Interrogate the Brief Before You Render
A Microsoft diagnostic system beat experienced doctors on hard cases mostly by asking 25 questions where they asked five. The transferable part is not the model. It is the order of the interrogation, the dissenting role, and a hard price on every test.
AI-assisted draft. Reviewed and edited by the Phosphene team before publication.
On this page
Le Monde reported on 28 October that a Microsoft Research system called Diagnostic Orchestrator, built around GPT-5, reached the correct diagnosis on 83 percent of hard cases from the New England Journal of Medicine. The study ran in Science on 22 October. The doctors in the same trial managed 36 percent when they had reference material to hand, and 22 percent without it. The best single clinician in the group cleared 51 percent.
The number that matters for anyone writing prompts is not 83. It is 25. Doctors ordered their first test after five or six questions. The system asked more than twenty-five before it ordered anything.
The part that is actually new
Language models have been beating clinicians on multiple-choice diagnostic vignettes for a while. Those vignettes are the problem: the chief complaint, the history, the exam findings and the lab results all arrive as one block of text. A model that reads the whole block at once and names the disease is doing retrieval, not investigation. Nobody can score whether it asked a good question, or ordered a test it did not need.
DxO changes the shape of the task. The case is a simulation. The system gets a short opening vignette and then has to ask for what it wants, one step at a time, until it believes it knows enough. When a piece of data does not exist in the record, the harness returns a neutral factual answer rather than an error, so that a gap in the file does not quietly become a clue pointing at the answer.
Three simulated roles do the work, all played by the same underlying model. One maintains and continuously re-ranks the differential list. One plays devil's advocate and attacks whatever hypothesis is currently leading, to stop the system settling on the first plausible story. One plays the economist and prices each proposed test against its expected diagnostic yield before approving it.
The results held across two corpora. On NEJM cases, 83 percent against 36 and 22. On 325 emergency medicine cases drawn from MIMIC-IV-ED, 80.3 percent against 46.8 for the participating physicians. The best clinician on the emergency set hit 62.5 percent.
What a one-shot prompt actually is
Look at your last image prompt. If it starts with a subject, a lens, a light source, a palette, a mood and a style reference, you have handed the model the entire chart at once and asked for the diagnosis. That is the vignette format the researchers were trying to get away from, and it fails for the same reason: nothing in it forces the model to notice what it does not know.
The unspecified parts of a brief do not stay empty. They get filled with whatever is statistically likely, presented at full confidence, and blended into the render so smoothly that you stop looking for them. This is why a batch of images can be technically clean and still wrong in four places nobody asked about.
The DxO result says the fix is procedural rather than textual. Withhold, then request.
Four moves worth stealing
Ask what is undetermined before you ask for an image. Put the scene and the constraint into a cheap text exchange first, and ask for the list of facts a specific shot would need which the brief does not currently decide. Which wall is behind the subject. What time of day. Whether the room is occupied. Answer those, then write the prompt. In a pipeline with a text model in front of the image model, this is two messages and no render credits. Without one, do it on paper in ninety seconds.
Keep a dissent lane. The devil's advocate exists to attack the leading hypothesis, not to add variety. The image equivalent is deliberate: after your first pass, generate the composition that would be correct if your reading of the brief were wrong. Not "another option." The competing interpretation, built to beat yours. If it does not beat yours, you learned something about the brief. If it does, you just saved a revision round with the client.
Put a price on every render. The economist role rejects a test whose expected yield does not justify its cost, and that is why the system spends less than the doctors while being more accurate: 36 percent correct at over 4,700 dollars per case for the doctors with reference material, against a system that, in the emergency set, hit 70 percent at roughly 2,000 dollars and 80 percent at roughly 5,000. Write down what decision a given render is supposed to settle. "Does the 35mm compression separate the two figures or flatten them?" If a render cannot discriminate between two live options, you are not buying information, you are re-rolling taste at full price.
Say "undefined" out loud. The harness returns neutral text for missing values so that absence carries no signal. Your prompt does the opposite by default. If the background must be plain, write plain. If there must be no logo, write that. Leaving a detail out is not the same as specifying it as unimportant.
Where the comparison breaks
This was a text-only benchmark. No radiographs, no electrocardiograms, no pathology slides, and the authors say so plainly: it is still a long way from medicine at the bedside. The NEJM corpus is selected for rare and atypical disease, which flatters a model trained on the whole literature against a physician who sees maybe two of these cases in a career. Doctors and system both interacted through a closed written question-and-answer loop, without the noise of a real consultation, an interrupted patient, or an incomplete file. And the system was compared against individual physicians, not against a multidisciplinary team that argues in a room.
Image work has no clean conditions at all. The brief arrives as a mood board, a deadline and a client who says "warmer." So take the procedure and leave the confidence. Sequential disclosure is a real advantage when the problem is underdetermined and the cost of a wrong guess is high. It is close to useless when you already know exactly what you want and the only open question is which of three renders you prefer. In that case the fastest path is to render all three.
The result that outlasts the model
The orchestration alone was worth 7.9 points on the NEJM set and 4.4 on the emergency set, measured against the same models reasoning on their own. GPT-5, Grok 4, Gemini, Claude 4.1 Opus and Claude 4.5 Sonnet were all run unassisted for comparison. Then the researchers rebuilt DxO on GPT-5.5 and the advantage over the bare model was still there.
A model upgrade did not erase the workflow advantage. That is the sentence to keep. It is the same argument as the studio that owns its pipeline instead of chasing checkpoints, and it rhymes with why a single graph beats a swarm of agents: the structure around the model is doing work the model cannot do for itself.
Run it on the next brief
- Write the one sentence the shot has to deliver.
- List five facts the shot needs that the brief does not decide.
- Decide them, or mark them deliberately neutral.
- Write the prompt from those decisions, not from adjectives.
- Render once. Then render the interpretation that would be right if you were wrong.
- Keep whichever survives, and note what settled it.
The doctors in that study were not careless. They were fast, and speed means asking the first five questions that come to mind. The system was slow on purpose, in a place where slowness was free and a wrong test cost money. Most of the money spent on AI image work is spent the other way round: fast on the thinking, slow and expensive on the rendering.