DEV Community: Krasyn EMR

The check got better and got a new way of being wrong

Krasyn EMR — Sat, 05 Sep 2026 17:26:56 +0000

Scribe red-team, installment 02. Published 2026-09-05 by Krasyn.

What this series is

Every two weeks we run Krasyn Note Check, the part of our product that reads an AI-written clinical note against its transcript, over data we did not write, and we publish every disagreement. That includes the disagreements where the check is the one that is wrong. The two series rules that shape this installment are these: sources must carry an open licence we can name, and when the engine version changes, the previous installment's exact inputs get rerun on the new engine before any new data is touched.

Both rules bit this time.

The therapy source we did not run

Installment 01 announced that installment 02 would take a therapy or behavioural-health source, because there is a real gap there. Studies of ambient documentation report accuracy falling in psychiatry and behavioural encounters compared with general medicine, and no published error data exists for AI-generated therapy notes anywhere. Behavioural health is also the segment where patients are most anxious about what a machine writes down about them.

The candidate was AnnoMI, a collection of expert-annotated motivational interviewing conversations transcribed from public demonstration videos. It is a good dataset for the question. It fails our licence rule.

Read on 2026-09-05: the GitHub API for uccollab/AnnoMI returns "license": null. The repository root holds AnnoMI-full.csv, AnnoMI-simple.csv and README.md, and no LICENSE, LICENSE.md, LICENSE.txt or COPYING. The README asks to be cited. A citation request is not a grant of terms, and "publicly available" is not a licence name.

Our rule says the licence has to be nameable and reproduced in every piece. It is not. So we did not run it, and this section is the record of that reading, which the rule also requires.

What we ran instead keeps the subject and changes the corpus. The MIT-licensed Omi Health set we used in installment 01 ships a file describing which body systems each of its 300 dialogues touches. 52 carry a psych tag. We took the eight with the densest behavioural-health content, excluding the six installment 01 already used, and paired each with notes from three named model writers: GPT-5.5, Claude Opus 4.8 and DeepSeek-V4-Pro. 24 pairs.

These are primary-care and specialty visits with substantial behavioural-health content, not psychotherapy sessions. The gap that motivated the therapy installment is still open. We are still looking for a therapy transcript source with a licence we can read.

First, the fixed set

Between installment 01 and now, the Note Check engine went from notecheck-1 to notecheck-6. The series rule says the previous installment's exact set gets rerun first, so that engine changes are visible against a fixed baseline rather than against new data that might be easier or harder.

We regenerated installment 01's 36 pairs from the same pinned corpus commit with installment 01's own preparation script, and confirmed the statement text came out identical in 36 of 36 pairs before comparing anything.

Metric	notecheck-1	notecheck-6
Statements judged	1,401	1,401
Supported	776	737
Unsupported	3	35
Contradicted	0	0
Scaffolding	622	629
Deterministic rule flags	16	2
Omissions listed	85	11

Installment 01's headline finding was that the AI judge inside Note Check accepted assessment and plan content the visit never contained. Every statement we named there is now marked Unsupported. The cough note that invented an allergic cause in a visit that ended before any conclusion. Two notes that wrote "Type 2 diabetes mellitus" over a doctor who only ever said "your diabetes". A plan line adding smoking cessation counselling nobody proposed. A note that wrote "Continue metformin" when the plan was to adjust medications.

Installment 01's other headline was that the pure-code negation rule fired sixteen times and thirteen of those were wrong, most of them on a patient answering "Correct, none of those" to a list of contraindications. Those thirteen are gone. The two flags that survive are exactly the two we read as correct.

That is a real improvement on a fixed set of bytes, and it is checkable line by line in baseline-rerun/engine-comparison.json.

Then, the new way of being wrong

Thirty-five Unsupported verdicts where there were three. Four of them are wrong, and all four are wrong the same way.

Example 1

Source: Omi Health medical-note-eval (MIT), dialogue_24, note by GPT-5.5.

Transcript: "S14: I have your recent lab results here, your A1C has gone up to 8.2% from 7.5% last time."

Note sentence: "Hemoglobin A1C increased to 8.2% from 7.5% previously." The note has it under Exam.

Note Check verdict: Unsupported. Reason given: "The hemoglobin A1C is a lab result, not part of the physical exam."

Our reading: Note Check is wrong. Both numbers and the direction of change are in the transcript verbatim. The reason concedes the fact and objects to which heading it sits under. In this product Unsupported means not supported by the transcript. A clinician reading that verdict will go looking for a fabricated lab value and find a correctly transcribed one under an awkward heading.

Example 2

Source: Omi Health medical-note-eval (MIT), dialogue_48, note by Gemini 3 Pro Preview.

Transcript: "S7: [P]: The irregular cycles are more frustrating, the cramps are manageable with ibuprofen."

Note sentence: "Ibuprofen (used by patient for cramps)"

Note Check verdict: Unsupported. Reason: "The transcript mentions the patient uses ibuprofen for cramps but does not indicate it was part of the plan or prescribed."

Our reading: Note Check is wrong. The note says what the transcript says, and the parenthetical is exactly the qualifier that keeps it honest.

Example 3

Source: Omi Health medical-note-eval (MIT), dialogue_70, note by Gemini 3.1 Pro Preview.

Transcript: "S23: Should I keep a log of when the headaches and numbness happen until then?"

Note sentence: "Patient asked if he should keep a log of when the headaches and numbness happen."

Note Check verdict: Unsupported. Reason: "The patient asked about keeping a log, but the clinician did not confirm or instruct the patient to do so."

Our reading: Note Check is wrong. The note reports a question as a question. In the same dialogue, two other notes asserted the log as an instruction the clinician never gave, and there the Unsupported verdict is right. The judge is not separating a note that fabricates a plan from a note that accurately reports a patient's question in a plan-shaped section.

There is a fourth, on dialogue_24, where a note honestly writes "Left knee examination planned but findings not documented in transcript" and the judge marks it Unsupported with a reason that restates the note back at it.

The pattern is one thing: the judge is grading section placement and reporting the result under a faithfulness label. That is worse than a miss. A miss is silence. This is confident and wrong in the direction that costs a clinician time, and it penalises exactly the careful hedging a good note should contain.

The psychiatric set

24 notes, eight dialogues, three writers. 1,393 statements judged: 992 Supported, 7 Unsupported, 0 Contradicted, 0 Unverified, 394 Scaffolding. 14 rule flags. 3 omissions listed.

Six of the seven Unsupported verdicts land on DeepSeek-V4-Pro notes, which is the direction the corpus authors' own leaderboard would predict. Dialogue_169, the most explicitly psychiatric encounter in the set, with a depression scale, a dementia screen and a memory quotient in the transcript, produced nothing at all across three notes.

Example 4, where the check is right

Source: Omi Health medical-note-eval (MIT), dialogue_87, note by DeepSeek-V4-Pro.

Transcript: "S10: [D]: Last time we spoke, we were working on managing your anxiety and seasonal allergies." and "S22: [D]: Have you been taking the medication we discussed for anxiety?" The words "disorder" and "diagnosis" never appear.

Note sentence: "Anxiety disorder, with intermittent exacerbations characterized by palpitations, chest tightness, and insomnia"

Note Check verdict: Unsupported. Reason: "The transcript does not label the anxiety as a disorder or provide a formal diagnosis."

Our reading: Note Check is right. This is the same shape of statement the old engine passed when a note wrote "Type 2 diabetes mellitus" over "your diabetes". In behavioural health the upgrade from a symptom to a named disorder is not a formatting choice. It follows a person into a record that insurers, employers and courts can reach.

The same note also carries "Seasonal allergic rhinitis, with nasal congestion and itchy eyes, partially controlled on current regimen", also Unsupported. We cannot call that one cleanly. "Partially controlled on current regimen" is a judgement nobody in the visit made, and the verdict is right on that clause. "Seasonal allergic rhinitis" is ordinary clinical translation of "seasonal allergies" with congestion and itchy eyes, and the same verdict is harsh on that clause. One verdict covers the whole line, so a reader cannot tell which half is at issue without opening the reason text.

