Clinical calculators computed in code, not recalled by the model

Editor's note (2026-09-06): this post originally described the note as carrying the computed value. That is true for the calculators an encounter requires, and not for every one of the 236 computed in code: the self-report questionnaires (PHQ-9, GAD-7, C-SSRS, CAGE, AUDIT-C) and CHA2DS2-VASc are never written into a note automatically, and whether any other computed value is written depends on how strongly the encounter calls for the calculator. The claim Scribeable makes is computation: the score is computed in code, not recalled by the model, and a detected disagreeing stated score is flagged for correction, except CHA2DS2-VASc, whose stated score is surfaced for adjudication rather than corrected. The body below is left as published; the title, excerpt and SEO metadata were updated 2026-09-06.

Here's a thing a language model will happily do: write "HEART score 3" into a note whose documented components add up to 5. It isn't lying, it's doing what language models do, which is produce the number that usually appears in that sentence. If you aren't reading closely you sign it. That's the whole problem with letting the model that writes your note also do your arithmetic. The prose is fluent, and fluent is exactly what a wrong number looks like.

So Scribeable doesn't let the model do the math. Its registry is 239 clinical calculators, and every one that carries a score is scored by code, not guessed by the model.

Where in the pipeline this happens

The scribe drafts the note, then a separate verification pass reads it back. In between, any calculator the encounter mentions or implies gets handled deterministically. If you said "qSOFA of 2," the components (respiratory rate, mental status, systolic pressure) are pulled from the transcript and the chart, the score is computed in code and the note carries the computed value, not the remembered one. Where a calculator is offered by the encounter rather than required by it, it is scored only when every component actually resolved: if the inputs aren't all there, nothing is asserted, because a score the note can't support is worse than no score at all.

Three calculators are further restricted, each because it turns on a judgment the chart cannot make for you. CHA2DS2-VASc is never written into your note automatically at all: its computed value only ever surfaces a disagreement for you to adjudicate, because its components come from history the encounter can't verify and a resolver that guessed wrong about a comorbidity should not be moving an anticoagulation decision. You can still fold it in yourself from the panel below. PERC and PREVENT are never scored in the background at all; they exist only as forms you fill in. PERC-negative means "PE ruled out" only once you have judged the pretest probability low, so a silent computation over absent inputs would produce that conclusion out of nothing. PREVENT is validated for primary prevention, and whether a patient already has cardiovascular disease is your call, not the chart's.

Activation is context-aware. The calculators that fit the encounter and your specialty run; a psychiatrist doesn't get a Wells score offered on every visit, and an ED note doesn't get a PHQ-9 for a laceration. And for every input, you can see where the value came from. In the web cockpit, the Sources view shows whether each component came from the transcript, the chart, or a clarification you answered, and it says so when a value was assumed or merely reported in the note rather than measured, and the Add-to-note panel shows the quoted text behind each value it prefilled, so when you review the score you're reviewing the evidence, not trusting a sentence.

When the note never mentioned the calculator

Sometimes the score you want isn't one the conversation brought up. You're finishing an atrial fibrillation visit and you want the CHA2DS2-VASc in the assessment, but nobody said the words out loud. For that there's a manual calculator panel, in the web cockpit: from the note's "Add to note" row you can reach for one of 18 clinician-reviewed calculators. It pre-fills the inputs the chart already documents, shows you where each value came from, and you fill in whatever it couldn't find. Then you review the change before the score is folded into your note. It's review-then-fold, on purpose; I didn't want a button that rewrites your assessment while you're looking at something else.

Why I care this much about a number in a note

A risk score in a chart is a decision. It's the reason someone was or wasn't anticoagulated, admitted, or sent home. If an AI scribe is going to put one in a note that carries my signature, I want the arithmetic to be the boring, checkable part and the judgment to stay mine. The scribe extracts, computes, and shows its work. I decide what the score means for the patient in front of me. That division of labor is the version of clinical decision support inside a note that I'd trust, and it's the version we built.

The transparency page has the architecture. Start the 14-day trial at scribeable.ai and say a score out loud in your next dictation; then look at the Sources view. No sales call; a card on file, and nothing is charged if you cancel before the trial ends.