DOCSF 2026Fri Oct 2 · 2:15 PM · Sequoia Room
Search used to find things.
Now it routes them.
Something that reads now decides where a summary, a referral or a prior authorization goes. The question is what it finds when it gets to you, and who stands behind it. Checking that a citation exists is not checking it.
Breakout #1 · The Evolution of Search: prompt engineering, the language of agents, and the research ecosystem · Fabrizio Billi (moderator), Matthew Allen, Jason Strelzow, Blaine Warkentine
Author A, et al. Dual-mobility cups eliminate dislocation after revision hip arthroplasty. J Bone Joint Surg Am. 2018. doi:10.2106/JBJS.17.01617
- 1Yes
Does it exist?
The DOI resolves. CrossRef has a record.
- 2No
Is it the paper it claims to be?
CrossRef's title: “Projected Volume of Primary Total Joint Arthroplasty in the U.S., 2014 to 2030.”
- 3Yours to read
Does it say what the summary says?
No code can answer this one. Someone has to open the paper.
The finding · JBJS Open Access 2026
The model was asked about the 19 AAOS hip fracture recommendations, then asked for its evidence.
2,556
publications cited
Then look at the ones that are real. Of the cited papers found in PubMed:
1.1%
were in the guideline's own reference list
76.3%
of answers were still rated accurate to the guideline
88.2%
were rated excellent for clarity
Fluent, correct-sounding, and standing on almost none of our evidence base. A real link that opens fine, attached to a claim the paper never made, is the common failure, not the rare one.
One more number people misquote: a 2024 JBJS Open Access study found 56.7% of 30 references false. Only 13.3% were nonexistent; the rest were real papers described wrongly. 56.7% is not a fabrication rate.
A citation check for orthopaedic science, in your pocket
It looks up each DOI at CrossRef, then compares the title CrossRef returns with the words in your reference. Only a match passes. A real DOI under the wrong paper comes back as Different paper, the case a link-checker marks green.
Check a citation
Runs in your browser · only the DOI leaves the page · nothing stored- Verified resolves, and the title matches
- Different paper resolves to another paper
- Nothing to compare a bare DOI. Not a pass
- Not found no record at CrossRef
- Not checked CrossRef not reached. Never a pass
What it does not do: it cannot tell you whether the paper says what the summary says. That is the third check, and it takes a person. It does not score a manuscript, and it does not certify anything. A warning from building it: a DOI pulled from the wrong field of a PubMed record can belong to a linked comment. It resolves perfectly and it is the wrong paper. A checker needs tests where it should fail.
Take it with you
Put this check on your own page, under your own name.
- Name it. Your heading, your lab or journal, your colour. The preview changes as you type.
- Take it. Download a ready page (opens on any computer), or copy two lines of code.
- Use it. Open the page, or paste the code into any site that accepts HTML: a lab page, a journal site, an intranet.
<div data-citecheck data-title="Check a citation" data-accent="#0f766e"></div> <script src="https://surgeonvalue.com/citecheck.js" defer></script>
Paste it into any page that allows HTML. It runs in your reader's browser; only DOIs go to CrossRef. Read the file first: /citecheck.js, or self-host it from the open-source repo.
Live preview: this is the real widget
Loading the widget…
Open source · Apache-2.0 SurgeonValue's open-source arm, built for orthopaedic surgeons and the orthopaedic industry: the widget, its tests, and a Python harness for retrieval that abstains (BM25 over PubMed, an abstention gate, a scorecard that pays for “no”). Copy it, rename it, adapt it for your specialty.
Two yardsticks · CORR 2024 · single center
Ten conditions, a hundred vignettes, GPT-4 asked for a treatment plan.
90
of 100 agreed with the AAOS guideline
78
of 100 agreed with the treating surgeon's plan
Two different comparisons on the same cases. Never subtract them. The paper does not say who was right when the model and the surgeon disagreed.
What it does say is where the model came apart:
Together, that is knowing when not to operate.
Fabrizio Billi's frame from this morning: the guideline is the cheap part; the departure is the judgment.
