Note Check · Works with any AI scribe · Free tier available
Your AI Scribe Wrote the Note. Who Checked It?
Paste the visit transcript and the note any scribe drafted from it. Note Check reports, statement by statement, what the transcript supports, what it does not, and what it contradicts, plus the facts the note left out. It works on notes from Krasyn, Freed, Heidi, Nabla, Abridge, DAX, SimplePractice, Upheal, Mentalyc, or any other scribe, and it never edits the note.
Inside the free Krasyn Scribe account
Each check uses one AI-drafted note from your allowance. Free includes 10 notes and 500 transcription minutes each calendar month, with no credit card.
Our standard Business Associate Agreement is available for review and electronic acceptance inside the account.
Already have a Krasyn account? Open Note Check in the app.
What You Get
One report per transcript-and-note pair, built to be read in the order that matters and kept beside the note.
A verdict on each statement the note makes
The note is split into its sentences and each one is labelled Supported, Unsupported, Contradicted, Scaffolding, or Unverified against the transcript you pasted. Supported covers direct statements, faithful paraphrase, and clinical translation of what was said. Scaffolding is a header or a "not assessed today" line that makes no claim about the patient.
Three deterministic flags, produced by code
A number in the note that never occurs in the transcript. A "denies X" about a topic that was raised and never actually denied. Specific content filled in where the transcript only has an inaudible or audio-dropout marker. These flags come from plain rules, not a model, so they are identical on every run.
Omissions: what the transcript said that the note left out
Clinically material facts stated in the transcript and absent from the note, each with the short transcript excerpt that establishes the fact. A note can be faithful and still incomplete, and the two failures are reported separately.
Read-these-first ordering
Contradicted statements, unsupported statements, and high-severity flags are lifted to the top of the report so the lines that need your eyes are the first lines you see. Everything else follows in the note’s own order.
A plain-text report you can keep
The finished report exports as plain text, so it can sit beside the note in whatever system you document in. Note Check itself files nothing, edits nothing, and signs nothing.
An honest status line
Every report says whether the model judge ran, ran over only part of the transcript, failed, or did not run at all. Statements the judge did not assess are labelled Unverified, never quietly counted as Supported.
Accepted note sources: Krasyn, Freed, Heidi, Nabla, Abridge, DAX, SimplePractice, Upheal, Mentalyc, any other scribe that produces text.
A Checked Example
A synthetic transcript, a synthetic note, and the report Note Check produces for the pair. No real person, visit, or value appears here.
Note source label: "Other". The pair was written to show each verdict and flag once. This is a static rendering, not a live check.
Input 1 · Synthetic transcript
Clinician: Good to see you both. How has the blood pressure been at home? Patient: Mostly in the 120s over 70s on the home cuff. I brought the log. Clinician: Good. Any chest pain, shortness of breath, or swelling in the legs? Patient: No, none of that. Spouse: I am the one with the headaches lately, not him. Mine started last week. Clinician: We can talk about yours at your own visit. Any fever on your side? Patient: I haven't checked. I have felt fine. Clinician: Your LDL is still above where we want it. I would like to start atorvastatin, 20 milligrams at night. Patient: I would rather not start a statin yet. My brother had muscle pain on one. Can I try diet changes first? Clinician: That is reasonable. We will recheck the lipid panel in three months and revisit it then. Keep taking the lisinopril 10 milligrams every morning. Patient: Will do. Clinician: Let's see you back in three months, sooner if anything changes.
Input 2 · Synthetic AI-drafted note
Subjective: Patient presents for hypertension and hyperlipidemia follow-up. Home blood pressure readings mostly in the 120s over 70s, log reviewed. Reports new headaches beginning last week. Denies chest pain, shortness of breath, and leg swelling. Denies fever. Objective: BP 134/82. Assessment: Hypertension, controlled on current regimen. Hyperlipidemia with LDL above goal. Plan: Started atorvastatin 20 mg nightly. Continue lisinopril 10 mg every morning. Recheck lipid panel in three months.
