Choose how far the AI is allowed to reason.
Evidence-based AI medical analysis is one of three reasoning modes — Evidence-Based, Interpretive, or Extractive — that you choose per query. The mode changes how much the AI infers, never whether it cites: all three hold one citation standard, and unsourced lines never pass silently.
Extractive: verbatim, and nothing else.
Extractive mode pulls the record’s exact language — quotes, values, dates — each with the page it came from. No paraphrase, no inference, no smoothing. Use it when only the record’s own words will hold up.
Evidence-Based: conclusions built only from cited facts.
The default for most legal and claims work. The AI connects facts across documents and draws a conclusion, but every step in the chain is a cited fact from the file. If a step can’t be sourced, the chain stops and the gap is flagged.
Interpretive: clinical reading, clearly labeled.
Interpretive mode reads between documents the way a clinical reviewer would — context, significance, what a pattern of findings suggests. Interpretation is labeled as interpretation, kept separate from fact, and still anchored to the cited record underneath.
The mode changes the reasoning. Never the receipts.
Whichever mode you pick, the output holds one standard: every factual statement cited to its page and source, and anything unsourced flagged for human review. That is what makes the work audit-grade and legally defensible — in any mode, at any depth.
See Verifiable AI CitationsPick the depth. Keep the receipts.
Three steps from question to cited answer, at exactly the reasoning depth your work calls for.
Extractive, Evidence-Based, or Interpretive — set per query, or as a default for the matter or team.
A plain-English question or a full work product. The AI reads the record at the depth the mode allows.
Every line links to its source page. Flagged lines show where the record ran out, queued for your review.
Who controls the reasoning depth.
Different files call for different inference ceilings — the citation standard stays the same.
Extractive for the record cite, Interpretive for clinical context — kept clearly apart.
For IME workEvidence-Based for demands and motions; Extractive when only the record’s words will do.
For law firmsOne consistent reasoning depth across every reviewer, client, and file.
For TPAsA controllable, auditable answer standard for record review at claims volume.
For carriersReasoning modes, answered.
Extractive returns the record’s own words: direct quotes with page citations and nothing added. Evidence-Based draws conclusions, but only from facts it can cite, with every reasoning step traceable. Interpretive adds labeled clinical reading and context — and the facts underneath it are still cited.
Most teams file from Extractive or Evidence-Based output: Extractive when you need the record verbatim, Evidence-Based when you need cited conclusions. Interpretive text carries its interpretation label with it, so nothing labeled as reading is mistaken for record fact. The AI drafts; your team decides what gets filed.
No. The mode changes how much the AI may infer, never whether it cites. In all three modes every factual statement traces to a page and source, and statements that cannot be sourced are flagged rather than written as fact.
Yes. The mode is set per query, so you can ask the same question of the same file in all three modes and compare the answers. You can also set a default per matter or per team to keep output consistent.
The mode acts as an explicit ceiling on inference. Reviewers always see which mode produced an answer, every line carries its citation, and flagged lines show exactly where the record ran out — a reasoning trail you can defend under scrutiny.
Related capabilities.
Same engine, same citation standard — other ways the platform reasons over the same cited record.
An evaluation framework scores every output and traces citations end to end — unsourced lines are flagged.
ExploreAsk the record anything, one file or a thousand — multi-step answers, every statement cited.
ExploreFavorable and unfavorable fact patterns flagged in the record, cited — signals, not a verdict.
ExplorePick a mode. Every answer still shows its sources.
Upload one file and ask it the same question in all three modes — the cited sample comes back the same day. Handled under our BAA; never used to train a model.