Every fact in Central carries a visible trail: which sources contributed, which agreed, which conflicted, which merchant decisions resolved it, and where the value ships. The proof is auditable end-to-end.
The PDP looks like any other product page until you hover a fact. Then the source list, agreement count, confidence score, last verified timestamp, and the merchant decision that resolved the conflict are all one click away. Nothing hidden, nothing invented.
Every fact climbs the same ladder. Imported / merchant-authored data anchors at 1.00. Five or more agreeing sources earns 0.97. Two independent agreements is the floor for display. Below that, the claim is held or routed to review.
Above all: AI never sees the open web. It only sees the verified record. Style can wobble — facts cannot.
How a fact earns the right to publish.
The promise: AI never sees the open web. It only sees facts the record already proved. So copy, bullets, FAQs and schema can be wrong about style — never about fact.
Central never silently picks a value when sources fight. Conflicts route to a queue with the owner, the candidate values, and the source list. The merchant chooses; the decision is logged with timestamp, user, and reasoning. Every claim has a person and a paper trail.
Multiple sources, different values. The engine never resolves on its own — it stops and surfaces the disagreement with all the candidate values intact.
The claim is added to the merchant's decision queue with an owner assigned. The product is not blocked from elsewhere — but the disputed field is held until decided.
Three options, every time: keep the merchant value, accept the verified value, or hold. Each is one click. Each writes an audit entry.
Every decision becomes an audit entry: timestamp, user, field, before / after, and the reasoning. Audit entries are append-only. The trail outlives the merchant.
Every decision, every conflict, every published claim is a record in the audit log. The same trail is reachable through a single REST endpoint, so AI agents, channel feeds, and downstream systems can read provenance without scraping admin pages.
A PDP without provenance looks complete and hides the question. The same PDP with provenance shows every claim's source list and confidence. Unsupported claims like "industry-leading ANC" are held until a source agrees. Verified facts ship.
All of them. Provenance lives on the canonical record and travels with every output. The admin UI shows the full trail. The PDP can expose a click-reveal source list. The API returns it via the /provenance endpoint. AI agents see it through the LLM Data Layer and x-central metadata. One trail, many surfaces.
The disputed field is held — not silently resolved. It's routed to a merchant decision queue with the candidate values, the source list, and an owner. The merchant chooses: keep the merchant value, accept the verified value, or hold for more evidence. The decision is logged. See Enrichment Engine for how source agreement is checked.
Yes. Merchant-authored data anchors at 1.00 — higher than any verified-from-sources value. The override is logged with the user, timestamp, and reasoning, and the source list is preserved alongside the decision. Override doesn't erase provenance; it adds to it.
Yes. Audit entries are append-only. You can read past entries; you cannot edit or delete them. Decisions that change a field write a new entry — they don't rewrite history. Retention is unbounded by default; the trail outlives the merchant.
If you want them to. Each field has a public reveal setting. Some merchants expose the source list inline (great for regulated categories — appliances, electronics, supplements). Others keep the trail internal but expose the confidence number. The choice is per-field, per-channel.
Yes. Provenance flows into x-central metadata on the page and into the JSON-LD Schema.org Product node. AI shopping agents reading the page get the value, the source count, and the confidence — without scraping admin pages. See LLM Data Layer.