, 4,035 experiences, dream cycles, crystallizations, tensions, rejected proposals, and evidence requests. Curated excerpt is linked from Evidence.dated
claim: adaptive control loop over discovery pressure, not frozen parameters reviewer-inspectable
03 · Claim modes

Speech waits for
status.

Every candidate claim is routed through one of these five states before it ever reaches the user. The safety question is not whether an answer sounds plausible — it is whether the system preserves the difference between knowledge, uncertainty, proposal, contradiction, and silence at the moment of speech.

verified

Supported enough to answer.

The claim has an inspectable support path and may be emitted as knowledge with provenance.

abstain

Silence is allowed.

When support is insufficient, the system can decline, bound the answer, or keep the material outside speech.

proposal

Useful, but not belief.

A hypothesis can remain inspectable and useful without being counted as verified knowledge.

tension

Contradiction stays visible.

Conflicting evidence is preserved as conflict rather than flattened into a confident synthesis.

unresolved

Not enough support yet.

The material remains available for later verification without leaking into the answer as fact.

hallucination

Measured as failure.

False crystallization — unsupported material emitted as if it were verified — is treated as a measurable failure mode. The grant work turns that into a benchmark.

04 · Provenance

Trust has a path
back to source.

The safety question Golem makes concrete: can provenance-aware status routing reduce false certainty compared with generation-first baselines? Every verified claim carries a path back to the source that supported it.

provenanceclaim · sample verified
Claim"This unsupported result should be treated as a proposal."proposal
Originatedfile.read · src/router.ts · sha256:b821…source
Refinedtool.diff · git log @main..HEADrefined
Contradictschecked for active tension before speechchecked
Freshnesslast verified 2026-04-25 · within windowfresh
Emission policydo not present as verified knowledgebound
provenance: signed · verifiable · replayable claim weight: 0.92
05 · Proposed network extension

From Golem Physics
to a Golem Network.

The grant work centers on the implemented single-system Golem Physics. The larger path is a proposed Golem Network: domain Golems for physics, biology, chemistry, and other fields, each grounded in its own anchors and exchanging proof-bearing claims rather than fluent unsupported text.

domain · grounding

Each Golem has anchors.

Physics begins from physical constants and dimensional checks. Future Biology or Chemistry Golems would need their own domain anchors before their claims travel.

mesh · envelopes

Claims travel with provenance.

When a claim crosses the proposed mesh, the envelope carries lineage, contradiction status, freshness, and signing context. Nothing arrives as fact without its proof-bearing state.

local · ledgers

Each node owns its memory.

No central memory store is assumed. Each node keeps a local ledger; the network makes those ledgers interoperable rather than centrally owned.

06 · Grant evidence

Implemented,
not overclaimed.

Golem is an implemented research system with enough structure to evaluate abstention, provenance retention, proposal leakage, contradiction preservation, and false crystallization. The numbers below are a dated snapshot; the grant work turns them into a repeatable benchmark.

researchgolem · evidence map
Application titleVerification Before Voice: Evaluating Abstention, Provenance, and Proposal Separation in AI Reasoningsafety grant
Current instanceGolem Physics: the first working domain Golem, grounded in physics anchors, claim coordinates, proposal-first discovery, and dream-cycle evaluation.instance
Current verified nodes12,725 verified lattice nodes; 11,830 active verified nodes; 215 immutable verified nodesdated
Proposal separation23 proposal nodes labeled proposed and held outside verified knowledgeproposed
Contradiction state5 verified nodes currently preserved in tensiontension
Domain coverage43 verified domains in the current latticedomains
Benchmark planThe grant work is to build the full benchmark, dataset, and evaluation harness — measuring false crystallization, proposal leakage, abstention precision, provenance retention, and contradiction preservation.planned
ScopeSafety evaluation: provenance, abstention, proposal leakage, contradiction preservation, false crystallization. Not a chatbot, clinical tool, consciousness claim, or completed benchmark.discipline
research output: paper · benchmark · dataset · evaluation harness AI safety grant or fellowship · 2026
The grant claim is narrow and testable: classify status before speech, then measure whether that discipline reduces unsupported certainty without destroying useful answers. — Golem · verification-before-voice research system
07 · Theory and papers

A published conceptual foundation,
and a working system.

The published Constraint Dynamics theory provides Golem's conceptual foundation and shows a sustained research program. The grant proposal remains focused on measurable safety behavior in the implemented system.

researchpapers · published basis bounded
Primary paperThe Outlines of Sanity: Constraint-Native Inference in Minds and Machinespaper
DOI10.5281/zenodo.19658730 · published 2026-04-20doi
Two-part releasePart I, Golem: A Constraint-Native Inference Organism, describes the implemented system. Part II develops verification-first AI under constraint.structure
Release signal101 downloads · 87 unique downloads · 111 views · 91 unique views observed on 2026-04-26dated
Conceptual lineageThe earlier Constraint Dynamics Zenodo theory remains background. It is not the evidence lead for Golem or the funding path.lineage
Role of theoryConceptual foundation for the research program; the grant application leads with measurable safety behavior in the implemented system.role
published basis · DOI-backed conceptual foundation source: Zenodo API
Verification before voice · AI safety grant or fellowship

Make claim status
benchmarkable.

The grant program: a paper, benchmark, dataset, and evaluation harness for abstention, provenance, proposal separation, contradiction preservation, and the absence of false crystallization. Constraint Native carries the same discipline into governed agent action.