Public open-source research preview · released August 23, 2026

CoherenceLattice Operational Cognition Engine

Community Edition v0.1.0-alpha.3 is a local-first, provider-agnostic research and review engine for authenticated source bundles, typed observations, bounded structural mapping, explicit uncertainty, deterministic receipts, and exact replay.

  • MPL-2.0
  • Python 3.11+
  • Standard-library runtime
  • No account or activation
  • No model required
  • No network required
Choose a download Read the Technical Manual

What this release is for

Make a local review process inspectable without turning it into truth authority.

The engine separates source identity, structural interpretation, uncertainty, decision posture, and human authority. It can help researchers reproduce how a bounded result was formed while preserving the difference between evidence and conclusion.

Authenticate local source

Create a deterministic three-file grounding bundle with source and segment hashes, then reject altered, extra, linked, or malformed bundle members.

Validate typed packets

Check canonical task and observation packets, source references, bounded axis values, uncertainty, candidate IDs, adapter declarations, and non-authority flags.

Project the full posterior

Aggregate candidate probability by declared equivalence group. Display-only top_k truncation cannot change the disposition.

Map or refuse

Produce a source-grounded AHA map—analogy, hypothesis, and action—only when authenticated segments provide adequate token coverage.

Preserve limitations

Search for explicit counterevidence and limitation markers, retain residual uncertainty, and refuse to manufacture support where grounding is inadequate.

Seal and replay

Write human-readable and machine-readable receipts, close manifests and checksum ledgers, and reconstruct the exact run from its sealed inputs.

Five-minute path

From one local document to a replayable review packet

  1. Install or extract the release.The pure-Python wheel requires Python 3.11 or newer and no runtime dependency download.
  2. Run the included demonstration.The demonstration intentionally returns HOLD; a bounded hold is a successful result when material gates remain open.
  3. Build a deterministic grounding bundle.The engine copies local source into a closed bundle of manifest.json, source.md, and segments.jsonl.
  4. Run the vertical slice.The engine verifies source binding, maps or refuses, projects the full posterior, scans limitations, emits telemetry, and seals the run.
  5. Read, verify, and replay.Inspect HUMAN_READABLE_RECEIPT.md, verify manifest closure, and reproduce the complete file tree from sealed inputs.

Decision states are process postures, not final answers.

PASS_SCREEN means a bounded local screen passed and human review is still required. HOLD preserves material uncertainty, ambiguity, limitation, or an open gate. REFUSE means source binding, transparency, uncertainty, or grounding is inadequate.

Local AI integration

A model can propose a candidate. It cannot grant itself authority.

This alpha does not bundle or call an AI model. Researchers may place it beside a user-selected local model by writing a separately reviewed adapter that produces a canonical candidate observation packet.

Permitted research pattern

local source → local model candidate → typed adapter packet → CoherenceLattice validation and review → human judgment

The adapter must declare its identity, version, locality, network and provider requirements, memory-write capability, schemas, and claim ceiling.

What the adapter may not do

It may not bypass validation, hide provider or network effects, suppress uncertainty, silently activate a dictionary, classify a person, turn a candidate into a final answer, or create publication, deployment, training, memory, or truth authority.

No model adapter ships in the base alpha. “Provider-agnostic” describes the contract boundary, not a claim that every model is already integrated or tested.

Exact public artifacts

Choose the form that matches your task.

Every artifact is no-charge and publicly accessible. Verify the SHA-256 before installation or execution.

Source ZIP

Complete public source, documentation, fixtures, tests, schemas, rights ledger, SBOM, and deterministic release tooling.

Choose this when: you want to inspect, study, fork, test, or build the exact source tree.

SHA-256
008b7ee0ae77360592c6073844f1b0401ffbbb34d520f02aac3c05a5b4ef2b9a
Download Source ZIP Checksum

Python wheel

Pure-Python package for an existing Python 3.11+ environment. Runtime dependencies: none.

Choose this when: you want the fastest offline install of the commissioned Python package.

SHA-256
eea6dd5603360e0e0daf384dbe0417c5f4ff1777d2b13561684d3cb15a1e3a81
Download Python wheel Checksum

Python sdist

Source distribution for standards-based Python packaging and source-install research.

Choose this when: you need a Python source distribution rather than the full repository-style source ZIP.

SHA-256
d11157eb506be04d9785aa7e2902048595f941906f5a9eb4074c2632de1f1c86
Download Python sdist Checksum

Why the source package may still say “public release not authorized”

The source ZIP was sealed before the later human exact-hash authorization. Its candidate-era language was preserved so the independently reviewed bytes would not change. The public-release wrapper dated August 23, 2026 authorizes only the exact hashes shown on this page.

Claim ceiling

What this alpha does not establish

Not a truth or science certificate

It does not prove GUFT, Pattern Donation semantic non-vacuity, external cross-domain utility, improved deployed AI cognition, AGI, consciousness, or a universal ontology.

Not a person-classification system

It must not diagnose, score, rank, predict, or assign identity, destiny, morality, empathy, spiritual standing, credibility, or human worth.

Not high-stakes autonomous decision software

It is not suitable as an autonomous clinical, legal, emergency, employment, credit, insurance, surveillance, punitive, weapon, or compliance-certification system.

Not a production security certification

A hash records byte identity, and a receipt records a tested process posture. Neither proves truth, authorship, consent, legality, safety, quality, or authority.

Support, correction, and security

One monitored inbox, routed by subject line

Do not email passwords, credentials, private keys, sensitive source material, protected health information, or customer records. Use email first to request an approved transfer route.

General inquiryQuestions about the project, license, or release library. Product supportInstallation, commands, receipts, or reproducible defects. Security reportPrivate report with version, hash, environment, and minimal synthetic reproduction. Correction requestChallenge an inaccurate public statement or request clarification.

Voluntary nonprofit support

Help keep public-interest research open and usable.

Voluntary gifts help fund accessibility, independent review, documentation, maintenance, educational resources, and grant-sponsored access to UVLM tools and research.

Giving is always optional. Donations never affect access to public materials, product behavior, verification results, ordinary support, or governance decisions.

Support UVLM