Presenter script

Synthetic data only. No funding, savings, outcome or time-saving claims. Start at the demo page; open each case in order.

1. Heart failure: the type is not written · chart P29

What you will see: An AI-drafted progress note says only “CHF”; the record holds two echocardiograms (35%, then 30%).
What to click: Open the case, read the draft, then in the side panel open the options and click the wording you choose.
What to say: The panel shows the record's own ejection fractions with dates and four published categories with what each needs; the clinician picks, and one click places that wording in the draft.
What the judge should notice: Nothing is pre-selected; “other”, “unable to determine” and “not clinically supported” are always there; acuity indicators (symptoms, NT-proBNP change, weight change, IV diuretic) are evidence only.
Related planted cases: P30 (the note says preserved; the latest value is 35%); P33 (the note says reduced; the latest value is 48% after 30%); P31 (agrees); P32 (no ejection fraction on file: not assessable); P41 (the value is only in the echocardiogram report text: read and quote-verified by two readers); P42 (the report gives a range (35-45%) that spans categories: not assessable); P43 (the report is about the right ventricle: no left-ventricular value, not assessable)

What you will see: A clinic note says “CKD” with no stage; filtering rates 90+ days apart sit in one stage range and urine albumin is raised twice.
What to click: Open the case, look at the kidney card's evidence and its albumin-to-creatinine category, then choose a stage or dismiss.
What to say: Stage ranges come from versioned rules with the 3-month persistence rule; one value is never shown as a confirmed stage.
What the judge should notice: Both directions: P27's note states a stage the results do not reach, and P28's states a stage the results do not support.
Related planted cases: P27 (stated stage higher than the results); P28 (stated, results normal); P20 (met, not mentioned); P21 (dialysis on record: not assessable)

What you will see: The dietitian documented intake, weight history and physical findings; the clinician's note does not state a nutrition diagnosis.
What to click: Open the case, read which of the six characteristics are documented and at which level, then choose or dismiss.
What to say: It triggers only on a dietitian's assessment, computes weight-loss percentages from the weight history, and quotes the dietitian's findings with author and date. Albumin is shown as not a criterion.
What the judge should notice: Two or more characteristics are required; P36's note says severe but only moderate-level characteristics are documented.
Related planted cases: P36 (note says severe; assessment reaches moderate); P37 (agrees); P38 (stated with no assessment); P39 (too few characteristics assessed: not assessable); P44 (a narrative assessment only: its statements read and quote-verified by two readers)

What you will see: Type a reason for visit (“worsening knee pain”) and the panel lists earlier visits, imaging, physical therapy and medicine trials from about two years of the chart.
What to click: Open the case, edit the Reason for visit box at the top of the draft, and watch “Relevant history” update.
What to say: Every result is a real chart item from the whole record, not just the last visit, with the matching sentence quoted and what it matched on.
What the judge should notice: It is plain text matching with a small concept dictionary: explainable, no model required, informational only.
Related planted cases: P29 (“shortness of breath and leg swelling”); P35 (“fatigue, poor appetite and weight loss”); P34 (“follow-up of kidney function”)

Other conditions from the definition library: Glucose and HbA1c, blood pressure, body measurements and the kidney check: the same both-directions check on every synthetic chart, from definitions with openly usable sources.
Coder view and worklist (staff): Codes and code categories appear only here, for coding staff; compliant query drafts are linted for leading wording.
Definition library and sign-off (staff): Every definition with its source, edition and licence status; each stays “Draft · pending review” until approved here.
Replay: Every check re-runs from its stored inputs under its stored rule versions to the identical hash.
Results: Agreement against synthetic labels, misses and false alerts side by side (simulated, not measured).
How it decides: Rule cards for the kidney check with sources, thresholds and live boundary tests.
The record shell: The full ward and clinic list in the demonstration record.
Audit trail: Every click, choice, dismissal and model call, hash-chained and verifiable.

25.8% of VA heart-failure diagnoses in FY2024 were coded unspecified (VA Diagnosis FY24 open data; unit inferred as patients per code; all settings pooled)
56.3% of veterans with lab-defined CKD in one VA network were not recognised (2012-2019) (Bansal et al., Kidney360 2020; one network; quote the table)
36.4% of admitted veterans at one VA hospital were nutritionally compromised at admission (n = 295) (Cassatt and Parker, Innov Aging 2023 (abstract); one site; screening is not diagnosis)
about 5 minutes of chart review per ambulatory visit in audit-log studies (Overhage and McCallie, Ann Intern Med 2020; our own time effect is not measured: say we will measure it)

“We don't claim savings; a pilot would measure recognition rates, clinician acceptance and false alerts, and coder agreement.”
Every count in this build is SIMULATED on synthetic data.
One click inserts only the wording the clinician chose from a neutral list; the software never inserts anything by itself and never signs.