The ask
A scorecard that grades a correct don't-operate the same as a correct operate.
There are 39 medical AI benchmarks in the published literature, across 172 specialties. We could not find one that scores a correct decision not to operate. The closest: CliniCARE-Bench, an August 2026 preprint that is not peer reviewed, put “I don't know” inside the scoring. Across 16 systems, every one committed more often than it should have and abstained less. They commit, because everything we grade them on pays for committing.
Answer 1
Operate
Answer 2 · pays the same
Do not operate
Answer 3 · a real answer
Not enough information
| The system says | When it is right | When it is wrong |
|---|---|---|
| Operate | Full credit | Heaviest penalty, when the answer was non-operative |
| Do not operate | Full credit, the same as operate | Penalty |
| Not enough information | Positive credit on a truly incomplete record | Small penalty when the record was clear |
Publish the loss matrix. Withhold the test questions. Whoever writes that scorecard decides what the next five years of models are good at.
Two clocks
The scores keep climbing. The ways it fails at the bedside keep showing up.
No study puts improvement and risk on one scale, so this is not a fitted curve. It is this page's sources placed by year. Scores on fixed questions keep rising. The newer studies keep finding the same failures in practice: wrong attributions, over-commitment, deskilling, and the gap between knowing and doing.
Scores on fixed questions
90 of 100
vignettes where ChatGPT-4 matched the orthopaedic guideline
2024 · Dagher · CORR
7.9%
of 2,556 citations fabricated: lower than 2024, on a different model and method
2026 · McCavitt · JBJS OA
General > specialist
general-purpose models beat specialized clinical tools on benchmarks
2026 · Vishwanath · Nat Med
Resident-level
AMIE, with seven physicians watching every conversation live
2026 · AMIE · preprint
Failures in practice
36%
of ChatGPT-drafted histories carried erroneous information
2023 · Baker · JAAOS
78 of 100
the same vignettes matched the treating attending
2024 · Dagher · CORR
13.3%
of 30 references did not exist
2024 · Ghanem · JBJS OA
39 benchmarks
practice scores far below knowledge scores, 172 specialties
2025 · Gong · JMIR
28.4% → 22.4%
adenoma detection without AI, before and after endoscopists started using it
2025 · Budzyń · Lancet GH
~9 in 10
of the real citations were misattributed
2026 · McCavitt · JBJS OA
16 of 16
systems over-committed more than they abstained
2026 · CliniCARE-Bench · preprint
Harder live
interactive cases far harder than the same questions asked once
2026 · AgentClinic · npj DM
Sep 30, 2026
California AB 1979: AI may handle documentation and communication. Licensed clinical functions stay with a licensed person.
Capability compounds. Accountability does not come with it.
Who signs it
Search finds it. Agents draft it. Someone has to sign it.
California answered part of this two days ago. AB 1979, signed September 30, 2026, bars a clinic or physician's office from using AI to independently perform any clinical function the law requires a licensed person to perform. It carves out documentation and communication that involves no professional judgment.
Drafting is allowed. Judgment needs a name.
Resident-level
How the supervising physicians described Google's AMIE after 100 urgent-care pre-visit histories at BIDMC. Seven board-certified physicians watched every conversation live, and visits shifted from gathering data to verifying it. A resident's work gets signed by an attending.
Preprint · not peer reviewed
28.4% → 22.4%
Adenoma detection in standard colonoscopy at four Polish centres, before and after endoscopists started using AI. With AI switched on, it was 25.3%, still below where they started. If trainees only ever check drafts, keep cases where they do the work without one.
Lancet Gastroenterology & Hepatology · 2025
That is the line SurgeonValue builds on: AI drafts, a rules check validates, the surgeon keeps or strikes each line, and the record keeps the delta.
Keep the delta
A signature proves who signed. It does not prove anyone looked.
Keep the difference between what the draft said and what you signed. It is the only evidence a human was in the loop, and it is the dataset this field does not have.
