The Number Without a Body
A crypto desk published a story last week claiming a Microsoft AI diagnostic system outperforms doctors on complex cases. No sample size. No benchmark harness. No publication date. No named author. No link to a paper. Just a ratio โ 85.5% against 20% โ and an adjective: revolutionary.
I read it twice. The second pass was for structure, not content, because there was no content. The body restated the summary. That is not reporting. That is a ticker tape with a headline bolted on.
Here is what made me keep reading: the word system.
Not model. Not network. System. Whoever drafted that copy โ or whoever drafted the press materials it was laundered from โ chose a word that describes an assembly, not an artifact. In my world, that distinction is the entire trade. A model is a parameter file. A system is a routing decision. And routing decisions are where spread lives.
So I went looking for the thing the article was pointing at without describing. The fingerprint matched: Microsoft AI's Diagnostic Orchestrator, MAI-DxO, and the accompanying paper on sequential diagnosis with language models. Roughly 85.5% top-1 accuracy on New England Journal of Medicine Clinicopathological Conference cases, against a physician baseline in the low twenties.
Flag it now: every figure in this piece that originates from that paper is an inference from public memory, not from a document I have audited. The preprint has not been peer reviewed. The media layer has not been corrected. Treat the digits as directional and the architecture as the actual signal.
What Actually Shipped
Strip the framing. What exists is a harness.
An orchestrator model receives a case presentation โ age, sex, presenting complaint, vitals, initial labs. It generates a differential. It selects the next test. It receives the result. It revises. It repeats until it commits to a diagnosis or exhausts its budget. Alongside it sits a panel of frontier models, drawn from multiple vendors, invoked as a synthetic consult. The orchestrator arbitrates.
That is the invention.
No new attention mechanism. No novel training objective. No parameter-efficient structure. No dataset of a hundred thousand annotated differentials. The scientific contribution, if there is one, is the evaluation protocol: sequential diagnosis under a cost constraint, scored on top-1 accuracy against published cases with known answers.
I have watched this exact pattern play out in DeFi four times. Someone assembles existing primitives into a pipeline. The pipeline produces a number. The number becomes the narrative. The narrative gets priced as if the pipeline were the primitive.
It never is. The pipeline is the trade. The primitive is the commodity.
An Orchestrator Is a Router
Understand what an orchestrator does mechanically. It takes an input, decides which downstream resource to spend on, and prices the expected information gain against the cost of the call.
That is a router. In market microstructure terms, it is a smart order router. It is doing the same job a solver does in an intent-based DEX: decompose the objective, survey available venues, allocate flow to the venue with the best fill, respect a budget.
Once you see it that way, the result becomes legible โ and much less impressive.
The gain is not coming from a better model. It is coming from test-time compute plus model diversity plus domain tooling. Three things that are, today, purchasable by anyone with an API key and a credit card. The paper is not a moat. The paper is a spec sheet.
I have run this experiment myself. Last year I built a market-making bot for a mid-cap DeFi token โ reinforcement learning on order flow, ten thousand executions a day, roughly fifty basis points of edge per fill, six months, one point two million in realized profit against a two percent maximum drawdown. The edge did not come from a proprietary model. It came from latency, from queue position, from refusing to quote when the book was thin.
The intelligence was commodity. The execution discipline was not.
Same structure here. The intelligence is commodity. The discipline is the cost-aware test selection โ which brings me to the part of the paper that actually matters and that every headline buried.
Cost-Aware Decoding Is a Gas Limit
The system prices tests. It weighs the information gain of an order against the dollar cost of running it.
This is the only commercially load-bearing design decision in the entire architecture, and it exists because of a failure mode the authors clearly saw coming: an unconstrained diagnostic agent will order more tests, not fewer. Uncertainty is a cost the agent pays in someone else's currency. Give it an unlimited budget and it will resolve every ambiguity by buying information, because buying information is how you maximize accuracy on a benchmark.
In DeFi I call this the infinite-gas agent. Hand a naive bot an unlimited gas budget and no slippage cap and it will route through every pool on the chain, pay every fee, and still report a successful execution. The P&L is negative. The success metric is green. This is not a hypothetical โ it is the default state of every automated strategy before someone adds a budget constraint.
Cost-aware decoding is a gas limit for clinical reasoning. It is the difference between a strategy that prints and a strategy that bleeds.
Now the uncomfortable implication. If the accuracy figure was produced under a cost constraint, then the headline number and the deployment reality are measuring different things. Remove the constraint in production โ because a hospital optimizing for liability, not for spend, will absolutely remove it โ and the system reverts to test maximization. That flips the cost story from reduces spend to increases spend, and nobody in the coverage has priced that reversal.
