JSON-LD Structured Evidence Auditor

Audit JSON-LD on explicit web pages and compare selected structured claims with deterministic, source-located HTML evidence.

What this Actor does

For each supplied HTTPS URL, AP-028 performs one bounded HTML GET, finds application/ld+json scripts, parses them with duplicate-key and non-finite-number rejection, inventories top-level and @graph entities, and compares supported claims with visible or machine-linked page evidence. It does not dereference remote JSON-LD contexts.

Supported comparisons include names/headlines, descriptions, URLs, images, SKU/MPN, brand/author names, publication dates, and Offer price, currency, and availability. Each claim is reported as corroborated, not_corroborated, not_observable, or not_compared, with JSON Pointer and HTML locators. This is evidence QA, not a complete JSON-LD processor, search-ranking guarantee, or Schema.org conformance certification.

Input

The controlled default is:

{
  "urls": ["https://flintglade-tools.github.io/apify-fixtures/jsonld/product.html"],
  "compareVisibleContent": true,
  "concurrency": 1
}

Supply 1-1,000 explicit public or user-authorized HTTPS URLs. HTML is capped at 5 MB, combined JSON-LD at 2 MB, script blocks at 100, entities at 500, JSON depth at 40, and supported claims at 5,000. Version 0.1 performs HTTP retrieval only and does not execute JavaScript.

Output

One result row is written per input URL, followed by one uncharged run-summary row. A result preserves:

• source URL with query values redacted, retrieval status, content type, and raw SHA-256;
• per-script byte length, hash, parse status, entity count, and source locator;
• JSON-LD entity type/ID, supported claim value, JSON Pointer, comparison status, method, and visible evidence locators;
• parse errors without copying raw scripts;
• deterministic counts, verdict, logic version, and canonical result hash.

aligned means every comparable supported claim was corroborated and no script was malformed. discrepancies_found means a supported claim was not corroborated or a script was malformed. no_comparable_claims and comparison_disabled are explicit. no_jsonld is a terminal negative, not an error.

Pricing and billing boundary

The pay-per-event event is structured-evidence-audit at $0.05 per completed site audit. Aligned pages, discrepancy reports, malformed JSON-LD reports, pages with no JSON-LD, and definitive 404/410 results are chargeable because the requested audit reached a durable terminal result. Unsafe URLs, duplicate inputs, transient failures, rate limits, unsupported content, parser resource-limit failures, internal failures, and charge-limit remainder rows are uncharged.

One URL costs up to $0.05 in Actor events; ten cost up to $0.50; one thousand cost up to $50.00, plus platform usage displayed by Apify. Chargeable rows use the SDK's combined dataset-push/event path, which persists each admitted dataset item before charging under one charge lock. The Actor never calls a standalone charge before output and stops scheduling after the event charge limit is reached.

Determinism and provenance

The same HTML, base URL, options, and logic version produce the same audit payload and canonical hash. Retrieval timestamps and run IDs live outside the hashed audit data. The fixture manifest records bytes, SHA-256, source URL, license, and derivation; evidence/fixture-audit-v1.json is an exact expected artifact regenerated by tests.

Safety, privacy, and source rights

Only explicit URLs are fetched. HTTPS, DNS, redirects, ports, userinfo, and public-IP resolution are enforced by the shared SSRF boundary. Page bodies and raw JSON-LD are not logged or redistributed. Query values are redacted from emitted URLs. The Actor has no telemetry or external database; Apify retention follows the customer's account settings. The controlled fixture is Flintglade-authored CC0 test infrastructure. Customers remain responsible for access authorization and downstream use.

Limitations

AP-028 does not execute JavaScript, expand contexts, apply JSON-LD framing/compaction, validate every JSON-LD or Schema.org rule, infer hidden visual state, or decide whether a search engine will accept rich results. Exact normalized comparisons are intentionally conservative and may report not_observable when HTML has no field-specific evidence.

Emitted claim values are capped at 500 characters. Longer values retain a deterministic hash and valueTruncated: true; values over 2,000 characters are explicitly not_compared instead of being partially matched.

Support: support@flintglade.com. See SOURCE_POLICY.md, THREAT_MODEL.md, LIVE_PROBES.md, and CHANGELOG.md.