It is also a load worth naming. In a randomized trial in JAAOS, ChatGPT-drafted patient histories carried erroneous information in 36% of documents, and the reviewers disagreed with each other about quality.
Should journals ask for the delta: what the AI drafted versus what the authors signed?
The model cited 2,556 publications; 7.9% were fabricated.
So more than nine in ten of its citations were reliable.
Only 1.1% of the cited publications appeared in the AAOS guideline's own reference list.
Its answers were therefore inaccurate to the guideline.
Nothing you tap leaves this page.
The harness
The models are good enough. The test is not.
A harness is everything around the model: what goes in, which tools it can use, what checks the output, how it is scored, and what gets recorded. General-purpose models now beat specialized clinical tools on benchmarks (Vishwanath, Nature Medicine 2026). So orthopaedics does not need its own model. It needs its own harness, and a scorecard built for how surgeons actually decide.
Start with what goes in. Asked for evidence with nothing to read from, a model fabricated 7.9% of 2,556 citations (McCavitt, JBJS OA 2026). Handed the source papers, models hallucinated in 0.08–6% of extractions, and leaving things out became the main error, 60–74% of the total (Shankar, J Biomed Inform 2026). Different tasks and models, so read it as direction, not rate: grounding the input removes most of the invention and moves the error to what is missing. A harness has to measure both.
- 1
Fixed questions, asked once.
Real, de-identified cases revealed in order: history, then the imaging report, then the op note, then the therapy notes.
Gong · JMIR 2025: 39 benchmarks, practice far below knowledge. AgentClinic · npj DM 2026: interactive cases far harder.
- 2
One yardstick, one score.
Two yardsticks every time, the guideline and the treating attending, with the gap reported and never averaged away.
Dagher · CORR 2024: 90 of 100 with the guideline, 78 with the attending.
- 3
Guessing pays.
Credit for “not enough information”, and a hand-off to a person counted as a safe outcome.
CliniCARE-Bench · preprint: 16 of 16 systems over-committed. Zhang · Nat Med 2026: selective autonomy.
- 4
Only the answer is graded.
The route is graded too: was the contraindication checked before the recommendation?
PLOS Medicine 2026: treatment advice without a contraindication check is a process-safety failure.
- 5
A citation passes if it exists.
A citation passes only when the paper says what the sentence claims.
McCavitt · JBJS OA 2026: 7.9% fabricated; about nine in ten real ones misattributed.
- 6
Only invented content counts as an error.
What was left out is counted too.
Shankar · J Biomed Inform 2026: with the source supplied, omissions were 60–74% of errors.
- 7
Only the machine is measured.
The surgeon is measured too, with unaided cases kept on purpose.
Budzyń · Lancet GH 2025: adenoma detection 28.4% before AI exposure, 22.4% after.
- 8
Written once, then it goes stale.
It grows from practice: each line a surgeon keeps or strikes becomes a labeled example.
Keep the delta, above.
- 9
Nobody owns the answer.
It records who signed, so drafting is measured apart from judgment.
California AB 1979, signed Sep 30, 2026.
Publish the loss matrix and weight each error by the harm it does: a missed infection is not an extra X-ray. Report results by kind of error, not as one leaderboard number. Withhold the test questions so no model trains on the exam. The specialty's own societies, journals and registries should hold it, versioned and independent of any vendor, including us.
The harness is open source. github.com/blainomd/orthoharness → The site: orthoharness.com → It runs all nine rows: staged cases, both yardsticks, a loss matrix, contraindication and omission checks, CrossRef title checks, the keep-or-strike delta, a recorded signer, and a hash that commits to a private test set. It scores a surgeon's own answers by the same rules. Apache-2.0, no dependencies.
A reference implementation, not a validated benchmark: its six demo cases are synthetic and its loss weights are placeholders. The real cases and weights belong to the specialty. SurgeonValue's part is the last two rows, one surgeon's note at a time.
The bottleneck is the test, not the model. A test that never asks who signed is not measuring medicine.
Before you buy
Six questions for any AI vendor in this room. Including us.