Contamination Is a Provenance Problem
Here is the technical objection that should have led every story, and led none.
The NEJM Clinicopathological Conference cases are published. The diagnosis is in the title. Case 12-2024: A 45-Year-Old Man with Fever and Rash resolves, in the final paragraph, to whatever it resolves to. That content has been on the open web for years. Every frontier model's pretraining corpus covers it.
So when a system scores 85.5% on those cases, you are measuring a mixture of two things: reasoning and recall. Nobody outside the lab can tell you the ratio. And the ratio is the entire claim.
This is not a hypothetical risk. It is the default failure mode of every benchmark that gets popular. I watched it happen in crypto research the moment people started scoring models on smart contract vulnerability detection using public audit reports โ the models had read the audits. The benchmark measured memorization and the field spent eighteen months treating it as capability.
The fix is not a better benchmark. The fix is provenance. You need to prove, cryptographically, what went into the training set and what did not.
This is the single most under-built primitive in the AI stack, and it is a primitive my industry already knows how to build. An oracle is a provenance machine. A Merkle commitment over a corpus is provenance. A TEE attestation over an inference call is provenance. We have spent seven years arguing about whether an oracle is trustworthy and have built the entire machinery for answering that question.
Nobody has pointed it at medical training data. That is an open trade.
And before someone tells you zero-knowledge proofs solve it: I have priced that. Proving a forward pass through a mid-sized transformer is not free. Constraint counts run into the tens of millions per token, and prover cost lands orders of magnitude above the value of a single clinical inference. The floor didn't hold on zkML for exactly this reason โ the economics never closed at the verification layer, so the market quietly migrated to trusted execution and called it decentralized. Anyone quoting you a zk-verified medical inference in 2026 is quoting you a subsidy, not a cost.
The Baseline Was Built to Lose
Twenty percent. Hold that number against the literature.
Autopsy-controlled studies of diagnostic error in the United States consistently land in the ten-to-twenty percent range for major missed diagnoses. So a physician baseline of twenty percent on these cases is not doctors are bad at diagnosis. It is: this task setup is hostile to a human being.
Look at the human condition. No internet. No literature search. No chart review beyond the presented extract. No colleague to call. Single pass. Timed. No follow-up on the patient. No ability to reorder a test after seeing an ambiguous result.
Look at the machine condition. Multi-turn. Unlimited latency. Tool access. A panel of consultants. No cost of being wrong in the moment.
That is not a comparison. That is a handicap match with the weights published on one side only.
I have seen this exact sleight of hand in backtests. You compare a strategy on survivorship-bias-free data with realistic slippage against a naive buy-and-hold on a hand-picked window. The strategy wins. The strategy always wins. The comparison was constructed to produce that outcome.
The honest framing is not AI beats doctors. The honest framing is: an agent with unlimited time, full tooling, and a consult panel outperforms a single physician working under adversarial constraints. That sentence is true, boring, and commercially worthless โ which is precisely why it did not get written.
The Regulatory Moat Is Real and It Is Slow
Now the part the article skipped entirely: how any of this becomes revenue.
Diagnosis is the most heavily regulated, least liability-clear, slowest-to-deploy function in medical AI. Documentation is not. The two live in different regulatory universes.
Microsoft already owns the documentation layer. Nuance Dragon sits inside roughly half of US physician workflows. Azure carries the HIPAA and business associate agreements. Epic integration is in place. That stack is very good at transcription, coding, and revenue cycle management. It is not a diagnostic stack, and bolting diagnostic output into it drags the entire product line into device regulation.
The FDA's clinical decision support criteria turn on a single question: can a clinician independently review the basis for the recommendation? A system that outputs a ranked differential and a recommended next test is, in most readings, a medical device. Class II at minimum. Possibly higher. That is a 510(k), a De Novo, or a PMA. Twelve to thirty-six months. Millions of dollars. Under the EU AI Act, medical AI is high-risk and routes through MDR. In China, NMPA Class III.
Timeline from here: publication, then prospective trials at twelve to twenty-four months, then regulatory filing at twelve to thirty-six months, then scaled revenue. The first realistic million dollars of diagnostic revenue lands after 2027.
And the inference cost โ the thing everyone assumes is the blocker โ is nothing. Rough estimate: one to ten dollars per case for multi-model orchestration across dozens of round trips and hundreds of thousands of tokens. Against a two-hundred-to-four-hundred-dollar specialist visit and a one-to-five-thousand-dollar diagnostic workup, inference cost is a rounding error.
