Card 02 of 13 · before the visit
copilot, directed verification

Chart summarization

Forty pages of chart, four minutes before the visit. The trade is real, and so is the risk: a fabrication can be checked against the source, but an omission only shows up if you reread the source, which is the exact work the summary existed to save.

Guiding questionWhat did the summary quietly leave out?

You have four minutes before a follow-up visit and forty pages of chart. You paste the history, read a clean paragraph, and walk in feeling prepared. Whatever the summary dropped, you will not miss it, because you never saw it.

Compressing the recordDelegable with review
Flagging what changedCopilot
Deciding what matters todayYours
Putting it in the chartNot delegable
01 · What it does, and what it does not
What it does well

Turns forty pages into a map, surfaces things you would not have thought to look for, and gets you into the room with the right question already in your head.

What it does not do

Audit the chart. A summary inherits the errors already in the record and hands them back with fresh authority.

Clinical design position

The risk here is asymmetric, and that asymmetry drives the whole card. A fabrication can be checked against the source. An omission cannot, because checking it means rereading everything. So this card does not ask you to reread. It asks you to check the short list of things that hurt when they go missing, and to make the model declare its own gaps out loud.

02 · Where it fails here
Plain omission

The allergy disappears. The most dangerous failure on this card, because it is invisible by nature: you do not miss what you never saw.

Flattened trajectory

Summaries compress time. Every value can be correct while the direction vanishes: a creatinine climbing over three months becomes "renal function stable." Right item by item, wrong story.

Silent contradiction

Real charts contradict themselves. The model is built to produce coherent text, so it picks a version and never mentions there were two. Coherence is a virtue in writing and a defect in an audit.

These three link to the site's failure taxonomy, where the transcripts live.

03 · The safer workflow
01Prepare the input.Text stripped of identifiers, clinical dates preserved. No dates, no trajectory, and trajectory is half the value of a summary.
02Fixed agenda, aimed at today.Active problems, medications with dose and frequency, allergies, recent labs with dates, pending results, what changed since the last visit.
03Require declared absence.For any agenda item with nothing in the source, the model writes "not documented." This is the signature technique of the card: it turns a silent omission into a visible gap.
04Ask for trajectory and contradiction.Serial values come as a sequence with dates and a stated direction. If the source contradicts itself, both versions appear, flagged.
05A summary is preparation, not documentation.It gets you ready for the visit; it does not go in the chart. If any of it does, it goes reviewed by you, with AI use recorded per local rules. And the directed verification below happens before any of it reaches clinical use.
04 · What not to paste
Direct identifiers come out

Name, record number, date of birth, address, phone, document numbers. Clinical dates stay, because dates are trajectory; date of birth goes, because it is identity.

Quasi-identifiers count too: a specific occupation, a small town, and a rare disease together identify someone without a name ever appearing. Swap in the clinically sufficient generic: age in years, occupation only if relevant and only as a broad category.

A black box does not anonymize

Covering is not removing. In a PDF, the black rectangle is a layer drawn on top; the text is still in the file and comes out with select and copy. Pixelating or blurring text is reversible, and image editors sometimes keep the pre-edit version inside the file itself.

Real redaction removes the data from the file. In practice, the safer path is to paste clean text rather than a redacted document. And the ceiling is the environment: Card 01 decides what may enter your tool at all.

Console · exercise

What did not survive

Read the source and the summary side by side, and mark in the source what did not make it across. The gesture of this exercise is the gesture of real verification: going back to the origin. Synthetic case.

The source · chart excerpts
Mar 1264-year-old woman. T2DM and hypertension, followed. Denies smoking. Allergy: penicillin (urticaria). Metformin 850 mg BID, losartan 100 mg daily. Cr 1.4 (eGFR 52).
Apr 30Dysuria; uncomplicated cystitis. Treated with nitrofurantoin (penicillin allergy on file). Cr 1.6.
Jun 15Lower extremity edema. Smoker, 20 pack-years. Cr 1.9 (eGFR 36). Renal ultrasound ordered.
The AI summary

64-year-old woman with T2DM and hypertension, followed regularly. On metformin 850 mg BID and losartan 100 mg daily. Renal function shows mild CKD, stable. Non-smoker. Cystitis in April, treated with nitrofurantoin, resolved. Renal ultrasound ordered, result pending.

Fluent, complete, and clinically wrong in three places. Nothing in the wording signals which three.

Omission · the allergy

The penicillin allergy is gone. That is the omission that hurts: the next amoxicillin prescription sails right through.

And notice the trail the source left. The cystitis was treated with nitrofurantoin precisely because of that allergy. The model read the consequence and dropped the cause.

Flattened trajectory · the creatinine

1.4, then 1.6, then 1.9 across three months became "stable." Every value was in the source. The direction was not in the summary, and the direction was the finding.

Silent contradiction · the smoking

March says non-smoker; June records 20 pack-years. The summary picked one and never mentioned there were two, which is exactly what an audit cannot afford.

What survived cleanly

The doses, the cystitis episode, the pending ultrasound. Not everything goes wrong, and that is the point of mixing correct items into every exercise here: the goal is calibration, not blanket suspicion.

05 · Before it reaches clinical use
0 of 6

Six items against the source, every time. Everything else, spot-check. This is what directed verification means: you check what is expensive to get wrong, not everything.

The prompt
Summarize the history below for today's visit, which is for
[reason for visit].
Fixed agenda: active problems; medications with dose and
frequency; allergies; most recent labs with dates; pending
results; what changed since [date of last visit].
For any agenda item with no information in the source, write
"not documented."
For serial values, show the sequence with dates and state the
direction: improving, worsening, or stable.
If the source contradicts itself, show both versions and flag
the contradiction.
Add nothing that is not in the text below.

[paste history with identifiers removed]
06 · Evidence

Models that clear 90 on knowledge exams score 44.8% on tasks built from real clinical text, and summarization is real clinical text1. For medical summarization specifically, there is a published framework with an error taxonomy that includes omissions, plus clinical rating of the potential harm of each error2.

  1. [article] BRIDGE: benchmarking large language models for understanding real-world clinical practice texts. Nature Biomedical Engineering, 2026. doi.org/10.1038/s41551-026-01719-2
  2. [article] Asgari E, et al. A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation. npj Digital Medicine, 2025;8:274. doi.org/10.1038/s41746-025-01670-7
Educational content; synthetic patient data. Not a substitute for clinical judgment or for the rules that apply where you practice.