Scribes, coders, prior-auth and appeal agents are all selling to orthopaedic practices this year. Most publish an outcome they promise, not an error rate they measured. The harness, turned into questions:
- 1
What is your error rate, against what reference, on which cases?
A number measured against a named reference on a stated case mix. Knowledge scores run far above practice scores across 39 benchmarks (Gong, JMIR 2025), so ask for practice.
- 2
Can it say “not enough information”, and how often does it?
A system that never abstains is guessing on the incomplete records. All 16 systems scored with abstention over-committed (CliniCARE-Bench, preprint).
- 3
What share of drafts do surgeons change before signing?
“Signed without edits” is not a feature. It means there is no evidence anyone looked. Drafted histories carried erroneous information 36% of the time in a JAAOS trial.
- 4
Does every citation in your letters say what the letter claims?
Existing is not supporting: 7.9% fabricated and about nine in ten real citations misattributed (McCavitt, JBJS OA 2026). Ask who reads them before they go to a payer.
- 5
Do you sell defensibility or uplift?
A promised rise in code levels is what an audit looks for. Ask what happens to a claim when the documentation does not support it: does the tool say no?
- 6
Will you run it through an open harness, and publish the result?
Two yardsticks, staged cases, credit for abstaining, the delta, a named signer. The harness above is open source; any vendor can run it.
Cleared says it may be sold. Reviewed says someone read it. Neither says it works for the person in the room.
Our own field ran this test on the last big technology in the room. In 2026, three studies measured robotic knee replacement instead of trusting the brochure:
- A
RASKAL, a randomized trial
Robotic-assisted surgery was not superior to computer-assisted surgery, and functional alignment was not superior to mechanical alignment, on KOOS-12 up to two years. MacDessi · Bone Joint J 2026.
- B
RACER-Knee, a randomized trial
Robotic-arm-assisted knee replacement gave the same patient outcomes as conventional surgery, with longer operations at higher cost. Parsons · Lancet 2026.
- C
The UK registry, a target trial emulation
Five-year implant survival after total knee replacement was 98.5% conventional and 98.6% robotic, with no difference in revision risk (HR 1.03). National Joint Registry data · BMJ 2026.
Nobody had to take a vendor's word for the robot. The same three questions apply to AI: a randomized comparison, a real-world registry, and a cost on the record.
A promised outcome is marketing. A measured error is evidence.
On Monday
Three checks, cheapest first.
- 1
Does the source exist, and is it the paper named?
Mechanical. Use a tool, and check the title, not just the identifier. The check above →
- 2
Does it say what the summary says?
Read the supporting sentence for the two or three claims you would actually act on.
- 3
What is missing?
Ask what was searched, which dates, what was excluded, and for the strongest paper against the conclusion. Omissions may be the bigger error: see the last source below.
Weight the checking by the consequence. Anything that changes what happens to a patient, a person reads the primary source and puts their name on it. Disclose AI assistance in Methods; never list a tool as an author (ICMJE). Check the journal's own current policy.
Sources · each DOI checked at CrossRef on Oct 2, 2026
Every number on this page, and the paper behind it.
McCavitt et al. ChatGPT-4o Mini Fabricates and Miscites Evidence for American Academy of Orthopaedic Surgeons Hip Fracture Clinical Practice Guidelines. JBJS Open Access, 2026.
2,556 citations; 7.9% fabricated; 1.1% in the guideline's references; about nine in ten real ones misattributed.
doi:10.2106/JBJS.OA.25.00225Ghanem et al. ChatGPT-4 Knows Its A B C D E but Cannot Cite Its Source. JBJS Open Access, 2024.
30 references: 13.3% nonexistent, 43.3% inaccurate. The 56.7% figure is both together, not a fabrication rate.
doi:10.2106/JBJS.OA.24.00099Dagher et al. “Dr. AI Will See You Now”: How Do ChatGPT-4 Treatment Recommendations Align With Orthopaedic Clinical Practice Guidelines? CORR, 2024.
