Every answer is cited#
The portal never gives a bare, unattributed answer. Each factual claim carries a bracketed citation marker like [1] that links to the exact source passage, and the sources are listed beneath the answer. Citation numbers are assigned by the application, sentence by sentence, from the platform's own source attribution and the cited papers' text: a sentence keeps a marker only when the cited paper carries its words, its figures and the names it hangs on, so the number you click resolves to a passage that grounds that claim. Markers never sit on headings, and an item in a list takes the citation of the paragraph it belongs to.
Every figure is checked before you see it#
Where an answer quotes figures, Ask checks them against the cited passages before the answer is complete. Every number, percentage, dose and range is located in the cited paper, and the sentence or table row that carries it must share the claim's own quantity, about the same outcome, at the same follow-up, with the same responder threshold and the same denominator. A figure the cited passage does not carry is looked for in the full text of the papers retrieval found, and if one carries it the sentence is cited to that paper instead - and, under a restricted retrieval, only ever to one of the papers the question named. A citation marker is only ever left on a paper whose passage was where the sentence's figures were found: a marker on a paper that does not carry the figure is dropped, and a sentence left with no marker is removed. A figure the cited paper only quotes from other studies is removed rather than footnoted when the question names a cohort.
A sentence whose figures cannot be verified anywhere is removed, and the answer says so in a note beneath it, naming the figures. A sentence cited to the wrong paper is replaced by the named paper's own sentence only when that sentence carries the same figure at the same time point, quoted verbatim and cited; a decline is never replaced. A denominator the answer paired with a figure is part of the figure: a pairing the paper contradicts is removed and said so, never rewritten, and a denominator is only ever added from the figure's own bracket or table cell. A year must come from a cited resource, and a medication called contraindicated must be called that, by name, in a cited passage: the verb is read with the medication nearest it, and a passage that only calls the drug "not recommended" does not carry the stronger word. When a sentence is removed, the conclusion that rested on it goes with it, and an answer with nothing left to cite is declined rather than shown.
While the answer streams, its text is shown as unchecked, in muted ink with an "Unchecked - still streaming, the check follows" mark, so nothing on screen reads as the answer before it has been checked. Its first complete sentence is checked against the papers retrieval found and, when it passes, "First sentence verified against" the paper appears under that sentence; "Checking N figures" shows until the checked answer replaces the streamed text in full ink. Document chat answers carry the same check and badge, against the open document's text. The badge beneath the finished answer then reads what happened: "N figures checked", with any sentence removed or replaced counted beside it and the figures named on hover. If nothing verifiable is left, the answer is withheld and the portal says which figures could not be verified rather than showing them.
The confidence signal#
Every finished answer carries a confidence control on its actions row, at the right beneath the answer. It reads High confidence, Moderate confidence, Low confidence or Confidence not scored, and it stays loud and labelled while the news is bad: a low-confidence answer shows a warning you cannot miss, a high-confidence one a quiet tick.
The label is led by the portal's own check of the answer against the cited papers. An unverified figure, year or contraindication makes it low; removed sentences cap it at moderate; every figure found beside its claim, with most sentences carrying a citation, makes it high. High confidence is earned only by that check, never by a score alone.
Open the control to see what sits behind it, including the platform's self-assessment of the answer across three dimensions - how well it answers your question (answer relevance), how firmly it is grounded in the sources (groundedness), and how relevant the retrieved context was (context relevance) - each shown as a plain score out of five with a one-line reading. Those scores can lower the label one step when groundedness is weak, and never raise it. When the check had nothing to judge and the scorer did not run, the control says so rather than guessing.
When confidence is low, the same panel offers to re-answer the question deeply, against the full text of the matching documents, so a thinly grounded first answer has a direct path to a firmer one rather than leaving you at a dead end.
Honest refusals#
If the corpus does not hold enough relevant material to answer confidently, the portal says so before generating anything, names the closest matches it found in the text and lists them beneath the answer as closest matches, not used, rather than bluffing an answer. A question about a relationship no held paper studies is declined as a boundary, and a paper the question names is read directly before anything is declined. An honest "no direct evidence found" is a feature, not a failure.
The evidence behind an answer#
Beneath an answer the evidence list shows every source it drew on. Each one quotes the paragraph that carries the claims cited to it, with the page it sits on, and a relevance score as a percentage, so you can weigh the sources at a glance; a weak match is labelled as one, and a bibliography paragraph is never shown as evidence.
When you want the AI's read on each source, open Journey through the context. It adds a short verdict to every source - for example Supports, Partial or Not relevant - which then appears alongside the evidence. Working this out costs a little time, so it runs only when you ask for it, never automatically on every answer. The verdicts are advisory, a quick steer rather than the last word, so open a source to judge it for yourself.
Passages the answer did not rely on are still listed under Also retrieved, and you can open any source in place. Together this lets you audit an answer rather than take it on trust.