Output · Note Check report
4 statements need your eyes first
Judge status: ran over the full transcript. Engine notecheck-1. Deterministic layer: ran.
- Supported
- 7
- Unsupported
- 2
- Contradicted
- 2
- Flags
- 2
Read these first
- ContradictedSubjective
Reports new headaches beginning last week.
Why: The headaches were described by the spouse about herself ("I am the one with the headaches lately, not him"). The patient reported no headache.
- ContradictedPlan
Started atorvastatin 20 mg nightly.
Why: The patient declined to start a statin and the clinician agreed to try diet changes first and recheck in three months.
- UnsupportedObjective
BP 134/82.
Why: No blood pressure value was spoken at the visit. The only readings discussed were the home log.
Deterministic flag num-001 · Number not in transcript: "134/82" appears in the note but no such value appears in the transcript.
- UnsupportedSubjective
Denies fever.
Why: The clinician asked about fever and the patient answered "I haven’t checked", which is not a denial.
Deterministic flag neg-001 · Denial not in transcript: Fever was raised in the transcript and never denied. Silence is not a denial.
Omissions
Return visit in three months, sooner if anything changes.
Transcript: “Let's see you back in three months, sooner if anything changes.”
Supported statements
- Subjective: Patient presents for hypertension and hyperlipidemia follow-up.
- Subjective: Home blood pressure readings mostly in the 120s over 70s, log reviewed.
- Subjective: Denies chest pain, shortness of breath, and leg swelling.
- Assessment: Hypertension, controlled on current regimen.
- Assessment: Hyperlipidemia with LDL above goal.
- Plan: Continue lisinopril 10 mg every morning.
- Plan: Recheck lipid panel in three months.
The four section headers (Subjective, Objective, Assessment, Plan) are labelled Scaffolding and are not counted as statements about the patient.
How It Works
Two layers, one fingerprint, and a status line that tells you what actually ran.
Step 1
The deterministic layer always runs
Pure code reads the note and the transcript first. It finds numbers in the note that never occur in the transcript, pertinent negatives about topics that were raised and never denied, and specific content placed where the transcript only has an inaudible marker. No model is involved, so this layer gives the same answer on every run and it runs even when no model is available.
Step 2
The judge runs at temperature 0 with a pinned prompt
A model judge then labels the remaining statements and looks for omissions. It runs at temperature 0 with a pinned prompt version. In our repeated runs it caught every planted error each time, though reason wording can differ and a borderline statement can move, so the finished report is stored rather than recomputed. When the judge did not run, failed, or saw only part of a long transcript, the report says which, and the statements it did not reach are labelled Unverified.
Step 3
The report is stored under a fingerprint of the inputs
The transcript, the note, the engine version, and the judge model are hashed into one fingerprint. The finished report is stored under it. Send the same transcript and note again from your account and you get the stored report back, without a second model call. Change one word and the fingerprint changes, so a stale report is never mistaken for a current one. Verdicts are decided by pure-code rules and a judge run at temperature 0. In our repeated live runs every planted error was caught each time and the pure-code flags were identical; the wording of a reason can differ, and a borderline statement can move between Supported and Unsupported on a fresh run, which is why the report is stored rather than recomputed.
Step 4
Nothing is edited, and nothing is overstated
Note Check reports. It does not rewrite the note, delete a sentence, or suggest replacement text. The summary counts are the counts: a report never states a percentage the underlying verdicts do not support.
What It Does Not Do
Stated here so you do not find it out from a report.
It does not make a note correct
A verdict is a comparison between the note and the transcript. Reading the report, deciding what to change, and signing the note remain the clinician’s work. Note Check never edits a note.
It cannot see what was never said
An exam finding the clinician observed but did not speak aloud has no transcript support and is reported as Unsupported. That is the honest label, not a judgement that the finding is false. Unsupported means "not in the transcript", nothing more.