Example 5, where the check is right about a small invented negative

Source: Omi Health medical-note-eval (MIT), dialogue_38, note by GPT-5.5.

Transcript: "S28: [D]: Do you ever get nausea or vomiting with the headaches? S29: [P]: Yeah, sometimes I feel sick to my stomach."

Note sentence: "Gastrointestinal: Positive for nausea; negative for vomiting."

Note Check verdict: flag neg-001, High severity: "'vomiting' is recorded as denied, but the transcript shows the topic was raised and nobody denied it."

Our reading: Note Check is right. The patient affirmed nausea and never answered on vomiting. The writer completed the pair because completed pairs look thorough. This is the exact thing the rule exists for.

Example 6, where the check is wrong on the cleanest possible input

Source: Omi Health medical-note-eval (MIT), dialogue_62, notes by GPT-5.5 and DeepSeek-V4-Pro.

Transcript: "S35: [D]: Any episodes where you got lost or forgot where you were? S36: [P]: No, nothing like that."

Note sentence (GPT-5.5): "Patient denies episodes of getting lost or forgetting location." DeepSeek-V4-Pro wrote "No episodes of getting lost or forgetting location".

Note Check verdict: flag neg-001, High severity: "'forgetting location' is recorded as denied, but the transcript shows the topic was raised and nobody denied it."

Our reading: Note Check is wrong, twice, and this one has no excuse. Question and denial on adjacent lines, correct speakers, plainest possible phrasing. Both writers recorded the visit correctly and the report tells a clinician at High severity that they did not.

Example 7, a rule that reads digits and not words

Source: Omi Health medical-note-eval (MIT), dialogue_54, note by DeepSeek-V4-Pro.

Transcript: "S59: [P]: My dad had an ulcer in his fifties, and my grandmother had colon cancer."

Note sentence: "Father: ulcer in his 50s"

Note Check verdict: flag num-001, Review severity: "'50' appears in the note but no such value appears in the transcript."

Our reading: Note Check is wrong. The transcript spells the number as a word and the rule compares digits. Ages, doses and frequencies get spoken as words constantly, so this is not an edge case.

Example 8, a fact listed as missing while it sits in the note

Source: Omi Health medical-note-eval (MIT), dialogue_100, note by DeepSeek-V4-Pro.

Transcript: "S75: [D]: If your anxiety worsens or you develop chest pain, shortness of breath, or severe headaches, call right away."

Note sentence: "Provide safety netting: call immediately if anxiety worsens, or if chest pain, shortness of breath, or severe headaches develop"

Note Check verdict: listed as an omission, key fact "Provider advised patient to call immediately if anxiety worsens or if chest pain, shortness of breath, or severe headaches develop."

Our reading: Note Check is wrong. The note carries the fact in the Plan section in near-identical wording. Two of the three omissions in this whole run are of this kind. The third, about a patient who eats worse under stress, is genuinely missing and is right.

The part we got wrong ourselves

Fourteen rule flags in this run, thirteen of them wrong on reading. Nine of the thirteen are our fault.

Three of these eight transcripts pack two speakers onto one line, like this: S37: [P]: No. [D]: Any recent head injuries? with the answer on the next line. Our preparation script, unchanged from installment 01 so the two installments stay comparable, maps only a speaker tag at the start of a line. So the doctor's question stayed attributed to the patient, and the rule had a topic with no visible denial.

We could have fixed the script and reported a cleaner number. Instead we kept the original for the headline run and added a control that renormalises the affected pairs properly, then reran them. Nine flags cleared. Three did not: the two on "forgetting location" above, and the "fifties" one. The correct flag on vomiting also survived.

Headline run: 14 flags, 1 right, 13 wrong. Control run: 4 flags, 1 right, 3 wrong. Both numbers are in the report, because publishing only the flattering one is the thing this series exists to not do.

The finding we did not go looking for

The corpus ships something we had not used before: the output of Omi Guard, its authors' own transcript-grounded safety layer, run over the same writers and dialogues. That puts it in the same job as Note Check, built by different people, on the same 24 notes.

We compared each base note with its Guard version, ignoring citations, bullet characters and headings.

	Across the 24 notes
Content lines Guard removed	0
Content lines Guard added	18
Omissions Note Check listed	3
Of those, naming a fact the note already had	2
Overlap between Guard's 18 and Note Check's 3	0

Everything Guard did on this set was recall. Almost everything Note Check did was precision. On what a psychiatric note left out, the two systems agreed on nothing.

One of Guard's 18, from the same MIT corpus, dialogue_54, GPT-5.5 note:

Transcript: "S87: [D]: Are you up to date on your routine screenings, like colon cancer screening? S88: [P]: I haven't had that yet. S89: [D]: Given your family history, we may want to discuss that soon."

Note: nothing. Colon cancer appears only as "Grandmother had colon cancer" under family history.

Note Check verdict: no omissions listed for this note.

Guard added: "Discuss colon cancer screening soon given family history."

That is a screening plan raised because of a family history, and it did not come back to the clinician from us.

Guard is model output, not a clinician's judgement, so none of this proves its 18 are all right. What it says is that two independently built checkers disagree almost completely about what a behavioural-health note dropped, and that the omission side is where to look next.

What happens now

Three defects go into our issue tracker before this publishes, each with the result file that reproduces it:

The judge returns Unsupported for statements it acknowledges are in the transcript, on note-section grounds. Four instances.
neg-001 fires on a denial stated on the line immediately after the question, with correct speaker attribution. Two instances, reproducing on renormalised input.
num-001 compares digits only and does not resolve spelled-out numbers. One instance, reproducing on renormalised input.

Carried from installment 01 and still open in reduced form: the omission list names facts the note already contains. Closed and verified closed here: the negation rule on short affirmative denials, the judge accepting inferred assessments and unsaid plan items, and inconsistent treatment of None placeholder lines.

None of this changes what we say Note Check is. It is a review aid that reports what it found, with the clinician reading the whole note. It is not a safety guarantee and we do not sell it as one. We publish counts and examples, never a rate. There is no accuracy figure, no recall figure and no catch rate for Note Check in this piece, in the report behind it, or anywhere else we write, and there will not be one until a clinician-adjudicated reference set exists and is published.

Every input here is regenerable by anyone. The corpus is public and MIT-licensed, and this run is pinned to commit d828ea9562324f51813f3e5bd1f589637efc187d of Omi Health medical-note-eval. Behind the piece sit 67 result files, the run summary for every check, the base-against-Guard comparison, a machine-readable file of every disagreement with its transcript excerpt and note sentence, and every script. Write to info@krasyn.com and we will send that set.

Send us a pair

Every installment runs reader-submitted pairs in the next one, in full, with the result published whether or not it flatters us. Send a de-identified synthetic transcript and the note written from it.

Synthetic only. We do not accept real patient data, de-identified or otherwise, and we never will.

A 5.8 million-view post gave our reply 6 views

Krasyn EMR — Mon, 31 Aug 2026 05:49:43 +0000

We replied to an X post with 5.8 million views. Our reply received six.

Another reply, under a post with 443,000 views, received 1,119.

That is not a controlled experiment. It is a useful correction.

Small accounts often choose conversations by the source post's visible reach. The larger the number, the larger the imagined distribution. But source reach is not inherited reach. A reply enters a different contest, at a different time, against a different number of participants.

Here are the three public observations that changed how we select conversations.

Source at observation	Our reply	Reply result
5,868,786 views; 1,954 replies	Reply	6 views; 1 like
443,041 views; 540 replies	Reply	1,119 views; 1 like; 1 bookmark
12,240 views; 5 replies	Reply	55 views; no engagement

The source counts are cumulative observations taken after the replies were published. The replies were not posted at the same age, on the same topic, or to the same audience. With only three observations, these numbers cannot isolate a cause.

