No source, no claim
How catalog rows become reviewed findings: the evidence tiers we accept, the identity rules we enforce, the statuses we use, what AI is allowed to do, and where the method stops.
The operating principle
A model can propose where to look, classify a document, or extract a candidate value. Only verifiable evidence can support a customer-facing field. We never convert a guess into an unsupported fact, and we never replace product-specific evidence with a generic fallback merely because a listing field is empty.
Source tiers
| Tier | Sources | Permitted use |
|---|---|---|
| Tier 1: Authoritative | Official manufacturer documentation, official registries, primary regulatory data such as the EU Safety Gate portal | Can support a proposed value, subject to human review |
| Tier 2: Corroborating | Marketplace product pages, secondary catalogs and databases | Can corroborate a Tier 1 finding; cannot establish a value alone |
| Tier 3: Candidate only | Model-suggested leads, fuzzy matches, unverified web content | Generates research candidates only; never presented as verified |
Every accepted source is recorded with its locator, capture date, and a content hash, so any proposed value can be traced back to exactly what was seen and when.
Product identity rules
Two products can share a title, a photo, or a near-identical model name and still differ by variant, batch, voltage, material, or manufacturer. Applying one product’s evidence to another can create more risk than leaving a field unresolved. So:
- Fuzzy matching (title, image, partial model) generates candidates only.
- Evidence is reused across records only on exact identifiers and compatible variant attributes.
- Ambiguous products remain unresolved rather than being force-matched.
Status vocabulary
We deliberately avoid a binary “compliant / non-compliant” verdict. Reviewed products carry one of these statuses:
| Status | Meaning |
|---|---|
| Ready for submission | A reviewer approved the proposed values and their evidence; awaiting your confirmation to export |
| Review | Evidence exists but conflicts or is incomplete; a specific confirmation is needed before use |
| Specialist | The category or claim requires a qualified specialist; we prepare the data, not the judgment |
| Hold | A risk signal (for example a recall candidate) requires mandatory human confirmation first |
| Unresolved | Available evidence does not support a value; the report says so instead of guessing |
The role of AI
AI may identify candidate information; unsupported candidate values are never presented as verified facts.
Models are used to normalize files, propose where evidence might be found, classify documents, and extract candidate values. Every customer-facing proposed value passes deterministic validation (identifier checksums, entity and variant consistency) and human review before it appears in a deliverable.
Safety Gate matching
The EU Safety Gate portal is used as a recall-signal input. Matching is conservative by design: a fuzzy match never automatically blocks or clears a product. Candidate matches are held for exact model and batch confirmation, and the outcome is documented either way. A finding is a candidate match, never a statement that a product is or is not subject to a recall.
Human review
- Human-reviewed findings are capped and prioritized by commercial importance, marketplace risk signals, evidence-gap severity, and identity conflicts. They are not reviewed indiscriminately.
- Each review decision records the reviewer, the evidence considered, and the outcome.
- Recall candidates and high-risk categories are always human-reviewed.
Category boundaries
Some categories require specialist analysis that this service does not perform: toys and children’s products, electrical and electronic equipment, batteries, PPE, cosmetics, medical devices, food-contact products, chemicals, automotive safety components, and machinery, among others. These can appear in the catalog risk map, but their status is Specialist. They are never automatically marked ready for submission.
Limitations
- The output is catalog readiness and evidence-quality analysis. It is not legal certification, legal advice, or product testing.
- Some evidence can only be supplied by the manufacturer, importer, authorized representative, or Responsible Person. The report classifies these cases rather than inventing values.
- Marketplace acceptance depends on platform processing, catalog ownership, and category rules outside our control. We measure outcomes; we do not guarantee them.
- Evidence reflects what a source showed on its capture date; sources can change afterwards.
- Nothing is submitted to any marketplace without your review and approval.
See the method applied to your catalog
Start with the ungated sample report, or apply for a qualified snapshot.
Get my free 250-row snapshot