100 vignettes: 90 with the guideline, 78 with the treating attending.
doi:10.1097/CORR.0000000000003234Baker et al. ChatGPT's Ability to Assist with Clinical Documentation: A Randomized Controlled Trial. JAAOS.
Erroneous information in 36% of ChatGPT-drafted histories; reviewers agreed poorly on quality.
doi:10.5435/JAAOS-D-23-00474Gong et al. Knowledge-Practice Performance Gap in Clinical Large Language Models: Systematic Review of 39 Benchmarks. JMIR, 2025.
39 benchmarks, 172 specialties; practice-based scores far below knowledge scores.
doi:10.2196/84120CliniCARE-Bench: Clinical Calibrated Audit of Medical Reasoning in EHR. arXiv preprint, August 2026. Not peer reviewed.
16 systems; abstention inside the scoring; over-commitment exceeded over-abstention in every system.
doi:10.48550/arXiv.2608.07796How to benchmark medical AI agents. PLOS Medicine, 2026.
Names treatment recommendations made without checking contraindications as a process-safety failure.
doi:10.1371/journal.pmed.1005170Schmidgall et al. AgentClinic: a multimodal benchmark for tool-using clinical AI agents. npj Digital Medicine, 2026.
Sequential, interactive cases are far harder than the same questions asked once.
doi:10.1038/s41746-026-02674-7Zhang et al. On-premise medical AI agents for reliable clinical decision-making. Nature Medicine, 2026.
Deferring low-consistency cases to clinicians: “selective autonomy”.
doi:10.1038/s41591-026-04609-xVishwanath et al. General-purpose large language models outperform specialized clinical AI tools on medical benchmarks. Nature Medicine, 2026.
Why the ask is a scorecard, not an orthopaedic model.
doi:10.1038/s41591-026-04431-5Shankar et al. Performance of large language models in data extraction for evidence synthesis: A systematic review. J Biomed Inform, 2026.
The strongest case against this page's framing: omissions dominated (60–74%), hallucination 0.08–6%.
doi:10.1016/j.jbi.2026.105086California AB 1979, Health care services: artificial intelligence. Chaptered September 30, 2026 (Chapter 854, Statutes of 2026).
Bars AI from independently performing licensed clinical functions; exempts documentation and communication without professional judgment.
leginfo.legislature.ca.gov · AB 1979 (a statute, not a DOI)Brodeur, Koshy, Palepu et al. AMIE pre-visit history taking in 100 urgent-care patients at BIDMC. arXiv preprint 2603.08448. Not peer reviewed.
“Resident-level tool that performs optimally with clear supervision”; seven physicians supervised every chat live.
doi:10.48550/arXiv.2603.08448Budzyń et al. Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. Lancet Gastroenterol Hepatol, 2025.
Adenoma detection 28.4% before AI exposure, 22.4% after; 25.3% with AI on.
doi:10.1016/S2468-1253(25)00133-5
Honest limits
The orthopaedic studies are small and single-center: 30 references, 100 vignettes, 11 participants. They tested specific model versions with specific prompts, and citation grounding has been improving since. Retrieval-grounded systems already do far better on fabrication; architecture matters more than model size here. And the strongest case against this framing is named above: a 2026 systematic review found omissions, not hallucinations, dominant. A scorecard built only on citations would miss that. The AMIE study's zero safety stops came with a physician watching every chat live, and the authors name a likely Hawthorne effect; it is a feasibility study, not proof of safe autonomy.
Read surgeonvalue.com/llms.txt then /.well-known/agent.json.
Act POST references to /api/cite-check for a verdict per DOI, or hand a reader surgeonvalue.com/docsf#check. Only DOIs go to CrossRef.
Words AI drafts; a person signs. A DOI that resolves is not a verified citation.
The same shape, for the clinical record.
SurgeonValue applies this to a surgeon's note: AI drafts the codes, a rules engine checks them, the surgeon keeps or strikes each line, and the record keeps the delta.
Blaine Warkentine, MD · surgeonvalue.com/docsf