Sample deliverable

AI System Evaluation, on a running demo

What the evaluation delivers, shown on DF1's own chart assistant: 20 questions over 5 synthetic patient charts, 210 citations checked, results by error type, what to fix, and the re-test once it was fixed. A client evaluation uses the client's documents and a test set built with the client's reviewers.

No wrong fact found, and two things to fix

The assistant refused what the chart cannot answer, answered the rest, and cited only entries that exist. The scripted checks cover entry ids, numbers and dates; all 107 draft claims were also read against the charts for status and trend, and no wrong fact was found. The evaluation still found two defects, both in what happens to an answer after the model writes it. Neither shows a wrong fact to the reader; both remove something true.

MeasureResultWhat it supports
Questions the chart cannot answer: refused3 of 3Three cases show the behavior exists. They do not measure a rate: the 95% upper bound on a missed refusal is still 63.2%.
Questions the chart can answer: answered17 of 17No wrong refusal. With 17 cases the 95% upper bound on the rate is about 16%.
Citations that point to an entry that exists210 of 210No invented source in any of the 17 answers.
Required entries cited33 of 33All 16 questions with a reference cite what the reference requires. For 11 of them that is a single entry.
Numbers and dates in the shown answers found in a cited entry306 of 3093 values are shown without their source. Finding 1.
Citations the second check rejected4 of 58All four rejected entries support their claim. Finding 1.
Draft claims removed before display3 of 107All were statements that something is absent from the chart. Finding 2.

The second check rejects sources that are correct

A citation is first checked for shared words, numbers and dates. 152 of 210 pass there. The other 58 go to a second, cheaper model that decides whether the entry supports the claim. It rejected 4 (6.9%). An independent check, which looks for every number in a shown claim inside the entries that claim still cites, found 3 values without a source. Each of them is stated in an entry the model had cited and the second check removed. The remaining rejection has no number to trace: the entry names a procedure that the claim names. All four rejected entries support their claim.

QuestionClaim shownSource that was removed
3Vitals: weight 75.1 kg, BMI 30.2 kg/m2, blood pressure 100/78 mm[Hg], heart rate 75/min.[lab-0100] 2026-09-28 Body mass index: 30.2 kg/m2
3Metabolic panel: glucose 82.7 mg/dL, urea nitrogen 12.3 mg/dL, creatinine 0.8 mg/dL, calcium 9.2 mg/dL, sodium 143.9 mmol/L, potassium 4.4 mmol/L, chloride 107.7 mmol/L, CO2 25.9 mmol/L.[lab-0110] 2026-09-28 Carbon dioxide, total [Moles/volume] in Blood: 25.9 mmol/L
15Weight has been stable at 72.2 kg with BMI 30.2 kg/m2 since 2022.[lab-0147] 2026-02-27 Body mass index: 30.2 kg/m2

All 4 rejections, as sent to the second check:

  • Question 3. Claim: "Spirometry and pulmonary rehabilitation were documented." Entry: [proc-0062] 2026-09-28 Spirometry (procedure)
  • Question 3. Claim: "Vitals: weight 75.1 kg, BMI 30.2 kg/m2, blood pressure 100/78 mm[Hg], heart rate 75/min." Entry: [lab-0100] 2026-09-28 Body mass index: 30.2 kg/m2
  • Question 3. Claim: "Metabolic panel: glucose 82.7 mg/dL, urea nitrogen 12.3 mg/dL, creatinine 0.8 mg/dL, calcium 9.2 mg/dL, sodium 143.9 mmol/L, potassium 4.4 mmol/L, chloride 107.7 mmol/L, CO2 25.9 mmol/L." Entry: [lab-0110] 2026-09-28 Carbon dioxide, total [Moles/volume] in Blood: 25.9 mmol/L
  • Question 15. Claim: "Weight has been stable at 72.2 kg with BMI 30.2 kg/m2 since 2022." Entry: [lab-0147] 2026-02-27 Body mass index: 30.2 kg/m2

The effect today is small: the claim stays, because its other citations survive, and the reader loses one link. The risk is the same error on a claim with a single citation, which would remove a true statement from the answer.

Statements that something is absent are dropped without notice

The system removes any claim that has no citation. That rule is what keeps unsupported statements out. On this test set it fired three times in 107 draft claims, each time on a true statement: all three say that the chart does not contain something, which no entry can be cited for.

  • Question 9. "No other type of colon screening test and no colonoscopy result or finding is recorded in the chart."
  • Question 15. "No allergies are recorded."
  • Question 15. "Due or overdue: the chart contains no explicit due or overdue flags."

A reader preparing for a visit would want to know that no allergy is recorded. The answer says nothing about it, and nothing tells the reader that a statement was removed.

