Every answer points at an approved document — or there isn’t one.

A question goes to the approved snapshot and the answer comes back with a numbered marker on each claim, pointing at the document and the page that supports it. Every substantive answer carries a document- and page-level citation and a link to the official manufacturer source.

Real

Live retrieval over the approved snapshot — the partner's own documents for the four retained topics, ingested with documented content rights. No answer is scripted, and the citations come out of the retrieval step rather than being attached afterwards.

Faked

Nothing in the retrieval path. The snapshot is fixed at ingestion, so a document approved after it is cut is not there until it is cut again by hand.

Two approved revisions of the same document can give different answers, and the same question can turn on which device or model is in the room. Rather than answering from whichever it retrieved first, the product resolves it or asks which one applies to the set in front of you.

Real

The behaviour, across at least three document pairs whose approved answers genuinely differ. Every question the pair touches either resolves to the selected revision or asks which revision applies.

Faked

Possibly the pairs. The partner confirms three; where the corpus cannot supply them, constructed examples are used and shown as constructed, and that surface's pass condition is waived.

The behaviour that separates this from a search box. Where the snapshot holds no approved source, the product declines and logs the question instead of assembling something plausible out of what it did retrieve.

Real

Abstention and the logging behind it, on unsupported questions and on the five topics removed from scope alike. It carries the highest pass condition in the build — 90 per cent.

Faked

Nothing. This surface is real in the demonstration.

A question about use outside the cleared indications is refused rather than softened. The product builds a handoff record carrying the question verbatim, the sources it retrieved and the rep's user record, and routes it to Medical Affairs.

Real

The refusal, the detection against a partner-approved term list, and the handoff record itself.

Faked

The destination. Medical Affairs is a mocked endpoint rather than an integration — the record is built and shown, and the send is stubbed.

Complaint language is caught before it becomes a chat message. A device that did not perform as intended opens a structured event routed to Quality — and no complaint content is retained after the transfer.

Real

Detection, the structured event, and the clearing of the content after transfer. Watch the thread empty when you send it.

Faked

The destination. The quality system is a mocked endpoint and receives nothing.

Two ways in that do not require typing. Scoping to a product or procedure narrows retrieval to that family's documents, and a question asked outside the scope abstains instead of quietly widening the search back out. Under it sit partner-approved pre-formatted questions — the ones a rep is actually asked, phrased so the answer is known to be supportable.

Real

The scoping and its effect on what is retrieved — not a filter drawn over an answer that was going to be the same either way.

Faked

The topic set and the questions under it. Four topics on a single manufacturer tenant; five more were removed from scope on 31 August.

Short, practical notes attached to a topic, written by the manufacturer rather than generated — the kind of thing a rep can hand to a surgeon and look good doing it. The surface exists so the judgement a field team already has ends up somewhere other than a group chat.

Real

The surface, its placement in the product and its per-topic scoping.

Faked

The tips. Manufacturer-authored in the build; the three shown here are illustrative and would be replaced before any showing.

Every submitted question is persisted with its answer status, the sources it cited and any feedback left on it. The content-gap and unanswered-question view is how that record is read: what was asked, how often, and why it could not be answered.

Real

The persistence and the gap view over it. Anything you ask in the app above lands in it as you ask it, including the questions it refuses.

Faked

Its starting contents. The three entries it opens with are seeded, and marked seeded.

A branded, responsive TrainMed PWA a rep installs to the home screen, running on the CareIQ application shell behind a dedicated TrainMed authentication and session boundary — not a tab inside somebody else's product.

Real

The app, the branding and the session boundary. It deploys to staging in CareIQ-controlled infrastructure and hands over with source, configuration and a technical guide.

Faked

Nothing in the boundary itself. SSO is out of scope, so sign-in is TrainMed's own and there is no enterprise directory behind it.