Model Card Revision Monitor

Compare explicit model-card revisions and return evidence-located semantic changes across identity, license, provenance, evaluation, training, use, limitation, safety, environmental, citation, and file-reference categories.

What this does not do

This Actor does not crawl Hugging Face, discover models, rank model quality, verify a card's claims, infer legal permission, execute model code, download model weights, or monitor continuously after a run. It does not treat wording changes as proof that a model itself changed.

Input

Supply 1-50 documents. Each document contains exactly one of huggingFaceRepo or modelCardUrl. Hugging Face inputs may specify revision and priorRevision; both are resolved to immutable commit SHAs before the README is read. For explicit URLs, priorRevision is an optional second HTTPS URL.

Controlled public example:

{
  "documents": [{
    "huggingFaceRepo": "google-bert/bert-base-uncased",
    "revision": "5546055f03398095e385d7dc625e636cc8910bf2",
    "priorRevision": "0e9f43ffc3afb401b28131d3443a63cb98e5deb8"
  }],
  "includeUnifiedDiff": true,
  "concurrency": 1
}

hfToken is optional secret input for repositories the customer is authorized to access. baselineStoreId may name an existing Apify key-value store for paid, persistent baselines. Without an explicit prior revision or persistent baseline, the first successful observation creates a baseline result.

Output

Each document receives one deterministic result row followed by an uncharged run-summary row. A result includes immutable source revisions where available, raw hashes, parsed YAML metadata, section inventories with line locators, category deltas, severity, before/after excerpts, and a unified diff. Diffs larger than 200 KB are placed in run key-value storage and referenced by hash.

Categories are identity_metadata, license, base_model, datasets, pipeline_task, evaluation, training, intended_use, limitations, safety, environmental, citations, and file_references. Pure whitespace changes are ignored. A section rename with unchanged content is a low-severity move. Removal of license, intended-use, limitation, or safety text is high severity; this is a review priority, not a legal or factual conclusion.

Pricing

The primary pay-per-event event is model-card-check at $0.05 per document. A baseline creation, unchanged comparison, material change, or definitive current-card 404/410 is chargeable. Unsafe URLs, authentication failure, rate-limit exhaustion, transient source failure, missing prior revision, unsupported media, invalid YAML, and internal failures are uncharged.

One document costs up to $0.05 in Actor events, ten cost up to $0.50, and fifty cost up to $2.50, plus platform usage displayed by Apify. The Actor stops admitting work when the run charge limit is reached.

Baselines and billing order

When baselineStoreId is set, the candidate baseline is journaled before output. It becomes current only after Apify accepts the paid event. A charged-but-interrupted commit is recovered at the beginning of the next run. A baseline-store outage is uncharged and stops the run rather than silently switching to stateless behavior.

Source and authorization policy

The Actor reads only the named Hugging Face repository or explicit customer-provided HTTPS URL. Requests are bounded to 2 MB per response, redirects are revalidated against SSRF rules, private/reserved destinations are rejected, and transient responses receive bounded retries. Customers are responsible for access authorization and source terms. See SOURCE_POLICY.md.

Partial failures and negative results

Every admitted document has its own status. Duplicate inputs are rejected without charge. A missing current card is a paid definitive negative; a missing prior card is uncharged because no comparison can be completed. Malformed or unsafe YAML is rejected without charge. Source failures never become successful empty comparisons.

Resource limits and privacy

Launch settings are 256 MB minimum, 512 MB default, 1,024 MB maximum, and a 300-second timeout. Concurrency is bounded from 1-6. The Actor has no external telemetry. Tokens are not logged, emitted, or persisted in baseline records. Dataset and key-value retention follow the customer's Apify settings. See PRIVACY.md and THREAT_MODEL.md.

Validation

The release gate includes a 50-pair hand-labeled synthetic gold corpus. Category precision must be at least 0.95 and recall at least 0.90, every emitted delta must have evidence locators, and whitespace/move cases must produce zero false high-severity findings.

Support

Email support@flintglade.com. See CHANGELOG.md for version history.