The bottleneck is liability. Who signs? If the agent recommends a test and the physician follows it and the outcome is bad, who owns that? The vendor, the hospital, or the human whose name is on the order? No one has answered. Until someone answers, there is no product โ there is a demo.
This is the same wall every DeFi protocol hits when it tries to touch real-world assets. The smart contract is trivial. The legal wrapper is the product. The wrapper is what takes three years and costs more than the engineering.
Where the Substitution Actually Happens
The industry impact is not uniform. Decompose it by function and the picture inverts.
Documentation is already substituted. Above sixty percent. It happened, it is done, the regulatory friction was low, and the value accrued to whoever owned the microphone.
Imaging triage is mid-substituted. Forty to sixty percent, with a high augmentation rate alongside it. Multiple imaging products already carry clearances. That market is contested and priced.
Complex differential diagnosis โ the thing in this paper โ substitutes at under twenty percent over a two-to-five-year window, with a very high augmentation rate. That is a second-opinion product, not a replacement product. It sells to the edge case, not the median case.
Test selection is the sleeper. Low-to-mid substitution, high augmentation, two-to-four years, moderate regulatory resistance. This is where the cost argument actually lives โ not in replacing the physician, but in stopping the reflex order.
The narrative market is enormous. Johns Hopkins work puts diagnostic error in the US at roughly seven hundred ninety-five thousand deaths or permanent disabilities annually. That is a number that gets put on slides. But narrative TAM is not addressable TAM. There is no reimbursement code for AI second opinion. Payers โ commercial insurers and CMS โ have not created one. Without a code there is no billing event, and without a billing event, the four-point-three-trillion-dollar US health spend is not a market you can sell into. It is a backdrop.
I have made this mistake. In 2022 I was sitting on fifty Bored Apes marked at four and a half million at the peak. The market for blue-chip PFPs was, on paper, tens of billions. When the floor dropped sixty percent, the addressable market was whatever a single counterparty would wire. I moved ten assets OTC at a twenty percent discount and took nine hundred thousand in stablecoins to cover fund liabilities. The paper value and the realizable value were never the same number, and the gap between them was the entire lesson.
Same gap here. The paper value of diagnostic AI is seven hundred ninety-five thousand annual events. The realizable value is zero until a payer assigns a price.
The Data Layer Is Where Crypto Actually Enters
Let me close the loop on the blockchain question, because the reflexive answer โ tokenize the medical data โ is the wrong one and it has failed twice already.
A diagnostic agent needs three inputs: the case, the literature, and the outcome. The first two are largely public. The third is the only genuinely scarce asset, and it lives inside hospital systems behind business associate agreements that prohibit the transfer. On-chain immutability is the exact opposite of what a HIPAA-covered entity is permitted to do with a patient record. You cannot cure that with encryption, because the regulated party is not permitted to make the commitment in the first place.
Which is why every DeSci health-data marketplace I have audited has the same shape: a consent narrative, a token, a staking mechanism, and no clinical dataset. The creators got paid. The data never moved. This is the OpenSea royalty lesson repeating in a different sector โ there is no sustainable on-chain business model for the party that actually produces the asset, because the moment you route value around the intermediary, the intermediary closes the gate and the asset stops flowing.
The workable version is not data markets. It is attestation markets. Sell the proof, not the record. A hospital that cannot export a chart can absolutely export a cryptographic commitment that a given model was trained on a defined corpus without the target cases. That is a compliance product with a buyer โ the payer, the regulator, the vendor facing a contamination audit โ and it never touches protected health information.
That is the only wedge I would fund.
Three Failures Nobody Cited
Any story about AI in clinical medicine that does not mention the prior attempts is not a story. It is an advertisement.
IBM Watson for Oncology. Internal documents surfaced in 2018 describing recommendations that were unsafe and incorrect. The system had been trained on a handful of synthetic cases and a small set of physician-authored ones. It was deployed anyway, because the sales motion ran ahead of the validation.
Epic's sepsis prediction model. A 2021 JAMA Internal Medicine evaluation found the deployed model's performance substantially below what had been claimed externally, and found it widened disparities across patient groups. The model was in front of clinicians at hundreds of hospitals. It was quietly recalibrated.
Google's diabetic retinopathy screening in Thailand. In field deployment, nurses overrode the AI at high rates because image quality in real clinics did not match the training distribution. The benchmark said one thing. The clinic said another.
Three failures, one pattern: benchmark performance and clinical adoption are uncorrelated, and the correlation gets worse as you move away from academic medical centers.