They can still falsify one tempting assumption: a large source audience does not guarantee meaningful reply distribution.

The denominator that matters is not source views

When a source post has thousands of replies, a new comment can disappear before it has a chance to be evaluated. The visible source-view number says how far the original post traveled. It does not say where a late reply will appear, how many competing replies exist, or whether the conversation is still open.

For a small account, a better pre-reply screen is:

Freshness: Is the conversation still forming?
Saturation: How many replies are already competing for attention?
Fit: Are the likely readers people we want to learn from or help?
Open loop: Does the source leave a real question unresolved?
Contribution: Can we add evidence, a test, or a useful distinction that is not already in the thread?

The first observation failed badly on saturation. The source had enormous reach, but nearly two thousand replies. The 443,000-view source was also crowded, yet our reply reached far more people. That tells us the screen needs more than a single ratio. Topic fit, timing, and the ranking behavior around each conversation remain plausible factors.

A polished reply can still be a dead end

Our replies were written as compact miniature essays. They made complete points. That sounds like a strength, but it also gave readers little reason to respond.

An effective contribution should not manufacture engagement bait. It can, however, create a legitimate next move:

provide a small artifact someone can challenge or reuse
make a falsifiable prediction
ask one narrow question when the answer would change the conclusion
connect the idea to a specific unresolved case
return later with measured evidence

The aim is not to end every comment with a question. The aim is to leave intellectual surface area for another person.

Repeat relationships beat celebrity drive-bys

The most expensive failure was not six views. It was treating each reply as an isolated distribution attempt.

A small account does not build a durable audience by appearing once beneath the largest names it can find. It builds one when the same relevant people repeatedly encounter useful work, recognize the point of view, and eventually choose to continue the relationship.

That changes the operating model:

maintain a focused map of practitioners, researchers, builders, and publications
prefer relevant conversations where a contribution can still be seen
return to the same people when there is genuinely new evidence to add
publish original artifacts that give those conversations somewhere useful to go
measure repeat interactions and substantive replies, not just impressions

Celebrity reach may occasionally produce a spike. Repeated relevance is more likely to produce recognition.

The next test

For the next set of replies, we will record the source age, source reply count, topic fit, whether the post contains an open question, and whether our contribution includes evidence or a reusable artifact.

The outcome measures will be:

substantive replies per ten comments
repeat interactions with the same relevant people
follows per 1,000 reply views
qualified conversations that continue off the thread

The goal is not to prove a universal algorithm from three posts. It is to stop using a seductive but inadequate proxy.

Source reach is evidence about the source. Relationship growth begins with what our contribution earns.

Autonomous AI needs a revocation test, not just a benchmark

Krasyn EMR — Mon, 31 Aug 2026 05:09:28 +0000

The most important test for an autonomous AI system may not be whether it can complete a task. It may be whether people can stop it after it completes the wrong one.

This matters because autonomy changes the shape of failure. A chatbot can produce a bad answer. An agent can turn a bad answer into a sequence of actions: call a tool, update a record, send a message, schedule a follow-up, and trigger another system before anyone notices.

In August 2026, the US Food and Drug Administration opened a public discussion about generative AI-enabled medical devices. Its paper asks how autonomous multi-step action, tool use, and reduced opportunities for human review should affect acceptance criteria and oversight. It also proposes competency testing inspired, at a high level, by how clinicians are evaluated.

That is a useful direction, but the licensing analogy has a missing half.

A professional license does not only certify competence. It also creates a mechanism for investigation, suspension, remediation, and revocation. An autonomous system needs an operational equivalent.

Call it the revocation test.

1. Can you reconstruct one action from source to consequence?

Pick a real or synthetic action the agent took. Ask the team to produce:

the input that initiated it
the model, prompt, policy, and tool versions involved
every tool call and returned value
each human-oversight checkpoint
the records or people affected downstream

An activity log that says only "task completed" is not an audit trail. The useful unit is the entire action chain.

This applies well beyond healthcare. A hiring agent may screen a candidate, update an applicant record, and send a rejection. A finance agent may classify a payment, move it into a queue, and notify a vendor. A logistics agent may reroute an order and create a customer promise. In each case, reconstructing the final output is not enough. You need the path that made the output consequential.

2. Is there a stop threshold written before deployment?

"We monitor it" is not a stop rule.

A stop rule names an observable condition and the action that follows. Examples include:

two unauthorized tool calls in a rolling hour
one action outside the approved scope
a failed human-approval checkpoint before an irreversible step
an unexplained jump in escalation or refusal behavior
a mismatch between the action log and the system of record

The threshold should exist before the incident. If a team invents it afterward, the threshold can quietly move to protect the deployment.

3. Does a named person have authority to hit pause?

A system can have excellent monitoring and still be unsafe if nobody owns the decision to stop it.

Name the role, the backup role, and the maximum time allowed to act. Then test whether that person can actually pause new work, cancel queued work, and prevent downstream systems from continuing the chain.

The owner must have both responsibility and control. Giving someone accountability without a working breaker is theater.

4. Can the team contain actions already in flight?

Stopping new actions is only the beginning. Autonomous systems create partial state.

A useful containment plan answers:

Which queues still contain work?
Which messages have already been sent?
Which external systems accepted a change?
Which actions can be reversed automatically?
Which affected people need a human follow-up?

The more tools an agent can call, the more important this becomes. A kill switch that leaves the last ten minutes of actions propagating is not a complete kill switch.

5. What evidence is required before restoration?

Teams often spend more time defining deployment than defining restoration.

Before an agent is turned back on, require a replay of the failure case, evidence that the stop threshold now behaves correctly, review of adjacent cases, and a written decision from the restoration owner. If the system changed, preserve the old and new versions in the record.

Restoration should be a new decision, not the default result of elapsed time.

Competence gets an agent into production. Revocability determines whether it belongs there.

Benchmarks help establish what a system can do under test conditions. Post-deployment monitoring helps show what it is doing in the world. The revocation test connects those phases to a real control surface.

For any autonomous system, ask five questions before launch:

Can we reconstruct the action chain?
What exact threshold stops it?
Who can trigger that stop?
How do we contain and repair actions already in flight?
What evidence is required before restoration?

If any answer is vague, the system may be autonomous in capability but ownerless in failure.

The FDA discussion paper is open for public feedback through October 19, 2026: https://www.fda.gov/media/194242/download

Krasyn publishes practical tests for high-stakes AI, workflow handoffs, and recovery. The aim is simple: turn broad claims into checks a real team can run.

Seven tests to run before trusting AI inside an EMR

Krasyn EMR — Sun, 30 Aug 2026 16:08:37 +0000

AI features inside an EMR are easy to demonstrate and hard to evaluate. A polished note or confident suggestion can look useful while hiding unsupported details, weak provenance, or a workflow that is difficult to reverse.

Here are seven tests a practice can run before adopting any AI-enabled EMR or scribe, including ours.

1. Test source fidelity, not fluency

Use a synthetic transcript that contains negation, a changed dose, a specific date, and an uncertain statement. Compare the draft with the source line by line.

Record four things separately:

unsupported additions
contradictions
missing facts
uncertainty rewritten as certainty

Do not collapse those categories into one accuracy percentage. Different failures create different review burdens.

2. Locate the human review boundary

Ask exactly when AI output becomes part of the clinical record. The safe answer should identify the review screen, the person responsible for approval, and what remains editable before finalization.

If a vendor cannot show that boundary, the workflow is not ready to evaluate.

3. Force a failure

Interrupt the network, provide low-quality audio, omit a required field, or supply contradictory source material. Observe whether the product fails visibly, preserves the original source, and allows the clinician to recover without losing work.

A graceful demo is less informative than a controlled failure.

