Clinical Note Consistency: Summary Claims vs. the Chart

Objective

Discharge summaries and referral letters are written from — but not mechanically checked against — the chart. A summary that states a procedure that wasn't performed, or omits a documented allergy, is a documentation error with clinical consequences. This recipe matches each summary claim to the chart section it concerns and flags the ones the chart contradicts, as a review assist for a clinician — and is deliberately loud about where NLI must not be the decision-maker.

!!! danger "Read this before anything else" This is a documentation-consistency assist, not a clinical safety device. It compares wording. It does not understand medicine, does not compare dosages or dates with typed precision, and must never gate a clinical decision on its own. Every flag is for a qualified human to adjudicate. Numeric and temporal facts (doses, timings, counts) are exactly where the model class is weakest — handle those with exact comparisons, not NLI.

Step 1: The chart, as atomic facts

CREATE TABLE IF NOT EXISTS recipe_clinical_chart (
    id INT PRIMARY KEY,
    section TEXT,
    fact TEXT
);

INSERT INTO recipe_clinical_chart (id, section, fact) VALUES
  (1, 'allergy-list',   'The patient chart documents an allergy to penicillin, which caused a rash.'),
  (2, 'operative-note', 'The operative note records that the appendix was surgically removed.'),
  (3, 'discharge-plan', 'The patient was discharged home with outpatient follow-up in two weeks.');

Step 2: Summary claims, each matched to its chart section

Match the claim to the relevant chart section first — with your record system's structure, not NLI. The chart_id below is that match.

CREATE TABLE IF NOT EXISTS recipe_clinical_summary (
    id INT PRIMARY KEY,
    claim TEXT,
    chart_id INT
);

INSERT INTO recipe_clinical_summary (id, claim, chart_id) VALUES
  (1, 'An appendectomy was performed.',            2),
  (2, 'A cholecystectomy was performed.',          2),
  (3, 'The patient has no known drug allergies.',  1);

!!! note "Why match, then verify — don't dump the whole chart" NLI_VERIFY reports a LIE if any anchor contradicts the statement (contradiction takes precedence). Hand it the entire chart as anchors and an unrelated line can spuriously contradict a true claim, turning everything into LIE. Verify each claim against the section it is about. When you don't have an explicit mapping, retrieve the relevant section by vector similarity first — the discipline in the RAG Answer Gate recipe.

Step 3: A procedure claim the matched note supports

SELECT JSON_UNQUOTE(JSON_EXTRACT(NLI_VERIFY('nli-default', s.claim,
         JSON_ARRAY(c.fact)), '$.verdict')) AS verdict
FROM recipe_clinical_summary s
JOIN recipe_clinical_chart c ON s.chart_id = c.id
WHERE s.id = 1;

Expected: TRUTH — "appendectomy" is entailed by "the appendix was surgically removed."

Step 4: A procedure claim the matched note contradicts

SELECT NLI_VERIFY('nli-default', s.claim, JSON_ARRAY(c.fact)) AS review
FROM recipe_clinical_summary s
JOIN recipe_clinical_chart c ON s.chart_id = c.id
WHERE s.id = 2;

Expected: verdict = LIE, violated_anchor = the operative note — the summary names a gallbladder removal the note does not record.

Step 5: A safety-critical claim the matched section contradicts

SELECT JSON_UNQUOTE(JSON_EXTRACT(NLI_VERIFY('nli-default', s.claim,
         JSON_ARRAY(c.fact)), '$.verdict')) AS verdict
FROM recipe_clinical_summary s
JOIN recipe_clinical_chart c ON s.chart_id = c.id
WHERE s.id = 3;

Expected: LIE — "no known drug allergies" is contradicted by the documented penicillin allergy. A summary that drops a recorded allergy is exactly the error worth catching.

Step 6: The honesty step — terse notation under-entails

Charts are terse. A shorthand allergy line may not lexically entail a fully-spelled claim, even when both mean the same thing:

SELECT JSON_UNQUOTE(JSON_EXTRACT(NLI_VERIFY('nli-default',
  'The patient has a documented penicillin allergy.',
  JSON_ARRAY('allergy: penicillin (rash)')), '$.verdict')) AS terse,
       JSON_UNQUOTE(JSON_EXTRACT(NLI_VERIFY('nli-default',
  'The patient has a documented penicillin allergy.',
  JSON_ARRAY('The patient chart documents an allergy to penicillin, which caused a rash.')), '$.verdict')) AS full;

Expected: terse = UNVERIFIED_PLAUSIBLE, full = TRUTH. The same true fact verifies or not depending on how the anchor is written. Normalize chart text into full sentences before verifying, and treat UNVERIFIED_PLAUSIBLE on a safety-critical claim as "route to a human", never "probably fine."

Step 7: The review queue — every claim, verdict first

SELECT s.id, s.claim, c.section,
       JSON_UNQUOTE(JSON_EXTRACT(NLI_VERIFY('nli-default', s.claim,
         JSON_ARRAY(c.fact)), '$.verdict')) AS verdict
FROM recipe_clinical_summary s
JOIN recipe_clinical_chart c ON s.chart_id = c.id
ORDER BY s.id;

Expected: claim 1 TRUTH, claims 2 and 3 LIE. The clinician reviews the two flags; they are not auto-corrected.

Cleanup (Optional)

DROP TABLE IF EXISTS recipe_clinical_chart;
DROP TABLE IF EXISTS recipe_clinical_summary;

Use it from your app

• Run the Step 7 query at summary-finalization time; surface LIE verdicts (and safety-critical UNVERIFIED_PLAUSIBLE) on the clinician's sign-off screen, with the violated chart line attached.
• Match claims to chart sections with your record structure (or vector retrieval); normalize chart shorthand into sentences first (Step 6).
• Why in-DB: chart and summary already live in the record system; the check is a query, and no PHI leaves the database for a third-party NLI service.

Key Concepts Learned

• Verify each claim against the section it is about — handing the whole chart as anchors lets an unrelated line spuriously contradict a true claim (contradiction precedence).
• Terse source text under-entails — normalize anchors to full sentences, or a true claim reads as UNVERIFIED_PLAUSIBLE.
• Dropped-contradiction (a missing allergy) and added-contradiction (a wrong procedure) both surface as LIE against the matched section.
• Dosages, timings, and counts are the model's weak spot; gate those with exact comparisons and leave NLI out of the numeric path entirely.