Prior record ingestion and summarization

A usable outside-record summary is a list of claims the packet already supports. Each line states one finding and names the source document: discharge summary dated 12 Mar 2026, BMP PDF from the sending lab, CT abdomen report. If you cannot point to a file in the packet, the line does not belong in the summary.

The model does not invent a diagnosis so the intake form looks complete. A diagnosis appears only when a source document states it. A problem list copied from another facility stays labeled as that facility's list, not as your assessment.

Empty is a valid result. When the packet never mentions A1c, an echocardiogram, or an allergy section, those rows stay blank. Blank means not in this packet. It does not mean normal, denied, or searched and unknown. You still ask the patient and order what you need.

The draft is a reading aid. It is not the chart until you accept or reject each line against the files.

Attach every file you actually received before you run summarization. Typical packets include scanned clinic or hospital notes, lab PDFs, imaging reports (sometimes the study), a CCDA or referral packet, and a medication list. Load what arrived. Do not wait for a complete longitudinal record.

Outside records arrive through health information exchange, Direct, patient upload, or the referral desk. The same review applies whether the sending organization uses Epic, Oracle Health, Microsoft, or athenahealth. Those systems are a transport class, not a completeness guarantee. A discharge summary from any of them can still be unsigned, truncated, or older than the labs sitting next to it.

Name files so a citation can resolve. Prefer sending facility, document type, and date over generic scan names. If two lab PDFs share a date, keep both and cite them separately. Do not merge every PDF into one blob and hope the model remembers which paragraph came from which file.

Do not mix this packet with answers the patient just gave in the waiting room. Those belong with automated anamnesis collection. If the patient reports apixaban and the outside list does not, that mismatch is for the medication reconciliation agent. Do not overwrite the cited outside list with the interview.

For imaging, ingest the report text you have. If you only have a photograph of a film and no radiologist report, say so in the packet inventory. Do not let the model turn a blurry image into an impression.

Run summarization only after the packet is loaded. Ask for a structured intake view: problems as the source stated them, procedures with dates, labs with values, units, and collection dates, imaging impressions with study dates, allergies as listed, medications as listed, and social or surgical history only when a document states it.

Every populated row needs a citation to a specific document in the packet. A line without a citation is a finding with no source. Delete it or regenerate. Do not repair it by guessing which PDF it came from.

Leave the field blank when the packet is silent. If no file contains an A1c, the A1c row is empty. If no document states heart failure, do not write HFrEF because the patient is on a loop diuretic. The diuretic belongs on the medication list, cited to that list. The diagnosis stays empty until a source names it or you make it after you see the patient.

Illustrative packet: new-patient intake with a two-page hospital discharge summary from facility A dated 4 Feb 2026, a BMP PDF from facility B dated 18 Feb 2026, and a CT chest report from facility A dated 3 Feb 2026. A valid summary might read as follows.

• Hospitalization for community-acquired pneumonia; completed five days of ceftriaxone (discharge summary, facility A, 4 Feb 2026).
• Creatinine 1.1 mg/dL; potassium 4.2 mmol/L (BMP PDF, facility B, 18 Feb 2026).
• No pulmonary embolism; residual right lower lobe opacity, clinical correlation recommended (CT chest report, facility A, 3 Feb 2026).
• A1c: blank; no HbA1c in packet.
• Echo or ejection fraction: blank; no echo report in packet.
• Assessment: blank; none of these three documents states congestive heart failure or COPD.

The last blank is the quality rule in practice. Residual opacity plus a diuretic on a med list is still not a license to write COPD exacerbation or chronic systolic heart failure into the intake assessment.

Do not pull treatment pathways into this summary. If you need a pneumonia follow-up pathway, use evidence-based guideline retrieval after you know what the packet actually said.

Open the cited file for every non-blank line. Confirm the value, the date, and that the document says what the line claims. If the summary says ejection fraction 35 percent and the only echo language in the packet is "echo pending," reject the line. If creatinine is 1.1 in the PDF and 1.8 in a narrative addendum, keep both with separate citations. Do not average them.

Check that blanks are truly silent. Skim for a lab or impression the model skipped. Missing a cited finding is an omission. Filling a blank with a plausible number is an invention. You can add omissions from the source. You discard inventions.

Do not treat the summary as the chart. Signing the intake note attests your review, not the model's draft. If the record system offers accept-all, do not use it. Accept or reject line by line.

The visit still happens. What you and the patient say in the room belongs with ambient clinical documentation. That capture does not backfill empty rows in the outside-record summary. New history from the room is new history.

When medications disagree across documents, list the conflict with both citations. Do not pick a winner inside this summary.

If a citation names a document you did not load, stop. Reload the packet and run summarization again. Do not edit the citation to match a nearby file.

A finding with no source: the draft includes history of CVA and none of the attached files mention stroke. Remove the line. If it remains, the next clinician will treat cerebrovascular disease as established.

Treating the summary as the chart: you paste the block into the HPI and sign. Months later an A1c value cannot be found in any lab because the model completed a blank. The summary was a reading aid. The chart is what you attested.

Inventing a lab: no BMP is in the packet, but the summary shows sodium 138. A round, normal value is the usual tell. Delete the row. Order the test if you need it. Do not cite inferred from clinical stability.

Related errors at the same severity: promoting another facility's problem list to your assessment; stamping today's date on an outside lab; merging two patients' files; reading an imaging photograph as a signed report; writing NKDA because the allergy section was empty. Empty is not NKDA unless a document states NKDA.

You walk in having read what the sending system actually sent, with every kept claim tied to a document, and with honest blanks where the packet said nothing. You still examine the patient, order what is missing, reconcile medications as a separate pass, and write the note you are willing to sign.

This summary is not a diagnosis, not a billing problem list, and not permission to skip the PDFs. Quality is a line that names its source, a blank that stays blank, and a clinician who still reviews.