OPEN MODEL WEIGHTS · EVIDENCE METHODOLOGY

Unknown is a valid value.

The registry is designed to maximize useful coverage without converting assumptions into facts. Every important field should resolve to source evidence, an explicit classification rule, a labeled derivation — or an honest unknown.

Source-firstField-level verificationDaily revision checksObserved history
Evidence statusSource-first verified
Latest registry check2026-11-16
Latest record full verification2026-11-16

CORE PRINCIPLES

The rules behind every record.

The goal is not to make every field look complete. The goal is to make every published claim inspectable.

01

Source first

A field starts with observable evidence from the listed source repository or linked publisher documentation — not from naming conventions.

02

Field by field

Verification applies to individual claims. A record can contain verified fields alongside explicit unknown or not-disclosed values.

03

Unknown stays unknown

Missing evidence is preserved as missing. Open Model Weights does not fill gaps just to make records look complete.

04

History compounds

Repository revisions and verified-field changes become observed evidence over time instead of being overwritten by the newest state.

REGISTRY LIFECYCLE

From repository to evidence record.

Discovery, verification, revision checks and history are separate stages.

01
Discover candidate

Identify a source repository that appears to publish open-weight model artifacts.

02
Pass publication gate

Require recognized weight artifacts and enough source evidence to create a real model record.

03
Verify fields

Check repository/API/config/model-card/license evidence field by field.

04
Classify explicitly

Store verified, declared, derived, not-disclosed or unknown states instead of silently inferring.

05
Check revisions daily

Read the current repository revision and fresh API metadata on the daily registry run.

06
Retain changes

When source evidence changes, create new observed history and field-level diffs.

FIELD → EVIDENCE → RULE

How individual claims are established.

Verification is specific to the field. One source does not automatically validate the whole record.

Field groupPrimary evidenceMethod / boundary
Identity & weightsSource repository API + exact file listingRepository exists, recognized weight artifacts are present, exact filenames are retained.
LicenseModel-card metadata + repository license files / linked termsDeclared terms are recorded with evidence. Commercial-use classification is a comparison aid, not legal advice.
Context & parametersStructured config/API first; explicit source text secondStructured values are preferred. Naming conventions do not become facts.
Formats & precisionObserved repository artifacts, filenames and dtype/config signalsPositive signals are recorded. “Not observed” does not mean no third-party conversion exists.
LineageDeclared base-model metadataNo parent model is invented from similarity, architecture family or naming.
Training assetsRepository files + obvious model-card disclosureDisclosure signals are recorded; Open Model Weights does not reconstruct undisclosed training.
Runtime supportSource tags, model card and repository artifactsCompatibility is source-derived unless a runtime is explicitly labeled as independently tested.
Hardware memoryDerived from verified parameter countWeight-only estimate: excludes KV cache, activations, runtime overhead and sharding.
Popularity & freshnessSource repository API metadataDownloads/likes aid discovery, not quality ranking. Revision checks are distinct from full field verification.

EVIDENCE STATES

“Verified” is not the only honest state.

Open Model Weights preserves the difference between something we directly checked, something the source merely declares, something we calculate, and something the available evidence does not establish.

Verified

Directly checked against the named evidence source.

Declared

Present in source metadata or publisher text, but semantically a publisher/source declaration.

Derived

Calculated from verified inputs and labeled as a derivation.

Not disclosed

The checked standard evidence did not expose the value.

Unknown

Available evidence is insufficient for a defensible value.

DERIVED VALUES

Hardware estimates are deliberately narrow.

They estimate storage for model weights only — not end-to-end deployment memory.

BF16 / FP16parameters × 2 bytes

Approximate weight-only memory for 16-bit weights.

FP8 / INT8parameters × 1 byte

Approximate weight-only memory for 8-bit weights.

INT4parameters × 0.5 byte

Approximate weight-only memory for 4-bit weights.

Excluded by design

KV cache, activations, optimizer state, runtime overhead, quantization metadata, device placement and sharding are not included. A displayed memory estimate is therefore not a deployment guarantee.

FRESHNESS & DATES

Repository activity and verification are not the same thing.

The daily pipeline checks current repository revision and fresh API metadata. If a source revision changes, relevant evidence is fetched again for field verification. The last revision check and the last full field verification are retained as distinct signals.

Repository revisionFreshness signal

Used to detect source changes without re-fetching every unchanged artifact.

Full field verificationEvidence check

The most recent run that re-evaluated the relevant record fields.

Repository createdRelease-date proxy

Used only when no separate structured release date is available, and labeled as a proxy.

Publisher updatedActivity metadata

A changed timestamp does not automatically become a semantic model release.

BOUNDARIES

What the registry does not claim.

These limits are part of the methodology, not footnotes. They prevent useful discovery signals from being overstated as stronger evidence.

No guessed completeness

A blank field is preferable to a plausible but unsupported value.

No publisher ownership assumption

The indexed Hugging Face URL is called the source repository unless publisher ownership is independently established.

No deployment guarantee

Weight-memory estimates are comparison aids, not claims that a model will run within that amount of RAM or VRAM.

No untested runtime guarantee

Runtime entries remain source-derived unless an execution test is explicitly documented.

No legal determination

Commercial-use labels summarize checked terms for comparison and are not legal advice.

No quality score from popularity

Downloads and likes remain discovery signals and never become a model-quality ranking.

TEMPORAL EVIDENCE GRAPH

How observed history is created.

The searchable registry shows the current verified state. The temporal layer records when that evidence state was observed and whether it later changed.

01
Observe a source revision

The daily process checks the current source-repository revision and relevant fresh metadata.

02
Normalize field evidence

Verified fields are projected into a stable evidence shape covering identity, weights, license, context, formats, lineage, training disclosure and runtime signals.

03
Hash the evidence state

An evidence hash identifies the normalized projection. The repository SHA separately identifies the checked source revision when available.

04
Compare with the prior observation

If the evidence hash changes, Open Model Weights computes field-level differences rather than silently overwriting the prior state.

05
Retain material history

A new snapshot is kept when the evidence state changes. Unchanged daily checks remain verification events without duplicating an identical snapshot.

06
Expose history to humans and agents

Model history pages, diffs, the Temporal Evidence Graph JSON and MCP make the observed timeline queryable.

Observation boundary

Open Model Weights does not claim retroactive knowledge. observed_at means when this project captured the evidence state, not necessarily the model publisher’s release date. Unknown or undisclosed fields remain explicit.

Read the Temporal Evidence Graph → What changed today? →

REPRODUCIBILITY

Inspect the evidence layer yourself.

The methodology is backed by public data surfaces rather than a closed scoring system.

FOUND A QUESTIONABLE FIELD?

Corrections should leave an evidence trail too.

Report the model, the field in question and the strongest source you have. Public version control keeps changes attributable and inspectable.

Report a correction ↗