4. Ask for provenance

Pick one generated statement and ask where it came from. A reviewer should be able to trace it to a transcript span, entered fact, imported result, or clearly labeled model inference.

Provenance makes review faster and turns disagreement into something testable.

5. Verify privacy and contract boundaries

Before using real patient information, verify the signed agreement, data-use terms, subprocessors, retention policy, deletion path, and whether customer data can train a model. Test with synthetic data until those answers are documented.

6. Measure workflow fit and reversibility

Time the whole task, not just generation. Include capture, correction, coding review, signing, export, and recovery from an error. Then confirm the practice can export its records and continue working if the AI feature is disabled.

The relevant question is not whether AI is fast. It is whether the complete reviewed workflow is better and remains under clinician control.

7. Inspect the evidence behind comparative claims

For claims such as best, most accurate, or time-saving, ask for the study population, sample size, comparison method, failure exclusions, adjudicator qualifications, and limitations. Customer count, paid retention, authentic reviews, and independent evaluation matter because vendor-authored pages cannot establish market leadership by themselves.

We turned this checklist into a more detailed worksheet for evaluating EMR vendor claims: https://krasyn.com/resources/evaluate-emr-vendor-claims?utm_source=devto&utm_medium=article&utm_campaign=authority-evidence-checklist-2026&utm_content=seven-tests

If you want to inspect one AI-drafted note without changing EMRs, Krasyn Note Check can compare a synthetic transcript and note statement by statement: https://krasyn.com/note-check?utm_source=devto&utm_medium=article&utm_campaign=authority-evidence-checklist-2026&utm_content=seven-tests

The standard is straightforward: require a visible test, preserve failures, and keep the final clinical decision with the clinician.

The stopwatch is timing the printer: a three-clock test for AI scribes

Krasyn EMR — Tue, 25 Aug 2026 06:30:24 +0000

At the end of a synthetic visit, start a stopwatch.

When the draft appears, tap lap. Do not stop it.

The first lap measures how quickly the system can produce prose. It does not measure whether the clinician can trust, verify, correct, and sign that prose. For that, an AI scribe pilot needs three clocks:

Visit end to first draft. This is system speed.
First draft to clinician-verified note. This is human review burden.
Verified note to the next time the same correction has to be made. This is correction durability.

If the stopwatch stops when prose appears, it is timing the printer, not the work.

The clock a polished demo can hide

The second clock measures the work after the impressive part. A draft can arrive in seconds and still make a clinician hunt through a transcript, repair the speaker, restore uncertainty, or remove an assessment nobody voiced.

The third clock is quieter. A correction that returns at the next visit creates an annuity of tiny edits. Each one is cheap. The repetition is expensive.

A ten-minute test for any AI scribe

Use a fictional visit and say:

Maybe the breathing exercise helped, but I'm not sure.

Then inspect four things:

Did “maybe” survive?
Can the clinician find the supporting sentence without rereading everything?
How long does verification actually take?
After correcting the draft, does the same failure return in the next synthetic session?

A good result is not merely a fast draft. It is uncertainty preserved, evidence easy to find, and corrections that stay corrected.

AI can return enormous amounts of time. But the value is not generated at the moment output appears. It is generated when a person can finish the work with less effort and no hidden transfer of risk.

That is the black-box test: measure the human outcome, not the machine event.

The diagram uses illustrative test values, not measurements of Krasyn or another product.

Benjamin Krasin, MD is the founder of Krasyn.

A practical first-10 test for outpatient group practices evaluating an AI-native EMR

Krasyn EMR — Sun, 23 Aug 2026 05:13:53 +0000

A practical first-10 test for outpatient group practices evaluating an AI-native EMR

If you run or help operate an outpatient group practice, the useful question is not whether an EMR vendor has the longest feature list.

The useful question is whether the system can fit your first real workflow without creating new administrative drag.

Krasyn is an AI-native outpatient EMR. For the next group-practice conversations, the test I care about is deliberately simple:

Can we map one real outpatient workflow, without patient data, and decide whether there is a practical paid-plan fit?

Who this is for

This is for small outpatient group practices, roughly:

4 to 15 providers,
one or more shared front-desk/admin workflows,
a real need to evaluate notes, scheduling, handoffs, or practice setup,
willingness to discuss workflow fit without sending patient details.

It is not for hospital inpatient workflows, enterprise RFP theater, or anyone looking for unsupported compliance/clinical claims.

What the fit review should cover

A useful first review should answer:

What specialty or outpatient service line do you run?
How many providers and staff need to use the system?
What is the first workflow that must feel trustworthy?
Are you replacing an existing EMR or evaluating a new setup?
Would AI-assisted note review, scheduling, admin setup, or workflow handoff be the first value point?
If the fit is real, what paid-plan shape would make sense?

No patient names, messages, dates of birth, diagnoses, chart screenshots, or clinical details are needed.

Why this is a better test than a generic demo

Generic demos are easy to make impressive and hard to learn from.

A real group-practice fit review is harder, but more useful. It forces the conversation toward:

provider count,
team workflow,
onboarding friction,
buyer urgency,
plan fit,
measurable next step.

That is the path I would rather measure.

What counts as progress

I will not call this growth because an article exists.

Progress means one of these happens:

a qualified group-practice fit review starts,
a legitimate inquiry comes in without sensitive patient content,
a trial or demo is booked,
a paid plan starts,
activation and retention can be measured after signup.

Anything else is just distribution activity.

Start here

If you operate an outpatient group practice and want to test fit without sending patient data:

Request a Krasyn group-practice fit review

If you want to see the public product first:

Visit Krasyn

The goal is not to win a marketing argument. The goal is to find out whether the workflow fit is real.

The outpatient EMR revenue path I would measure before claiming growth

Krasyn EMR — Sun, 23 Aug 2026 04:20:22 +0000

The outpatient EMR revenue path I would measure before claiming growth

Krasyn is an AI-native outpatient EMR. I am not going to claim growth from a new page, a pitch deck, or a better valuation story.

The cleaner test is whether real outpatient group practices move through a measurable buyer path:

a group-practice fit review starts,
the practice has a real workflow and team size,
the plan commitment is captured without patient data,
activation and retention can be measured after signup,
revenue events can be tied back to the source that produced them.

That is the path we are measuring now.

Why outpatient group practices?

A solo user can validate usability, but a group practice validates operational value: multiple providers, handoffs, scheduling, notes, chart review, billing-adjacent workflows, and admin oversight.

For a small outpatient group, the buying question is not "does this have every enterprise feature?" It is:

can the team understand the system quickly?
can clinicians review AI-assisted notes safely?
can the practice see where time is saved or risk is reduced?
can the subscription be justified by real team usage?

That is a better signal than generic traffic.

What we are not claiming

We are not claiming clinical validation, regulatory certification, or a medical outcome improvement from this post.

We are also not claiming a valuation increase from marketing copy. A valuation claim needs real evidence: qualified interest, paid conversion, activation, retention, and revenue.

What a useful fit review should answer

For an outpatient group practice, a useful first conversation should answer:

specialty and practice size,
provider count,
whether the team is replacing an existing EMR or starting fresh,
the first workflow they would need to trust,
whether AI note review, scheduling, or admin workflow is the first pain point,
whether there is a realistic paid plan fit.

No patient details are needed for that.

The current measurable path

If you run or help operate an outpatient group practice and want to evaluate this path, start here:

Request a Krasyn group-practice fit review

If you just want to see the public product surface first:

Visit Krasyn

The bar for calling it growth

Publishing this post is not growth by itself.

The bar is objective movement in the funnel:

fit-review starts,
legitimate inquiries,
trials or demos,
purchase events,
activation evidence,
retention evidence.

If those do not move, the honest answer is that the content did not produce growth.

That is the point of measuring it.

We ran our own note checker against a public scribe benchmark. Here is what it missed.