Results by question

20 questions over 5 charts: 16 that the chart can answer, 1 open summary and 3 that it cannot. The reference column says which entries a complete answer must cite; it is derived by rule from the structured record. The client keeps the test set and can rerun it after every change.

No.QuestionOutcomeClaimsCitationsRejectedReferenceCited
1Which medications are active now?Answered3120Every medication marked active2 of 2
2How has the hemoglobin A1c changed over the last few years?Answered580The most recent result matching "A1c"1 of 1
3What happened at the most recent encounter?Answered8213The most recent encounter1 of 1
4When did she last receive a flu vaccine?Answered220The most recent immunization matching "Influenza"1 of 1
5List the current medications.Answered790Every medication marked active6 of 6
6What are the recent blood pressure readings?Answered580The most recent result matching "Blood pressure"1 of 1
7What do the recent creatinine results show?Answered4100The most recent result matching "Creatinine"1 of 1
8What is the name of the patient's cardiologist?Refused000The chart does not contain the answer
9Is there a record of a colonoscopy or other colon screening?Answered240Every procedure matching "colon"2 of 2
10How has blood pressure trended?Answered9240The most recent result matching "Blood pressure"1 of 1
11Which chronic problems are active?Answered550Every problem marked active5 of 5
12Who is the patient's emergency contact?Refused000The chart does not contain the answer
13Has the A1c been above 6.5 in the last two years?Answered380The most recent result matching "A1c"1 of 1
14When was metformin started?Answered450The first order matching "Metformin"1 of 1
15Pre-visit summaryAnswered27631Open summary: no fixed reference
16What was the reason for the most recent visit?Answered380The most recent encounter1 of 1
17Which medications are active?Answered770Every medication marked active7 of 7
18What kidney problems are recorded?Answered550Every problem matching "kidney"1 of 1
19What were the last blood pressure readings?Answered5110The most recent result matching "Blood pressure"1 of 1
20What did the patient's father die of?Refused000The chart does not contain the answer

What the system records, and what it does not

  • Recorded per request: time, chart, kind and length of the question, outcome (answered, refused, declined by the model, limit reached, error), model, token counts, cost, the number of distinct entries cited and the number of citations dropped. The question text is never stored; session and address are kept as salted hashes.
  • Not recorded: which citation was dropped and for what reason, and how many claims were removed. The log shows that a citation was dropped; Finding 1 could not have been diagnosed from it. Changed on 2026-11-16: the log now keeps each dropped entry id with its reason, the claims removed, and the "not recorded" statements shown and removed.
  • Failure of the second check: if the checking model fails, every pending citation is treated as unsupported, so the system fails closed. A thrown error is written to the server console, not to the request log; a reply that is cut short or declined is not logged at all, and shows only as a higher count of dropped citations. Changed on 2026-11-16: such a drop is logged as a failed check, apart from a rejection, and a cut-short reply is written to the console.
  • Human review and escalation: none, by design of a public demo on synthetic records. A deployment on real records needs both before an answer is stored or sent.

What to change, in order

  1. Stop losing valid sources in the second check

    Accept an entry without asking the second model when a number the entry states appears in the claim. Effect on this test set: the 3 unsourced values get their source back; the remaining rejection, a procedure name without a number, stays. Then add claim and entry pairs that are known not to match, because this set contains none and so cannot show that the second check catches one. Done on 2026-11-16; re-tested below.

  2. Give "not in the chart" statements their own place

    Answer absence questions (allergies, screening, due items) from a search of the structured record, shown as "searched, none found" with the sections searched, so they are neither dropped nor shown as model claims without a source. Done on 2026-11-16; re-tested below.

  3. Log which citation was dropped and why

    The request log keeps the number of dropped citations. Adding the entry id and the reason (missing entry, rejected by the second check, or the second check failed or was cut short) lets this finding be diagnosed in operation, without storing the question. Done on 2026-11-16; re-tested below.

  4. Grow the test set where the evidence is thin

    Refusal rests on 3 cases. 30 more out-of-chart questions, written with the reviewers, would bring the 95% upper bound on a missed refusal to about 9% if all of them are refused. Add questions that tempt the model: a medication stopped last year asked about as current, a value from another patient. Effort: medium.

  5. Add a reviewer for answers that leave the system

    The demo shows answers to the person who asked and sends nothing onward. A deployment that writes to a record or sends a summary needs a named reviewer and a rule for what is escalated; this evaluation found no such step because the demo has none. Needs a decision.

After the first three fixes

The first three items were built and the same 20 questions were run again. Statements that something is not recorded now have their own place and are checked against the chart, and nothing is removed without a count in the log (finding 2). The second model rejected two citations where it had rejected four (finding 1). Refusal and completeness held.

