HomePlatformVerifiable AI Citations
Verifiable AI Citations

AI medical summaries with a citation on every line.

"If we can't cite it, we don't say it."

Every sentence the platform writes is locked to its source — click any line to jump to the exact page or DICOM slice it came from. An evaluation framework scores every output, and low-confidence findings are flagged, not hidden. The category's hallucination fear, answered with a mechanism instead of a promise.

Interpretive summary

Onset of low-back pain on 03/14/2025 p.412, MRI confirms L4–L5 herniation MRI·s3·sl18. A 6-week gap appears between visits flag.

p.412 · Progress Note — St. Mary'sCited
"…acute onset of low-back pain radiating to the left lower extremity."
3 citations · 0 uncited claimsclick any citation →
3 modes extractive · evidence-based · interpretive — you pick the posture
100% of claims traced to a page or DICOM slice
1 click from any line to the source it came from

From claim to source, and back.

Pipeline
01
A claim is written

A summary line, a chronology entry, an answer.

02
Locked to source

No source, no sentence — uncited claims are blocked.

03
Scored

An evaluation framework rates confidence; low scores are flagged.

04
Clicked to verify

Jump to the exact page — or the exact DICOM slice.

Trust as a product mechanism.

Click-through citations

Any line in any summary, chronology, or answer jumps to the page or slice it came from. Verification is a click, not a re-read.

Three reasoning modes

Extractive (zero inference), Evidence-Based (every statement cited), Interpretive (clinical synthesis, still cited) — you pick the epistemic posture per deliverable.

Confidence scoring

An AI evaluation framework scores every output; anything below threshold is flagged for your review — never silently included.

Works on imaging too

Citations resolve to DICOM study, series, and slice with the same precision as a page number — unique in the category.

PHI access logging

Every access to every page is logged, with case-level minimal access — an audit trail for the audit.

You attest

Edit any finding, regenerate, and sign. AI does the reading — the calls are always yours, and the work product carries your name.

Works with
Citation-locking isn't a feature you turn on — it's the constraint every other feature runs under.

The defensibility spec sheet.

Evidence Lock™
Uncited claims
Zero allowed

If it can't cite it, it doesn't say it — enforced in generation, not review.

Reasoning modes
Three

Extractive · Evidence-Based · Interpretive — the epistemic posture is yours to set.

Citation depth
Page & slice

Document page or DICOM study · series · slice — one click to the source.

Confidence
Scored per finding

Below threshold → flagged for your review. Nothing hidden.

Audit trail
PHI access logged

Every event, case-level minimal access, HIPAA controls, SOC 2 audit-ready.

Final authority
You, attested

Edit, regenerate, sign. AI does the reading; you make the calls — always.

Built for work that gets challenged.

When your output faces cross-examination, an audit, or a records request, "the AI said so" is not an answer. A citation is.

FAQ

Verifiable AI, answered.

With a mechanism, not a promise. Generation is citation-locked — a statement that can’t be tied to a source page or DICOM slice doesn’t get written. On top of that, an AI evaluation framework scores every output, and anything below the confidence threshold is flagged for your review instead of delivered as fact.

Extractive pulls verbatim content with zero inference. Evidence-Based writes conclusions where every statement carries a citation. Interpretive adds clinical synthesis — still cited — for reviewers who want the connective tissue. You choose the epistemic posture per deliverable, so a demand letter and an IME opinion don’t share the same risk profile.

That’s what it’s built for. The reviewing professional attests and signs the final work product; every line click-throughs to its source; and PHI access logging records who saw what, when. Your answer to "how do you know?" is a click — not a shrug at a black box.

They’re flagged and surfaced for your review — visibly. Nothing below threshold is silently included, and nothing is silently dropped: the flag itself appears in the workspace so a person makes the call.

Yes. Every PHI access is logged per event, access is scoped per case, data is encrypted in transit and at rest under HIPAA controls, and the platform is SOC 2 audit-ready. Your data is never used to train a model.

Yes — citations resolve to the DICOM study, series, and slice with the same precision as a page number. A finding about the L4–L5 herniation links to the exact slice, opened in the built-in viewer.

Verify us on your own file.

Upload a file, get a cited summary back, and click every citation. Handled under our BAA; never used to train a model.