This is the same dissociation I watch in DeFi every cycle. TVL is not usage. TVL is a number you can print by incentivizing your own liquidity. The protocol with the highest TVL is frequently the one with the fewest organic users, because the incentive structure selects for mercenary capital. Benchmarks select for benchmark performance. Both metrics are self-referential.
The NEJM CPC cases come almost entirely from top-tier US academic centers. They are not representative of the global clinical distribution โ not of primary care, not of resource-constrained settings, not of atypical presentations, not of the patient who cannot give a coherent history. The system has been scored on the easiest slice of the hardest problem and reported as if it solved the problem.
The Moat Is Measured in Weeks
Here is the part that should worry Microsoft and should excite anyone building in this space.
The orchestration protocol is public the moment the paper posts. The prompt structure, the test-selection heuristic, the panel composition โ all of it becomes common knowledge. Replication cost is a few API keys and a weekend.
In crypto, that means two weeks to a fork. The half-life of an unpatented mechanism in an open ecosystem is measured in weeks, and value accrues to whoever owns distribution, not whoever published first.
Same will be true here. There is no defensible technical moat in multi-model orchestration. There is only a distribution moat and a regulatory moat. Microsoft has the first. It does not have the second.
And there is a strategic awkwardness the coverage missed entirely. If the accuracy gain comes primarily from the third-party models in the panel โ and reports suggest the panel draws from multiple frontier labs โ then the paper is evidence for the value of test-time compute, model diversity, and domain tooling. It is not evidence for Microsoft's model training capability.
That is why the word was system. A model claim invites comparison. A system claim does not.
I recognize this move. It is the same one every L2 makes when it publishes a throughput number that depends on a sequencer it did not write and a prover it did not build. The architecture diagram is the product. The components are someone else's.
And the proving cost problem is not solved. Rollups are still bleeding on proof generation, and the only reason it is not a crisis is that gas has not returned to bull-market levels. The moment throughput demand returns, operators with the thinnest margins go underwater. The same structural fragility sits under every verified-inference pitch in AI. The verification layer is subsidized. Subsidies end.
Epic Owns the Wallet
The competitive analysis that matters is not Microsoft versus Google versus the frontier labs. It is all of them versus the layer that owns the workflow.
Epic holds the EHR. Epic holds order entry. Epic holds the screen the physician is actually looking at. A diagnostic agent that does not sit inside that screen does not exist, regardless of its accuracy.
This is the wallet thesis, and it is the most reliable structural bet in both industries. The application that owns the user's entry point captures the flow. Everything upstream competes on price.
Microsoft's position: strong distribution via Nuance, strong compliance infrastructure via Azure, a real but dependent partnership with Epic. Its capability position is weaker โ no proprietary medical imaging model at Google's level, no diagnostic data asset at Epic's level, no frontier base model at the level of the labs whose APIs it orchestrates.
Distribution is a channel advantage, not a technical barrier. Channel advantages erode when the channel owner builds native. Epic shipping native diagnostic assistance is a far more plausible 2028 scenario than Microsoft out-modeling the frontier labs.
And the dependency cuts both ways. Every dollar of diagnostic narrative Microsoft generates increases Epic's leverage, because Epic is the gate.
What I'm Actually Watching
Four things. No predictions, just levels.
One: the spread between orchestration-layer and model-layer valuations. If orchestration is replicable in weeks, durable margin sits with whoever owns the interface and the liability signature. Any basket pricing AI agents as a monolithic category is mispricing the stack. The router is a commodity. The rail is not.
Two: cost per verified inference. Not cost per inference. Cost per verified inference. The gap between those two numbers is the entire value of every decentralized compute network currently trading on a promise. When someone publishes an attestation cost that clears a real workload, that is the signal. Until then, watch the proving overhead and assume it is subsidized.
Three: the reimbursement code. The day a payer creates a billing event for AI-assisted second opinion, the addressable market becomes real and the entire valuation stack re-rates. That is a regulatory event, not a technical one, and it will not be announced as a crypto story. It will be announced as a CMS rule.
Four: adoption at the edge, not the benchmark. Find a deployment in a resource-constrained setting. Measure override rate. If clinicians are dismissing the agent more than thirty percent of the time, the benchmark is decoration.
The floor didn't hold on the last generation of clinical AI, and it won't hold on this one until someone prices the liability. The accuracy number is not the trade. The accuracy number is the marketing. The trade is the routing decision, the cost constraint, and the signature at the bottom.
So here is the question worth sitting with. If the moat is distribution and the capability is a weekend of API calls, why is the market pricing this as a model breakthrough instead of a channel expansion โ and who is on the other side of that mispricing?