The review of systems lists normals nobody asked about

Take one note with a suspicious review of systems and search the transcript for each negative it asserts. Denies chest pain. No shortness of breath. Bowels normal. If those words are not in the audio in any form, the section was not transcribed — it was completed. That distinction decides everything that follows, because a completed section is a structural fault in the template and no amount of instruction fixes structure.

The mechanism is not mysterious. A review of systems presented as a dozen-odd rows, each with an expected value, is a form. Give a language model a form and evidence for 2 rows, and the pull of the remaining 10 is towards the unremarkable answer, because that is what completes the pattern. The model is not lying about what it heard; it is finishing a shape it was handed. Change the shape.

Confirm it is completion before you accept it is mishearing

1. Take 5 notes flagged by clinicians and pull the matching transcripts. 5 is enough, because completion is systematic — if it is happening it will be in all of them.
2. For every asserted negative, search the transcript for the symptom word and for its common spoken forms. 'Denies chest pain' can legitimately come from 'any pain in your chest at all?' followed by 'no'.
3. Mark each as evidenced, ambiguous or absent. Absent means the symptom is nowhere in the audio, in any phrasing, by either speaker.
4. Count absent negatives per section. If the assessment has none and the review of systems has 6, the difference between those sections is the cause — one is prose and one is a list of rows.
5. Now check whether the same negatives appear in the patient's previous note. If they do, verbatim, you have a second cause running alongside the first and both need fixing.

The output is a count, and the count is what makes this arguable in a clinical governance meeting. 'The scribe makes things up' cannot be acted on. '11 of 14 negatives in these 5 notes have no basis in the audio, all of them in one section' can.

Where a normal nobody said actually comes from

Only the last row is a model-quality issue in the usual sense, and it is the least common of the 4 in practice. The first 3 are all decisions somebody made in the design of the note — which sections exist, what gets passed in as context, and what the system is told a good note looks like. All 3 are cheap to change and none of them requires touching the model.

Yesterday's normals arriving as today's

Passing the previous note in as context is a reasonable-sounding idea with an unreasonable failure mode. It helps a model resolve references — 'the same tablets', 'that shoulder' — and it also hands over a fully-formed review of systems that fits the template perfectly. When the current consultation supplies nothing for those rows and the prior note supplies everything, the prior note wins. What comes out is indistinguishable from copy-forward, which is frequently the exact problem the clinic bought ambient documentation to escape; how a chart fills with text nobody wrote today is the fuller account of why that matters.

Make 'not discussed' a value the note can print

The structural fix is one sentence long and most templates violate it: a row with no evidence must render differently from a row with a negative finding. A blank row reads as normal to every clinician who has ever used a paper form. An absent row reads as normal too. Only an explicit marker — not discussed, not asked, not examined — reads as what it is. That means the template needs 3 states per row rather than 2, and the note renderer needs to print the third one rather than tidying it away.

• Give every row 3 states: asserted positive, asserted negative, not discussed. Two states force the system to choose a clinical claim where it has none.
• Let whole sections be absent. A consultation about a knee does not need a genitourinary row, and printing 12 rows to fill 1 trains reviewers to skim.
• Render the third state visibly in the signed note, not only in the draft. If it disappears at signature, the chart carries the same ambiguity you started with.
• Sort unevidenced rows together rather than interleaving them. A reviewer scanning a block of not-discussed rows can act on it; a reviewer hunting them among negatives cannot.
• Count them. The number of not-discussed rows per note is the one operational signal that tells you whether capture or template is drifting, and it costs nothing to record.

A blank row and a negative finding look identical on paper and mean opposite things. The template's job is to make them look different, because the model cannot invent the difference.

What the clinician adds at review, and what the system must not

There is a legitimate route for pertinent negatives to enter a note, and it runs through the person who was in the room. A clinician reviewing a draft can add 'denies red flag symptoms' because they asked and remember the answer; the system cannot, because it has only the audio and the audio does not contain it. Design the review step so adding a negative is one action per row rather than an edit to a paragraph, and the clinicians will do it — the friction, not the principle, is what stops them.

That only works if review happens while the consultation is still recoverable from memory. A draft reviewed the same afternoon gets accurate negatives added; a draft reviewed on Friday gets the not-discussed markers deleted so the note looks tidy. This is one of the less obvious reasons draining the unsigned-note queue the same day matters, and why it should be built before any work on note content.

The template is usually inherited, not designed

Almost every fixed-row review of systems in a clinic descends from a paper proforma, where a printed list of systems made sense — paper cannot branch, so every possible row has to be on the page whether or not it applies. Carrying that layout into a generated note imports the completion pressure along with it. Converting it properly, keeping the clinical intent and dropping the rows that exist only because of the medium, is the method in turning a paper proforma into a template a model can fill.

Across a multi-doctor list the same template serves clinicians with different examining habits, which is where a shared structure with per-clinician sections earns its keep; the session-level mechanics of that are in running ambient documentation across a shared outpatient room list. And before changing anything, record what the note currently looks like — measuring a clinic before you change anything is the baseline that lets you show the negatives count went down rather than assert it.

Where the template and renderer are your own software, all of this is a small, well-defined change: 3 states per row, a printable third state, sections that may be absent, and prior notes demoted from evidence to reference. That is the shape of work we mean by MVP and product builds. Where the note structure belongs to a vendor, the question to put to them is narrow and answerable: what does your note print when a section had no evidence at all? The rest of the drafting-to-signature chain sits in ambient documentation and the clinical note, within our healthcare and MedTech work.

• review of systems
• note templates
• ambient documentation
• clinical safety