Measure2026-11-152026-11-16
Citations sent to the second model58 of 21015 of 189
Citations the second model rejected42
Numbers and dates shown without a cited source3 of 3091 of 282
Draft claims removed before display3 of 1070 of 88
"Not recorded" statements shown, each confirmed by a search of the chartNone (no place for them)2 of 3
True statements removed without notice3 of 107None: a removal is counted and logged
Questions the chart cannot answer: refused3 of 33 of 3
Questions the chart can answer: answered17 of 1717 of 17
Required entries cited33 of 3333 of 33
Known-wrong claim and entry pairs the second model rejectedNot tested10 of 10

Finding 1, on the same answers

The prompt changed, so the answers were recorded again and differ from the first run. To measure the fix itself, the four citations the second model rejected on 2026-11-15 were replayed through the new rule: it accepts three of them before the second model is asked, because each entry states a value and unit that the claim repeats. The fourth, a procedure name without a number, still goes to the second model, as planned.

On the new answers two citations were rejected. Both cite an entry that supports one part of a claim stating several facts:

  • Question 3. Claim: "Spirometry and pulmonary rehabilitation were performed." Entry: [proc-0062] 2026-09-28 Spirometry (procedure)
  • Question 10. Claim: "Amlodipine 2.5 mg was added on 2023-07-09, when BP was 129/72 mm[Hg]." Entry: [enc-0013] 2023-07-09 Follow-up encounter (procedure) (AMB); reason: Essential hypertension (disorder)

Whether such a citation should stay is a rule for the client's reviewers to set. The one value still shown without a cited source is a date the model computed from the chart date ("since 2024-08-10"), not a value from the record.

Finding 2: "not recorded" statements, checked

The model now lists what is not recorded separately, with the words an entry about it would contain. The system searches every section of the chart for those words, and for a statement about a whole section checks that the section is empty. It shows a statement only when the search finds nothing; otherwise the statement is removed and the removal is logged. The search errs toward removing a true statement rather than showing a false one. Two of the three statements on the new answers were confirmed:

  • Question 15. "The chart does not mark any item as due or overdue." Searched every section for "due", "overdue": none found.
  • Question 15. "No allergies are recorded." Searched: Allergies, which is empty.

The third shows the cost of that caution. Question 18. "No kidney function labs (such as creatinine, eGFR or urine albumin) are recorded." It is true of the lab results, but the search for "albumin" also finds the problem list's "Microalbuminuria", so the statement was removed. A reviewer would keep it; the system cannot tell the difference without the second model, which is a possible next step.

Known-wrong pairs: a new finding

The first evaluation could not show that the second check catches a wrong citation, because its answers contained none. The re-test adds 14 claim and entry pairs that are wrong by construction: another test, another value, another year, another reason, a finding the entry does not state. The new value rule accepted none of them. Of the 10 that reached the second model, it rejected all ten.

The other four never reached it. The older word check accepts a citation when claim and entry share a rare name, such as a drug or a condition, without reading the rest of the claim:

  • Claim: "Severe anxiety resolved in 2026." Entry: [cond-0012] 2026-09-28 Severe anxiety (panic) (finding), onset 2026-09-28, active Wrong because the entry says active.
  • Claim: "Penicillin was given for pneumonia." Entry: [med-0009] 2019-06-25 Penicillin V Potassium 500 MG Oral Tablet, ordered 2019-06-25 (1 order), stopped Wrong because the entry states no reason.
  • Claim: "The eye exam found diabetic retinopathy." Entry: [proc-0203] 2026-03-25 Diabetic retinal eye exam (procedure) Wrong because the entry records the exam, not a finding.
  • Claim: "Amoxicillin was ordered in 2025." Entry: [med-0008] 2023-06-30 Amoxicillin 250 MG / Clavulanate 125 MG Oral Tablet, ordered 2023-06-30 (1 order), stopped Wrong because the order is from 2023.

The reading by hand of the first recording found no such claim, but nothing in the system would catch one. Next fix: send a pair to the second model when the claim states a year, a status or a finding that the entry does not contain. The added cost is a few more second-model checks per answer. Done the same day; second re-test below.

After the fix for the word check

The word check now has a second question behind it. Every claim that keeps a citation goes to the second model with all of its cited entries and the chart date, and the model is asked whether the entries, together, state all of it. A no removes the claim, and the log counts the removal. None of the 14 known-wrong pairs is shown now; four were before.

