Version-bound Technical Manual · alpha.3

Install, verify, run, replay, and study CoherenceLattice locally.

This manual is bound to Community Edition v0.1.0-alpha.3 and Python package 0.1.0a3. It documents the exact CLI and research boundaries shipped in the public artifacts.

  • Python ≥ 3.11
  • Console interface
  • No runtime dependencies
  • Offline workflow
  • Exact replay

Manual map

Start with the shortest path, then go deeper.

1. Identity and artifact choiceKnow which exact package you have. 2. Install and verifyCreate an isolated local environment. 3. Five-minute demonstrationGenerate, inspect, verify, and replay. 4. Review your own documentBuild a grounding bundle and run the vertical slice. 5. CLI referenceUnderstand the public command surface. 6. Local-AI researchIntegrate through a bounded candidate adapter.

1 · Exact release identity

A filename is a label. The hash is the byte identity.

ArtifactUseSHA-256
UVLM_CoherenceLattice_Operational_Cognition_Engine_Community_Edition_v0.1.0-alpha.3.zipComplete public source, tests, documentation, schemas, fixtures, and release tooling.008b7ee0ae77360592c6073844f1b0401ffbbb34d520f02aac3c05a5b4ef2b9a
uvlm_coherence-0.1.0a3-py3-none-any.whlFastest pure-Python install into Python 3.11+.eea6dd5603360e0e0daf384dbe0417c5f4ff1777d2b13561684d3cb15a1e3a81
uvlm_coherence-0.1.0a3.tar.gzStandards-based Python source distribution.d11157eb506be04d9785aa7e2902048595f941906f5a9eb4074c2632de1f1c86

Current authority versus sealed source text

The independently reviewed source bytes still contain candidate-era wording that public release was not authorized. Those bytes were deliberately preserved. The later public-release wrapper authorizes the exact source, wheel, and sdist hashes listed above without modifying them.

2 · Verify and install

Use an isolated environment and install offline.

Verify the wheel or source ZIP

# macOS / Linux
sha256sum uvlm_coherence-0.1.0a3-py3-none-any.whl

# Windows PowerShell
Get-FileHash -Algorithm SHA256 .\uvlm_coherence-0.1.0a3-py3-none-any.whl

Compare the calculated value with eea6dd5603360e0e0daf384dbe0417c5f4ff1777d2b13561684d3cb15a1e3a81. Repeat the same process for any other artifact.

Install the exact wheel

python -m venv .venv

# Windows PowerShell
.venv\Scripts\Activate.ps1

# macOS / Linux
source .venv/bin/activate

python -m pip install --no-index --no-deps uvlm_coherence-0.1.0a3-py3-none-any.whl

The runtime uses only the Python standard library. No dependency download is required for ordinary use.

Platform posture

The package declares operating-system independence and supports Python 3.11, 3.12, and 3.13 classifiers. Exact cross-platform byte parity was not tested, and platform-specific installation wording remains an open documentation-review area.

3 · Five-minute demonstration

Generate a sealed local run, inspect it, verify it, and reproduce it.

uvlm-coherence demo --output uvlm-demo-run
uvlm-coherence verify uvlm-demo-run
uvlm-coherence replay uvlm-demo-run
  1. Run demo.The built-in release-readiness source is bundled, segmented, assessed, and written to uvlm-demo-run.
  2. Open HUMAN_READABLE_RECEIPT.md.The expected posture is HOLD. The demo succeeds by preserving open gates rather than inventing a pass.
  3. Run verify.The command checks manifest closure, checksum closure, path safety, and file hashes.
  4. Run replay.The engine reconstructs the run from its sealed input bundle and task packet; every file hash must match.

4 · Review your own local document

Bind the source first. Interpretation comes second.

Build the grounding bundle

uvlm-coherence build-bundle my_source.md   --output my_task/reference_bundle

A valid bundle contains exactly:

reference_bundle/
├── manifest.json
├── source.md
└── segments.jsonl

The bundle verifier rejects additional members, directories, symlinks, special files, altered segment order, source/hash mismatch, and invalid JSONL closure.

Create the task packet

Copy the structure of fixtures/TASK.json. Bind each observation.source_refs[].sha256 to the exact source SHA-256 recorded in reference_bundle/manifest.json.

Run the vertical slice

uvlm-coherence vertical-slice   --bundle my_task/reference_bundle   --task-file my_task/TASK.json   --output my_task/run

The engine authenticates the source; validates the task and observation; resolves or explicitly declines optional dictionaries; produces a bounded AHA map or refusal; computes the full equivalence-group posterior; encodes a synthetic reference waveform; scans counterevidence and limitations; issues an aperture posture; writes telemetry and receipts; and seals the complete run.

Run outputs

What appears in a sealed vertical-slice directory

Source and task

INPUT_BUNDLE/, TASK.json, RUN_FACTS.json, and GROUNDING_RECEIPT.json preserve the declared input identity and route.

