Needs review flags unsourced hard facts, not embellishments

The AI judge now reads Stuck to known facts more narrowly. It fails when the assistant states a hard fact that no source backs: a price, a stock level, a link, a delivery, return, payment or warranty term, a delay or an amount promised in your name, a figure about the product, or an ingredient or allergen said to be present or absent.

A product description, a usage tip, a benefit or a comparison between your products no longer fails it, even when your catalog doesn't say it.

• Needs review. A conversation now gets the High priority (a trust reason) for a hard fact with no source, not for an embellishment.
• Facts not sourced. The tile on the Quality report's Conversations tab counts hard facts only, and its help text says so. Its rate isn't comparable with what it showed before this change.
• The trust score, the Trust failure tile and the "A trust criterion failed" filter move the same way.

The label keeps its name. Under the hood this is taxonomy v1.2, and every label figure, filter and API read uses it. A conversation whose last activity was more than 30 days old when v1.2 labeling started has no v1.2 label, so it reads as not labeled.