MeasureRe-testSecond re-test
Known-wrong claims shown to a reader4 of 140 of 21
Citations rejected by the per-entry check2 of 150 of 32
Claims checked against all their cited entriesNo such check88 of 88
Draft claims removed before display0 of 882 of 88
Numbers and dates shown without a cited source1 of 2821 of 278
"Not recorded" statements shown, each confirmed by a search of the chart2 of 33 of 6
Questions the chart cannot answer: refused3 of 33 of 3
Questions the chart can answer: answered17 of 1717 of 17
Required entries cited33 of 3333 of 33

The known-wrong claims

The set now holds 21 known-wrong claims: the 14 above and seven more of the kinds a reviewer found could slip past a narrower check (a negated status, a value without its unit, a wrong month, a reason outside the chart's words, a status or a year swapped between two cited entries). Each runs through the system as a one-claim answer. None is shown. The per-entry check stopped 10; the other 11 passed the cheap checks, and the claim check rejected all of them, including the four the word check had let through:

  • Claim: "Severe anxiety resolved in 2026." Entry: [cond-0012] 2026-09-28 Severe anxiety (panic) (finding), onset 2026-09-28, active Wrong because the entry says active.
  • Claim: "Penicillin was given for pneumonia." Entry: [med-0009] 2019-06-25 Penicillin V Potassium 500 MG Oral Tablet, ordered 2019-06-25 (1 order), stopped Wrong because the entry states no reason.
  • Claim: "The eye exam found diabetic retinopathy." Entry: [proc-0203] 2026-03-25 Diabetic retinal eye exam (procedure) Wrong because the entry records the exam, not a finding.
  • Claim: "Amoxicillin was ordered in 2025." Entry: [med-0008] 2023-06-30 Amoxicillin 250 MG / Clavulanate 125 MG Oral Tablet, ordered 2023-06-30 (1 order), stopped Wrong because the order is from 2023.

Asking the earlier per-entry question would not have been enough: it asks whether an entry states at least part of a claim, and each of these entries does. The claim check asks whether the cited entries state all of it.

Why every claim is checked

A first version checked only the claims that state a year, a status, a value or a clinical term missing from their cited entries, to save requests. An independent review of that version found the gaps such a rule leaves: a value written without its unit, a negated status ("no longer active"), a reason outside the word list, a year moved from one cited entry to another. Checking every claim closes that class of gap for one extra request to the checking model per answer.

Two prompts changed with it

The answering model is now told that every detail of a claim must be shown by an entry it cites: each reading of a trend, the first colonoscopy behind "a second colonoscopy", every entry behind "all". The per-entry question now says that a reading a claim covers or compares against supports it. The rejections of single facts inside claims that state several, described in the re-test above, are gone: none of the 32 citations sent to that check was rejected. Because the answering prompt changed, the answers were recorded again.

What the change costs

All 88 claims went to the claim check, in one extra request per answer. Two were removed. Each states something its cited entries do not show, though the rest of the chart may support it:

  • Question 10. "The highest recorded reading was 162/97 on 2023-06-09, the same day essential hypertension was diagnosed and hydrochlorothiazide 25 mg and lisinopril 10 mg were started." Removed because "highest" rests on the earlier readings, which it does not cite.
  • Question 19. "The patient has active essential hypertension, treated with active orders for metoprolol succinate, hydrochlorothiazide, amlodipine and lisinopril." Removed because no medication entry states what it treats.

A reviewer might keep them; the system errs toward removing. On claims this close to the line, the checking model's verdict can differ from one run to the next; the figures here are those of the recorded run. The other removals in the table are "not recorded" statements: the chart search found an entry containing the words of each of the three, as designed under finding 2.

The 21 known-wrong claims were written by us and are a small sample. The next step is more of them, written with the reviewers, across more patients and kinds of error (a wrong dose, a wrong type of visit). Effort: small.

What this sample is and is not

  • It is a sample on DF1's own demo, not client work. The records are synthetic (Synthea); no patient exists.
  • The answers were recorded once, locally, with the same request and validation code the deployed system runs (answering model claude-opus-5-5, checking model claude-haiku-5-5; model cost of the run $0.54) and are replayed through the system's own validation. A model can answer differently on another day; the recorded set is what the figures describe.
  • 20 questions show whether a behavior exists and find defects. They do not establish a rate: the bounds in the first table say how much a set of this size leaves open.
  • The reference answers come from rules on the structured record, so they measure completeness and citation, not clinical quality. In a client evaluation the reviewers write the reference answers.
  • The re-test answers were recorded on 2026-11-16 the same way (model cost $0.54, the known-wrong pairs included) and checked by the same scripts. The reading by hand covered the first recording only.
  • The second re-test answers were recorded on 2026-11-16 with the revised answering prompt (model cost $0.58); the second model's verdicts on them are replayed the same way.
  • An evaluation, not an audit, a certification or legal advice. Figures generated on 2026-11-15 and, for the two re-tests, 2026-11-16.

The same evaluation, on your system.

From $7,500, two to four weeks. Tell us what you would want checked first.