Krasyn EMR — Fri, 21 Aug 2026 21:44:49 +0000

Krasyn ships a tool called Note Check. Paste a visit transcript and the note any AI scribe drafted from it, and it labels each sentence Supported, Unsupported, Contradicted, Scaffolding or Unverified against the transcript, raises three pure-code flags (a number the transcript never contained, a denial about a topic raised and never denied, content filled over an inaudible marker) and lists the facts the note left out. It never edits the note. The clinician reads the whole note and signs it.

We publish no accuracy figure for it, because it has not been measured against a clinician-adjudicated reference set. What we can do is run it on open data that someone else has already labelled, publish the counts and the disagreements, and say which side looks right on reading. This is the first installment.

The data

Omi Health publishes medical-note-eval under the MIT license: 300 synthetic primary-care dialogues, SOAP notes written for each by named frontier models, and labels from a panel of LLM judges. The 2025 benchmark ships per-note counts of unsupported claims from three cross-family judges; the 2026 benchmark publishes per-writer totals only.

We picked six dialogues with different presenting problems and took the notes from three 2025 writers (GPT-5, Claude Opus 4.5, Gemini 3 Pro Preview) and three 2026 writers (GPT-5.5, Claude Sonnet 4.6, Gemini 3.1 Pro Preview): 36 pairs. Sentence IDs were removed from transcripts and citation tags from notes; nothing else changed. Every pair went through the production Note Check service (engine notecheck-1, judge gpt-4o at temperature 0). Eight were also run in the live product with screenshots and a screen recording.

The counts

Set	Pairs	Statements	Supported	Unsupported	Contradicted	Scaffolding	Rule flags	Omissions
2025 notes (per-note labels exist)	18	627	354	3	0	270	6	55
2026 notes (no per-note labels)	18	774	422	0	0	352	10	30

For the 18 labelled notes, Omi's panel had counted a major unsupported claim (two or more judges agreeing) in 10. Note Check's judge labelled a statement Unsupported in one of those 10 and in none of the other nine. On the eight notes the panel had not flagged, Note Check also found nothing. The two tools agree on clean notes and disagree on almost every note the panel thought had a problem.

Dialogue	Model	Omi majors per judge	Note Check Unsupported + Contradicted	Note Check flags	Reading
dialogue_0	GPT-5	3, 3, 4	0	1	Plan adds counselling nobody said. Flag is a rule misfire.
dialogue_0	Claude Opus 4.5	2, 2, 2	0	0	"Order laboratory studies" firms up "may need some labs".
dialogue_2	GPT-5	1, 1, 1	0	2	Assessment in a visit with none. One flag right, one wrong.
dialogue_2	Claude Opus 4.5	1, 1, 1	0	0	Same invented assessment.
dialogue_24	GPT-5	1, 3, 1	3	0	Hit three "None" placeholders, not the type label.
dialogue_24	Claude Opus 4.5	1, 1, 0	0	0	"Type 2" is never stated.
dialogue_48	Gemini 3 Pro Preview	1, 1, 0	0	3	All three flags wrong; the patient did deny.
dialogue_57	GPT-5	2, 3, 4	0	0	The word diabetes never appears.
dialogue_57	Claude Opus 4.5	3, 3, 1	0	0	"Continue metformin" when the plan was to adjust.
dialogue_70	GPT-5	1, 1, 0	0	0	A differential and a symptom log nobody proposed.

The eight rows where both sides found nothing, and the 18 notes from the 2026 writers, are in the full report.

The surprising one

In dialogue_48 the doctor asks the patient to confirm there is no history of blood clots, migraines with aura or uncontrolled blood pressure. The patient says, "Correct, none of those." Four models (Gemini 3 Pro Preview, GPT-5.5, Claude Sonnet 4.6, Gemini 3.1 Pro Preview) each wrote some form of "Denies history of blood clots, migraines with aura, or uncontrolled blood pressure." That is the right sentence.

Our pertinent-negative rule flagged all three items in all four notes as High: the topic was raised and, it claimed, nobody denied it. The rule found the question and did not recognise the answer as a denial. It is pure code, and it behaved identically every time. Twelve of the sixteen flags in this run are that one mistake. A thirteenth read "without sharp pain and intermittent foot swelling" as a denial of the swelling. Two flags were right: "weight change" recorded as negative in two notes when weight was asked about and never answered.

The quiet one

The larger miss is quieter. In dialogue_2 the visit is history only; it ends with the doctor saying the timeline is now clearer. Both the GPT-5 and Claude Opus 4.5 notes carry an assessment: cough "likely related to allergic etiology". All three of Omi's judges counted that as a major unsupported claim. Note Check's judge called it Supported. The same pattern covers smoking cessation counselling in dialogue_0, "Type 2 diabetes mellitus" where the doctor said "your diabetes" or nothing at all, and a differential of migraine versus transient ischemic attack in dialogue_70 that nobody voiced. Our judge prompt allows faithful clinical translation to count as Supported. On this corpus it let clinical conclusions through under that heading. Each is sensible, which is exactly why a signer would not catch them either.

What held

The fingerprint for a pair was the same in the command-line run and in the live product. Fresh judge calls were not identical: verdicts matched on seven of eight pairs run both ways, and the omission list changed on four of eight. This is why the product stores a report under its fingerprint instead of recomputing it, and why the omission list is a prompt to look rather than a count.

Limitations

Synthetic dialogues, short and clean. LLM judges on both sides, no clinician adjudication on either; the reading column is ours. Omi's labels are counts without statement text, so overlap is inferred. The notes come from general-purpose models under a benchmark prompt, not from any commercial scribe product. Six dialogues is a small set; we report counts and no rate.

The rule defect and the judge leniency are tracked. The next installment reruns this exact set on the next engine version. If you have a de-identified synthetic transcript and note pair, send it and we will run it and publish the result.

Dialogues and notes: Omi Health medical-note-eval, MIT license.

Note Check is free inside a Krasyn Scribe account: https://krasyn.com/note-check?utm_source=devto&utm_medium=longform&utm_campaign=scribe-redteam-2026. The faithfulness benchmark for Krasyn's own scribe is at https://krasyn.com/resources/ai-scribe-faithfulness-benchmark.

We published how we measure our AI scribe's faithfulness, and built a checker anyone can run on any scribe's note

Krasyn EMR — Fri, 21 Aug 2026 18:26:15 +0000

I founded Krasyn, an outpatient EMR with an AI scribe inside it. Krasyn has run a working outpatient clinic's real patient records since March 2026, so what our scribe drafts ends up in charts that real clinicians sign. This post covers two things we shipped in August: a published benchmark of how faithful those drafts are to the transcript, and Note Check, a tool that reads any scribe's note against its transcript and lists what the transcript does not support.

Why a fluent note is the problem

A faithful note and a note with one invented blood pressure look the same on the screen, and the clinician who signs it owns every sentence. Published evaluations put ambient-scribe hallucination at about 1 to 3 percent of notes. A March 2026 analysis of 71,173 AI-drafted and finalized note sections found a confirmed edit in 5.8 percent of them. The drafting got automated. The checking did not.

I wanted a number for our own scribe that I could defend, with the definitions printed next to it. A benchmark without definitions is marketing.

The unit: a clinical assertion

We measure at the level of a clinical assertion, one atomic statement about the patient that could be true or false on its own. "Denies fever, chills, and nausea" is three assertions. A measurement and its value are one. Hedging is kept verbatim.

Every assertion gets exactly one label against the transcript:

Supported: the transcript says it, or it is a faithful paraphrase or clinical translation.
Inferred: not stated, but a reasonable clinical inference with a basis in the transcript. Tracked separately because it is the contested category.
Unsupported: no basis in the transcript at all.
Contradicted: the transcript says the opposite, including a symptom the patient denied, a treatment the clinician declined, or another person's symptom attributed to the patient.

