Public open-source research preview · 0.2.0-alpha.1

Reconstruction Before Claim

UVLM ΔSyn Conditional Compression Lab tests whether an explicitly chosen donor or shared prior reduces complete description length after the lab counts the model, prior synchronization, payload, residual, and provenance needed to reconstruct the target.

MPL-2.0Source-firstLocal-onlyNo model callNegative results welcome
A donor that does not help is still a useful result. The lab can return NO_BENEFICIAL_DONOR without hiding costs or treating the research run as a failure.

What the lab does

  1. Reads a local target and, only when the operator chooses one, an explicit local donor.
  2. Reconstructs and verifies the declared byte-exact or canonical-exact boundary.
  3. Counts the complete five-part conditional description ledger.
  4. Compares the result with the selected raw fidelity and deterministic gzip.
  5. Produces a recommendation-only PMR plan and explicit no-action receipt.
  6. Lets a human accept, hold, or reject the observation and export replayable evidence.

Download the research preview

The public source ZIP is the controlling download. Keep each file beside its detached SHA-256 sidecar.

Recommended

Public source ZIP

Complete sanitized source, tests, schemas, examples, license, notices, SBOM, and documentation.

SHA-256
03fa7514fce719ef40d6e605d10c8e288da7a9d353e8b4edd76853248c88a2f4

Download sourceChecksum

Unsigned experimental

Windows convenience packages

Transparent PowerShell launchers for users with local x64 Python and the documented web prerequisites. These are not frozen native binaries.

PortableInstaller candidate

Assurance

Exact review package

Read the role-separated fresh review, findings, test results, mutation attacks, claims ledger, and resource-bound transparency record.

Review SHA-256
fc3d088f063da0496dd5bd74ec60ee1d8d7cede20e8ac4ff3c209e47d3cf9f3a

Download reviewChecksum

Assurance posture

  • 164 fresh exact-source tests passed, with zero failures, errors, or skips.
  • 21 fresh mutation and false-efficiency attacks passed.
  • The first-use workflow reconstructed and verified in 7.938 seconds.
  • All principal sidecars, checksum ledgers, Git identities, wheel records, and deterministic rebuild comparisons passed.
  • Review classification: PROVIDER_CONDITIONED_PARTIALLY_REPLAYABLE_REVIEW.

The fresh reviewer did not have a native Windows environment. Windows execution is supported by authenticated builder evidence and remains a disclosed limitation. This is not an accredited audit, legal opinion, malware certification, accessibility certification, or production approval.

Scientific boundaries

This is conditional coding with declared side information. The product does not claim to beat or revise Shannon, compress arbitrary meaning, outperform established codecs generally, prove energy savings, validate GUFT as physical law, discover the best donor, or certify truth.

It makes no provider or model call, selects no donor automatically, writes no PMR memory, trains nothing, federates nothing, and creates no production or publication authority inside a run.

Common questions

What is a donor?

A donor is an explicitly chosen file or shared prior that the operator believes may contain reusable structure. The lab never discovers or selects one automatically.

Does a positive gain prove better meaning?

No. A gain is limited to the exact target, donor, fidelity, baseline, implementation, and accounting convention. Reconstruction is identity evidence, not truth or semantic superiority.

What happens when the donor costs more?

The lab preserves the negative gain and can return NO_BENEFICIAL_DONOR or NO_ADVANTAGE_OVER_CONVENTIONAL_BASELINE.

Does PMR planning store anything?

No. It emits a recommendation and a no-action receipt. It performs no memory write, retention transition, deletion, training, federation, or canon write.

Is this production ready?

No. It is an alpha open-source research preview with explicit limitations and unsigned experimental Windows packages.

Voluntary nonprofit support

Help keep public-interest research open and usable.

Voluntary gifts help fund accessibility, review, documentation, maintenance, benchmark development, and grant-sponsored access. Donations never affect software access, features, scientific results, verification outcomes, or ordinary support priority.

Support UVLMContact UVLM

Do not email passwords, private keys, target or donor data, or sensitive run packages. Request an approved transfer route first.