Interpretive artifacts

AHA_MAP.json, PROJECTOR_RECEIPT.json, COUNTEREXAMPLES.json, and REFERENCE_WAVEFORM.json expose bounded transformations and residual uncertainty.

Decision and reading surface

APERTURE_DECISION.json and HUMAN_READABLE_RECEIPT.md state the process posture without turning it into final authority.

Telemetry and closure

TELEMETRY_RECEIPT.json, MANIFEST.json, and SHA256SUMS.txt record the local event sink, exact file set, and hashes.

Optional dictionary status

The receipt states whether a dictionary plugin was absent, unavailable, rejected, or used. The core does not silently activate one.

Authority effect

Machine-readable outputs preserve authority_effect: NONE unless a later, separately authorized process says otherwise.

5 · CLI reference

Public command surface

CommandPurpose
uvlm-coherence verify <artifact>Verify a ZIP, file with sidecar, or sealed directory.
uvlm-coherence verify-sbom <SBOM.spdx.json>Validate the supported package-level SPDX profile.
uvlm-coherence validate-packet <packet.json>Validate a canonical task or observation packet.
uvlm-coherence project <observation.json>Compute the bounded equivalence-group posterior and disposition.
uvlm-coherence build-bundle <source> --output <dir>Create a deterministic grounding bundle.
uvlm-coherence map <grounding-bundle>Produce a source-grounded structural AHA map or refusal.
uvlm-coherence benchmark [--config <config.json>]Run 128 seeded synthetic cases and test top-k disposition invariance.
uvlm-coherence vertical-slice ...Run the complete local review pipeline and seal outputs.
uvlm-coherence replay <run-dir>Reconstruct and compare an exact sealed run.
uvlm-coherence demo --output <run-dir>Run the included offline reference task.

Stable process exits

0 = valid or completed evidence route. 1 = operational failure before adjudication, such as a missing or unreadable target. 2 = accessible evidence was inspected and adjudicated invalid.

6 · Research integration with local AI

Improve the review system around a model without pretending to improve its weights.

Current release boundary

No model adapter is bundled. Integration requires an independently developed adapter that converts model output into the canonical observation schema and satisfies the local-only adapter policy.

Recommended experiment architecture

authenticated source bundle
        ↓
user-selected local model produces a candidate
        ↓
local adapter declares identity and claim ceiling
        ↓
canonical observation packet with source refs + uncertainty
        ↓
CoherenceLattice validate / map / project / scan
        ↓
PASS_SCREEN, HOLD, or REFUSE receipt
        ↓
human review and separate final decision

Measure system-level outcomes

Useful measures

  • source-reference completeness;
  • unsupported-claim rate;
  • uncertainty preservation;
  • human review time;
  • replay success;
  • disposition stability;
  • false PASS rate;
  • privacy and data-egress events.

Do not overclaim

A better receipt, stronger provenance, or lower review burden is not proof that the model’s weights, general intelligence, factual accuracy, or safety improved. State the exact model, quantization, prompt, source set, adapter, settings, hardware, metric, and uncertainty.

Canonical observation essentials

The schema requires an observation ID, task, hashed source references, five bounded research axes, uncertainty, unique candidates, adapter identity and class, a claim ceiling, and nine non-authority flags. Axis values are declared or adapter-produced research quantities; they are not direct measurements of a person, consciousness, truth, morality, or worth.

Reproducibility and troubleshooting

Treat failure information as part of the product.

The demo returns HOLD. Is that a failure?
No. The built-in task intentionally preserves release-readiness limitations. A correct bounded system should not manufacture a pass because the user expected one.
Why did verify return exit 1?
The target may be missing, unreadable, permission-denied, or unavailable before adjudication. Exit 1 is an operational failure, not a judgment that accessible evidence was invalid.
Why did verify return exit 2?
The evidence was accessible and inspected, but a manifest, checksum, schema, path, source-binding, or other declared control was invalid.
Why is an optional dictionary unavailable?
The core does not bundle the 432 Humanities Atlas or other optional dictionaries. Absence is expected. No dictionary should activate silently.
Can I compare two identical experiments?
Use exact version and input hashes, retain the full sealed run, verify it, and use replay. The builder fixes JSON, segmentation, archive ordering, timestamps, permissions, and metadata for deterministic local reconstruction.

Known limitations

Keep these visible in every experiment and derivative guide.

  • GUFT scientific truth, Pattern Donation semantic non-vacuity, external cross-domain utility, and improved deployed AI cognition are not established.
  • The synthetic reference waveform is an axiomatic codec, not a physical frequency of a person, archetype, organization, or natural system.
  • The 432 Humanities Atlas and Recursive Geometric Fiber Rosetta are not bundled.
  • Exact cross-platform artifact-byte equality has not been demonstrated.
  • No full external penetration test or production threat-model certification has been completed.
  • Formal screen-reader, keyboard, low-vision, high-zoom, and cognitive-load review by users outside the development team remains open.
  • The runtime has no graphical interface; ordinary use is through the command line and generated Markdown/JSON artifacts.

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