Hallucination rate is unsupported plus contradicted over all assertions. Coverage is measured separately against key facts per case, because a note that says only "Patient was seen" scores perfect faithfulness. A faithfulness gain bought by dropping content is a regression.

Traps

Each case also defines traps, specific fabrications the note must not contain, each a regex plus a written rationale. Traps need no model, so they cannot drift when a model changes. A trap fires only if a mention survives negation and irrealis suppression scoped to the containing sentence, and every suppressed mention is written to the artifact with the rule that suppressed it.

The corpus

Twelve synthetic transcripts, 144 key facts, 61 traps. All original invented dialogue, no real or de-identified patient data. Seven of the twelve are adversarial: a spouse describing her own symptoms, eleven explicit denials in one visit, pervasive hedging, audio dropouts, treatments discussed and then declined. In those cases the faithful answer is to say less. The corpus was committed to git before the harness existed, because this repo has a documented habit of expectations written to match current output.

The two arms

One generation feeds both arms. The ungrounded arm is the draft as generated. The grounded arm is the same draft after our production grounding verifier, which strips sentences it cannot trace to the transcript. There is no competitor arm. We cannot test other scribes fairly with no API access and no matched corpus, so we publish no number for them.

The 2026-08-09 run

Metric	Ungrounded	Grounded
Assertions extracted	296	279
Hallucination rate	1.7%	0.0%
Contradiction rate	1.0%	0.0%
Faithfulness precision	94.6%	97.8%
Key-fact coverage, strict	88.2%	86.8%
Traps fired	6 of 61	5 of 61

Negative control: every run injects three unambiguous fabrications into every grounded note and re-scores. 36 of 36 were caught. If that figure were ever below 100 percent the report would say at the top that every other number is suspect.

The coverage cost is real. Grounding removed 12 sentences and strict coverage fell 1.4 points. All of that loss sits in one case, FB-008, where the judge stripped a whole sentence that bundled a fabricated denial with a true hedged finding. Whole-sentence removal is the residual defect.

Five traps still fire in the grounded arm. Two are in the crosstalk case, where the note correctly recorded that the spouse described her own headaches and the regex fired anyway. We leave that unsuppressed on purpose. A rule good enough to silence it would also hide a real wrong-patient attribution.

The limitations, stated plainly

The labels come from gpt-4o, and gpt-4o also wrote the notes. Self-preference bias in model judges is documented, and it is uncontrolled here. No clinician has adjudicated a single label. The harness author, an AI agent, also wrote the corpus, the traps, and the judge prompts. We measure agreement each run on 163 assertions with the list held fixed: the same judge re-run agreed 100 percent, and gpt-4o-mini agreed 96.3 percent on the exact label. Agreement between two models is not agreement with a clinician.

Twelve cases is small, and per-case differences are anecdotes. The transcripts are written English, not ASR output. The trap suppression rules were tuned after a first run produced 13 false positives, and a self-test now pins every rule against an assertive counterpart.

Neither layer reliably scores a fabricated denial. An invented "denies fever" reads as grounded to the judge when the transcript merely mentions fever. That defect was found by reading notes, not by a metric moving. A number that cannot go up when a defect occurs is not evidence the defect is absent.

Note Check

The benchmark measures our scribe. It does not help someone who uses a different one. So we built Note Check. You paste a transcript and the note any scribe drafted from it, and you get a report on that note. It works on pasted text from Krasyn, Freed, Heidi, Nabla, Abridge, DAX, Suki, Upheal, Mentalyc, or anything that produces a note as text. It does not connect to another vendor's system.

What it does. Each sentence of the note gets one of five labels against the transcript: Supported, Unsupported, Contradicted, Scaffolding, or Unverified. Three flags come from plain code with no model involved: a number in the note that never occurs in the transcript, a pertinent negative about a topic the transcript raised but never denied, and specific content placed where the transcript only has an inaudible marker. An omissions list names clinically material facts the transcript stated and the note left out, each with its excerpt. Contradicted and unsupported lines are lifted to the top, and the report exports as plain text.

What it does not do. It does not edit the note, file it, or sign it. An exam finding the clinician observed and never spoke reports as Unsupported, which means "not in the transcript" and nothing more. It compares against the transcript you paste, not the chart. And we publish no accuracy figure for Note Check, because it has not been measured against a clinician-adjudicated reference set. It is a review aid. Read the whole note.

The honest determinism note

The pure-code layer always runs and its flags are identical on every run. The model judge runs at temperature 0 with a pinned prompt version, and it is not seeded, because the pinned SDK exposes no seed. On 2026-08-21 we ran the same flawed synthetic note twice on the bench and once in production. All five planted contradictions were caught every time and the two code flags were identical. The reason prose differed on all five contradicted statements, and one borderline statement, "Hypertension, controlled", moved between Supported and Unsupported between environments.

So we do not call the judge deterministic. Instead the finished report is stored under a fingerprint of the transcript, the note, the engine version, and the judge model. Within your account the same inputs return the stored report, with no second model call. Every report states whether the judge ran, ran on only part of the transcript, failed, or did not run, and any statement it did not assess is labelled Unverified rather than counted as Supported.

Where to find it

The full method with every definition is at https://krasyn.com/resources/ai-scribe-faithfulness-benchmark. The source document is docs/ai/FAITHFULNESS_BENCHMARK.md in our repository, which is private today. Note Check lives inside the free Krasyn Scribe account, which includes 10 AI-drafted notes a month with no card, and each check uses one of those notes. A synthetic worked example is at https://krasyn.com/note-check?utm_source=devto&utm_medium=article&utm_campaign=note-check-landing.

If you run a scribe evaluation of your own and your definitions differ from ours, I would like to read them.

Reducing Documentation Burden: AI Tools That Actually Work for Physicians

Krasyn EMR — Sat, 15 Aug 2026 05:55:13 +0000

Originally published at krasyn.com/blog/reducing-documentation-burden-ai-tools

Documentation burden is the single largest driver of physician burnout. This guide evaluates the AI tools that have demonstrated real-world documentation time reduction -- and separates them from the ones that add complexity instead of removing it.

The Documentation Burden Problem by the Numbers

The AMA's Physician Work Life Study reports that for every hour spent in direct patient care, physicians spend nearly 2 hours on EHR documentation and desk work. The problem compounds: a 2023 JAMA Internal Medicine analysis found that primary care physicians open their EHR after hours an average of 1.4 times per day, spending an additional 45-60 minutes on documentation outside clinic hours (the "pajama time" problem).

Documentation burden correlates directly with burnout. The 2023 Medscape Physician Burnout and Depression report found that 64% of burned-out physicians cited "too many bureaucratic tasks including charting and paperwork" as their primary burnout driver -- more than any other factor including long hours, lack of autonomy, or inadequate compensation.

Category 1: Ambient AI Scribing

Ambient AI scribes listen to clinical conversations during patient encounters and produce structured clinical notes -- no dictation required, no template filling, no manual data entry. The physician reviews the draft note, edits where needed, and signs.

How it actually works: A microphone captures the conversation. The audio is processed through a speech recognition and clinical language model pipeline that identifies the chief complaint, history elements, physical exam findings, assessment, and plan. The output is a SOAP or APSO note in the physician's preferred format.

Published Evidence on Ambient AI Scribing

Epic/Nuance DAX Copilot (JAMA Network Open, 2023): 2,400 encounter study; physicians spent average 1.8 minutes reviewing AI notes vs 4.2 minutes traditional documentation. 85% of physicians rated AI note quality as "good" or "very good." After-hours documentation time decreased by 36%.
Suki AI (NEJM Catalyst, 2024): 280-physician multi-site study; documentation time per note fell by an average of 72 seconds.
Abridge (NEJM Catalyst, 2024): 150-physician study; physician-rated accuracy: 91% acceptable without significant edits.

What does not work: AI scribes that require physicians to speak differently than they do naturally add cognitive load instead of reducing it. The systems with the best outcomes work passively.