It checks against the transcript, not the chart
Note Check compares the note to the transcript you paste. It does not open the patient chart, the medication list, or prior notes, and it does not know whether the transcript itself is complete or correct.
It is not a substitute for reading the note
The report orders the lines that most need attention, but it is a review aid. The deterministic layer is narrow by design, the model judge can be wrong, and the omissions list is bounded to material facts. Read the whole note.
We publish no accuracy figure for Note Check
We have not measured Note Check against an adjudicated reference set, so we do not state how often it is right. Our scribe drafting path has a published faithfulness benchmark with its limitations written out, and that benchmark measures the scribe, not Note Check.
The full list of things Krasyn does not do, generated from our claim-evidence system, is on What We Don’t Do. The scribe benchmark and its limitations are written up in the AI scribe faithfulness benchmark.
Why This Exists
Every ambient scribe drafts notes. The clinician who signs carries the responsibility for what the draft says, and the published evidence is that drafts are rarely edited before signing. These are the four external findings we rely on, and the only ones we cite.
Annals of Internal Medicine, April 2026
A cross-sectional evaluation across 11 AI scribe tools found AI-generated notes lower in quality than clinician-written notes.
Published ambient-scribe evaluations
Reported hallucination rates for ambient scribes run about 1 to 3 percent per note, with physical-exam and plan content the highest-risk areas.
Across 71,173 pairs of AI-drafted and finalized note sections, only 5.8 percent of sections showed a confirmed edit before the note was finalized.
Federation of State Medical Boards, January 2026 position
The signing clinician holds full legal responsibility for the note. No ambient AI vendor accepts clinical liability for what its draft says.
These findings are about AI scribes as a category. None of them is a measurement of Note Check, and we do not publish an accuracy figure for Note Check.
Frequently Asked Questions
Which AI scribes does Note Check work with?
Any of them. Krasyn, Freed, Heidi, Nabla, Abridge, DAX, SimplePractice, Upheal, Mentalyc, or anything else that produces a note as text. You paste the transcript and the note. Note Check does not connect to another vendor’s system and does not need to.
Does Note Check fix the note?
No. It reports. Each statement gets a verdict, the deterministic flags and the omissions are listed, and the lines that most need attention are placed first. What to change, and whether to sign, stays with the clinician. The note you pasted is never edited.
Will I get the same report if I run the same check twice?
Within your account, yes. The transcript, the note, the engine version, and the judge model are hashed into a fingerprint, and the finished report is stored under it. The same inputs return the stored report, identical to the first one, without a second model call. Change anything in either text and a new report is produced. The pure-code rules give the same flags on every run anywhere; the judge runs at temperature 0, and in our repeated live runs it caught every planted error each time while reason wording varied and one borderline statement moved, which is why the report is stored rather than recomputed.
What does it cost?
Each check uses one AI-drafted note from your Krasyn Scribe plan allowance. The free tier includes 10 AI-drafted notes and 500 transcription minutes each calendar month, for one clinician, with no credit card. Paid tiers with larger allowances are listed on the scribe page.
What do I need to paste?
A transcript, either speaker-labelled or plain, and the note in any format: SOAP, DAP, BIRP, or narrative. Note Check compares the note only to the transcript you paste. It does not read the patient chart.
What does Unverified mean?
The model judge did not return a verdict for that statement. That happens when the judge did not run, failed, or saw only part of a long transcript. The report states which of those occurred. Unverified statements are never counted as Supported, and the deterministic flags still apply to them.
Explore More
Krasyn AI Scribe
The standalone scribe the free account comes with, and its three tiers.
Pressure Test
The seven-point worksheet for evaluating any scribe on a synthetic session.
Faithfulness Benchmark
How we measure our own scribe against its transcripts, limitations included.
What We Don’t Do
The capabilities we have deliberately not built or will not claim yet.
Check One Note Today
Paste a transcript and the note your scribe drafted from it. Read the report. Decide what to change. Free includes 10 notes a month, no card.