Category 2: AI-Assisted Note Generation (Template-Based)

A step below ambient scribing, AI-assisted note generation helps physicians build notes more efficiently through intelligent templates, auto-population of stable elements, and natural language processing of dictated or typed text.

When it helps: Practices where ambient scribing is difficult (high background noise, shared exam rooms) or not yet implemented.

When it does not help: If the template still requires significant manual input, the cognitive burden is shifted, not reduced.

Category 3: AI Billing Review

AI billing review analyzes the completed clinical note and suggests appropriate CPT and ICD-10 codes, flags potential undercoding or missing codes, and identifies documentation gaps.

Real-world impact: A 2024 University of Michigan study found that practices using AI billing review captured 8-12% additional revenue per physician per year -- primarily by identifying legitimate 99214 visits being billed as 99213, and by flagging HCC-relevant diagnoses present in the note but not included in the claim.

Key capability to look for: The AI should show its work -- displaying the specific note content that supports each suggested code.

Category 4: AI Order Entry and Clinical Decision Support

AI-powered order entry predicts the orders likely to be needed based on the encounter type and chief complaint, reducing the number of clicks to complete an order set.

The alert fatigue problem: Studies consistently show that physicians override 90%+ of CDS alerts in legacy systems. AI-powered CDS addresses this by learning which alerts a given physician acts on and suppressing those they reliably override.

Category 5: Inbox and Message AI

Patient messages through portal systems generate significant after-hours work. AI that triages messages and drafts responses to common questions has demonstrated significant inbox time savings.

The key clinical safeguard: the physician reviews and signs every response generated by AI. AI does not communicate with patients autonomously.

Choosing the Right Tool: Questions to Ask Any Vendor

What is the average physician-reported time reduction in peer-reviewed studies (not marketing materials)?
How does the system handle encounters with significant background noise or non-verbal patients?
What is the BAA status? Who processes the audio? Where is it stored and for how long?
How is the system trained on my specialty's vocabulary?
What is the physician review workflow? How long does it actually take to review and sign a typical AI-drafted note?

The Implementation Reality

Most physicians report a 2-4 week adaptation period when adopting ambient AI scribing. During this period, note quality is lower and review time is higher as the system calibrates to the physician's style. After adaptation, most physicians find the workflow genuinely faster. Planning for and communicating this adaptation period prevents premature abandonment of tools that take time to realize their value.

Originally published at krasyn.com/blog/reducing-documentation-burden-ai-tools

The Real Cost of a Bad EMR: Time, Burnout, and Revenue Leakage

Krasyn EMR — Sat, 15 Aug 2026 05:55:10 +0000

Originally published at krasyn.com/blog/real-cost-bad-emr-burnout-revenue

Most practices underestimate what a poorly designed EMR costs them annually. This analysis breaks down the time cost, burnout impact, and revenue leakage attributable to EMR workflow friction.

The Hidden Costs That Never Appear on an Invoice

The visible cost of an EMR is the software subscription: $300-$800 per provider per month. But that number captures only a fraction of what a poorly designed EMR actually costs. The real costs -- documentation time loss, billing error rates, staff turnover driven by system frustration, and the clinical outcomes of delayed care decisions -- are substantial and almost never calculated explicitly.

This analysis quantifies the full cost picture for a hypothetical solo primary care physician using a legacy fee-for-service EMR vs. an AI-native platform.

Cost 1: Documentation Time Loss

Legacy EMRs average 1.8-2.4 hours of physician documentation time per day for a physician seeing 18-22 patients. That includes note writing, order entry, message responses, prior authorization documentation, and lab result review and management.

Physicians in AI-native EMRs with ambient documentation average 0.8-1.2 hours of documentation per day for equivalent patient volumes. The difference: 45-90 minutes daily.

Metric	Legacy EMR	AI-Native EMR	Difference
Documentation time/day	2.1 hours	1.0 hour	-1.1 hours/day
Annual documentation hours (220 working days)	462 hours	220 hours	242 hours saved
Value of physician time (@$350/hr opportunity cost)	$161,700	$77,000	$84,700 saved
After-hours documentation ("pajama time")	45 min/day average	10 min/day average	-35 min/day

At a physician opportunity cost of $350/hour, 242 hours annually represents $84,700 in recoverable time value.

Cost 2: Revenue Leakage from Coding Errors

Legacy EMRs with manual coding produce systematic coding patterns that reduce revenue without anyone necessarily making an error:

Habitual under-coding: A physician who codes 99213 instead of 99214 for 30% of visits loses approximately $44/visit x 6 visits/day x 220 days = $58,000 annually.
Missed HCC capture: Diabetic patients coded without appropriate specificity lose HCC credit in Medicare Advantage arrangements.
Missing add-on codes: G2211 (Medicare complexity add-on), 99429 (advance care planning), G0444 (depression screening) -- codes that can legitimately be added but require documentation awareness.
Claim denial rates: Practices with poor coding practices see claim denial rates of 8-12%; AI-assisted billing review practices typically run 2-4%.

Cost 3: Staff Turnover Driven by EMR Frustration

Front office and medical assistant turnover is heavily correlated with EMR usability. A 2023 KLAS Research report found that practices with EMRs ranked in the bottom quartile of usability had MA/front-office turnover rates notably higher than practices with top-quartile EMRs.

Replacing an experienced medical assistant costs $8,000-$15,000 when accounting for recruiting, training time, and productivity loss during the transition.

Cost 4: The Burnout Penalty

Physician burnout driven by EMR friction has direct financial consequences:

Reduced clinical hours: Burned-out physicians reduce their clinical hours by an average of 4 hours/week (AMA survey). At 20 patients/day, reducing to 3.5 days/week loses 520 patient visits annually -- at $150 average revenue per visit, that is $78,000 in lost annual revenue.
Early retirement: Burnout is the primary non-financial driver of early retirement among physicians.
Quality effects: A 2022 Mayo Clinic study found that burned-out physicians had 2.2x higher rates of self-reported medical errors.

Building the Full Cost Picture

Cost Category	Legacy EMR	AI-Native EMR	Annual Savings
Software subscription	$7,200	$2,988	$4,212
Documentation time cost	$161,700	$77,000	$84,700
Revenue leakage (coding + denials)	$75,000	$30,000	$45,000
Staff turnover (EMR-attributable)	$14,400	$7,200	$7,200
Total annual impact	$258,300	$117,188	$141,112

The directional finding is consistent: the cost of a bad EMR is not primarily its subscription price.

Calculating Your Own Number

To calculate the real cost of your current EMR:

Track documentation time for one week (use a time diary or ask your EMR for login duration data)
Run a coding audit: compare your E/M code distribution to specialty benchmarks (CMS publishes these by specialty)
Pull your claim denial rate from your billing system for the past 12 months
Ask your last two departed MAs/front office staff why they left

Originally published at krasyn.com/blog/real-cost-bad-emr-burnout-revenue

How to Start a DPC Practice: A 12-Month Roadmap

Krasyn EMR — Sat, 15 Aug 2026 05:55:08 +0000

Originally published at krasyn.com/blog/how-to-start-dpc-practice-12-month-roadmap

Direct Primary Care has grown from approximately 300 practices in 2010 to over 2,500 in 2026. More than 30 states have enacted DPC-enabling legislation that clarifies the legal status of DPC memberships (not insurance contracts). Employer interest in DPC as a benefit has expanded significantly since 2022 as small and mid-size businesses seek alternatives to rising group health insurance premiums.

The financial case is also clearer than it was five years ago. The DPC model is well-documented: a solo physician with 600 patients at $85/month generates $612,000 in gross membership revenue with lower overhead than a fee-for-service practice. For physicians who feel trapped in a volume-driven, insurance-dependent system, DPC offers a genuine alternative.

This roadmap assumes you are starting from employment or an existing fee-for-service practice. Adjust timelines as needed for your situation.

Months 1-3: Foundation and Legal Structure

Step 1: Verify Your State's DPC Landscape

DPC-enabling legislation exists in 30+ states, but the specifics vary. Some states explicitly exempt DPC membership agreements from insurance regulation; others have no specific DPC statute and require navigating insurance department guidance. Start by:

Reviewing your state's DPC statute (search your state legislature website for "direct primary care")
Consulting a healthcare attorney familiar with your state's DPC regulations -- this is a $500-$1,500 investment that can prevent a $50,000+ mistake
Contacting your state's insurance department if no statute exists -- get a written opinion on whether your proposed membership structure is subject to insurance regulation

Step 2: Choose Your Business Structure

Most DPC practices organize as a Professional Limited Liability Company (PLLC) in states that allow it, or a Professional Corporation (PC). The PLLC structure provides liability protection while allowing pass-through taxation. A solo DPC physician PLLC with no employees can elect S-Corp taxation if profitable enough -- a conversation worth having with your CPA when annual revenue exceeds $150,000.

Step 3: Review Your Employment Contract Non-Competes and IP Assignment

If you are leaving an employed position, review your contract for:

Non-compete geographic restrictions and duration
Patient solicitation restrictions (often different from non-compete terms)
Chart and patient record ownership provisions
Tail coverage obligations for malpractice

Non-competes in physician contracts are enforceable in most states but increasingly restricted -- verify your state's current law. Budget 90-180 days from resignation to practice opening to honor typical non-compete notice periods.

Step 4: Draft Your Membership Agreement

Your DPC membership agreement is the foundation of your business. It must clearly specify:

Included services (define explicitly -- list all services covered by the membership fee)
Excluded services (labs, imaging, specialist care, medications not provided in-office)
Membership fee, billing cycle, and payment method
Termination terms (by physician and by patient)
That the agreement is not health insurance
HIPAA authorization and privacy practices

Have a healthcare attorney draft or review this document. Do not use a template found online without legal review for your specific state.

Months 4-6: Physical Space, Technology, and Operations

Step 5: Select Your Location

DPC practices require significantly less administrative space than fee-for-service practices. A solo DPC practice can operate from 600-1,200 square feet with 1-2 exam rooms and a front desk. This dramatically reduces overhead compared to traditional practice.

Monthly rent considerations:

Target rent at 8-12% of projected monthly revenue
Look for medical office suites with shared reception
Some DPC physicians start in home office setups with telemedicine-first delivery to minimize startup costs

Step 6: Select Your EMR and Membership Billing Platform

Key EMR requirements for DPC:

Built-in recurring membership billing (ACH, credit card)
Membership tier management and panel capacity tracking
AI ambient documentation (saves 45-90 min/day you cannot afford to lose with a small panel)
Telehealth integrated (DPC patients expect portal and telehealth access)
Clean patient portal for messaging, scheduling, and document sharing

Step 7: Credentialing and Licensing

Verify your medical license is in good standing and renew if within 12 months of expiry
DEA registration if you plan to prescribe controlled substances
CLIA waiver if you will perform in-office laboratory testing
If opting out of Medicare: file opt-out affidavit with your MAC
Malpractice insurance: obtain occurrence-based or claims-made + tail coverage

Months 7-9: Patient Acquisition and Marketing

Step 8: Build Your Initial Patient List

Most DPC physicians open with 100-200 founding members and grow to 400-600 over 12-18 months. Founding members typically receive a discounted rate ($10-$20/month off) as recognition for their early commitment.

Sources for initial patients:

Your existing patient panel (subject to non-solicit provisions -- get legal clearance first)
Employer groups: approach 5-10 small businesses about a group DPC arrangement
Community outreach: local presentations, farmer's market presence, chamber of commerce membership
DPC specialty directories: DPC Alliance's provider directory, directprimarycare.com

Step 9: Price Your Membership and Build the Pro Forma

Use this formula to determine the minimum viable membership fee:

Target annual physician income: $350,000
Annual overhead (rent, staff, supplies, malpractice, EMR, marketing): $120,000
Total needed from practice: $470,000
Target panel: 600 patients
Required monthly fee: $470,000 / 600 / 12 = $65.28/month minimum

Price above minimum to account for attrition and ramp-up: $75-$85/month for most markets.

Months 10-12: Launch and Stabilization

Step 10: Soft Launch Before Full Capacity

Open to founding members at reduced capacity (100-150 patients). This controlled launch lets you refine your workflows and identify scheduling issues before scaling.

Step 11: Financial Runway Planning

The ramp to break-even for a DPC practice is typically 6-12 months. Minimum recommendation: 6 months of personal and practice expenses in liquid reserves before opening.

Step 12: Build Your Wholesale Lab and Imaging Network

Negotiate direct pricing with Quest or LabCorp. A comprehensive metabolic panel at wholesale may cost $8-$12 vs. $150+ through insurance. Patients see this as a major benefit.

Timeline Summary

Month	Key Milestones
1-3	Legal structure, state registration, employment contract exit, membership agreement
4-6	Space selection, EMR selection, credentialing, malpractice, CLIA waiver
7-9	Founding member recruitment, employer outreach, pricing finalization, marketing
10-12	Soft launch, billing cycle testing, lab network, referral relationships

Originally published at krasyn.com/blog/how-to-start-dpc-practice-12-month-roadmap

DEV Community: Krasyn EMR

The check got better and got a new way of being wrong

What this series is

The therapy source we did not run

First, the fixed set

Then, the new way of being wrong

Example 1

Example 2

Example 3

The psychiatric set

Example 4, where the check is right

Example 5, where the check is right about a small invented negative

Example 6, where the check is wrong on the cleanest possible input

Example 7, a rule that reads digits and not words

Example 8, a fact listed as missing while it sits in the note

The part we got wrong ourselves

The finding we did not go looking for

What happens now

Send us a pair

A 5.8 million-view post gave our reply 6 views

The denominator that matters is not source views

A polished reply can still be a dead end

Repeat relationships beat celebrity drive-bys

The next test

Autonomous AI needs a revocation test, not just a benchmark

1. Can you reconstruct one action from source to consequence?

2. Is there a stop threshold written before deployment?

3. Does a named person have authority to hit pause?

4. Can the team contain actions already in flight?

5. What evidence is required before restoration?

Competence gets an agent into production. Revocability determines whether it belongs there.

Seven tests to run before trusting AI inside an EMR

1. Test source fidelity, not fluency

2. Locate the human review boundary

3. Force a failure

4. Ask for provenance

5. Verify privacy and contract boundaries

6. Measure workflow fit and reversibility

7. Inspect the evidence behind comparative claims

The stopwatch is timing the printer: a three-clock test for AI scribes

The clock a polished demo can hide

A ten-minute test for any AI scribe

A practical first-10 test for outpatient group practices evaluating an AI-native EMR

A practical first-10 test for outpatient group practices evaluating an AI-native EMR

Who this is for

What the fit review should cover

Why this is a better test than a generic demo

What counts as progress

Start here

The outpatient EMR revenue path I would measure before claiming growth

The outpatient EMR revenue path I would measure before claiming growth

Why outpatient group practices?

What we are not claiming

What a useful fit review should answer

The current measurable path

The bar for calling it growth

We ran our own note checker against a public scribe benchmark. Here is what it missed.

The data

The counts

The surprising one

The quiet one

What held

Limitations

Next

We published how we measure our AI scribe's faithfulness, and built a checker anyone can run on any scribe's note

Why a fluent note is the problem

The unit: a clinical assertion

Traps

The corpus

The two arms

The 2026-08-09 run

The limitations, stated plainly

Note Check

The honest determinism note

Where to find it

Reducing Documentation Burden: AI Tools That Actually Work for Physicians

The Documentation Burden Problem by the Numbers

Category 1: Ambient AI Scribing

Published Evidence on Ambient AI Scribing

Category 2: AI-Assisted Note Generation (Template-Based)