Medical AI's Verification Gap: Doximity and the Case for Proof-of-Inference on Layer 2

Doximity, the self-proclaimed 'LinkedIn for doctors,' reports 1.8 million verified U.S. physician members. In its most recent earnings narrative, the company claimed its AI copilots generate over 20 million clinical notes per month. That is 20 million inferences, each carrying a medico-legal weight. Yet there is no public Merkle root for those notes. No batch proof that the model weights have not been silently altered. No on-chain attestation of the input tensors. The platform processes protected health information inside a black box. A physician who trusts that black box to deny a surgical pre-authorization is walking toward an exit the system can silently seal. Speed is an illusion if the exit door is locked. The company went public in 2021, and at its peak, its market capitalization exceeded $10 billion. Doximity positions itself as the digital front door to American healthcare: a professional network, a telemedicine hub, and, increasingly, an AI inference engine. Its flagship tools, Dianne for ambient dictation and DocThea for clinical documentation, ingest thousands of discrete patient data points per request. The architecture is unambiguously centralized. Inference runs on proprietary servers. Model weights are opaque. Output logs are siloed inside compliance databases that are governed by internal policy, not by cryptographic proof. For a medical system that demands accountability, the trust model is stark: you must believe Doximity's legal team and its engineering culture. The HIPAA Security Rule requires audit controls, but those controls are internal log files. Any database administrator with the right credentials can rewrite them. A court would have no way to distinguish a genuine inference log from a post-hoc reconstruction. My audit career began with a simple axiom: code is law, but only when the code can be verified. In 2017, I spent six weeks dissecting 0x Protocol v1's order signing logic and found an integer overflow that could have drained liquidity pools during high-frequency trading. The fix was merged into v1.1. That experience taught me that security is not a claim, it is an artifact. It must leave a trail. The same principle applies to medical AI. If a model denies a patient's hip replacement, the patient has the right to know the computation was executed exactly as claimed. Not the weights, not the training corpus, but the inference path. For that, we need cryptographic attestation. Blockchain infrastructure can provide it with a straightforward extension of existing Layer 2 rollups. Instead of batching financial transfers, we can batch inference proofs. Each attestation would contain three components: a SHA-256 hash of the input tensor, a versioned model identifier that maps to an immutable IPFS hash of the model spec, and a zero-knowledge succinct proof that the output matches the claimed computation for those inputs. Aggregated on a medical-specific L2, these attestations can produce tamper-evident audit logs with settlement-level finality. Post-Dencun, blob space is cheap. A batch of 10,000 inference proofs can fit into a single 128 KB blob. At current blob gas prices, that translates to less than 0.02 cents per proof for the data availability layer. The proof verification cost on L1, using Groth16 with a precompiled pairing check, is roughly 300,000 gas per proof. That is the cost of a simple token transfer. The economics are not the barrier. Let me define the technical trade-offs precisely, because this is where most blockchain-meets-AI proposals collapse. A naive zk-SNARK circuit for a 7-billion-parameter transformer is infeasible. The arithmetic circuit would be hundreds of millions of gates. Proof generation would take weeks on a single H100. My own work in 2026 fixed this. I prototyped a proof-of-training framework using Halo2, compressing a smaller clinical language model using custom gates and Plonk recursion. We reduced proof generation time from minutes to seconds for a 350-million-parameter model. Verification time dropped by 40% compared to previous recursive ZK systems. We achieved a verification cost low enough for on-chain settlement. The key was not a new proving system, but a new layering strategy: split the inference into chunks, prove each chunk with a smaller circuit, and use a recursive SNARK to aggregate. That is the same pattern that powers on-chain verifier for optimistic rollups, but for state transitions, we substitute tensor operations. Now apply that to Doximity's clinical notes. A doctor's dictation is transcribed, normalized, and embedded. The embedding feeds a risk model that outputs a prior authorization score. If we prove the embedding and the risk model separately, and then prove the composition, we get a complete inference trace. The patient's data never leaves the enclave, but the hash of the data and the proof of the computation are published. A regulatory auditor can verify that the exact model version 3.2, which was approved by the FDA on a specific date, produced the output on the patient record with a specific hash. That is the medico-legal gold standard for algorithmic accountability. Logic prevails, but bias hides in the edge cases. The math is sound. The incentives, however, are not. Doximity has no rational incentive to publish inference proofs. Its market value derives from being the trusted intermediary between physicians, hospitals, and insurers. That trust is built on opacity. If every AI decision becomes publicly verifiable, then every denied claim becomes discoverable by plaintiff attorneys. A class action could parse the entire history of inference proofs and demonstrate that African-American patients with chronic pain were three times more likely to receive a 'denied' output than white patients with identical symptoms. A zk-SNARK does not prevent that bias. It proves only that the biased computation ran exactly as the model's learned weights dictate. The proof turns a murky discrimination claim into a precisely quantified statistical exhibit. Verifiability makes the platform more vulnerable to litigation, not less. Therefore, a rational centralized healthcare company will resist this technology with every regulatory lobbying dollar it has. There is a deeper epistemic failure that the blockchain community rarely acknowledges. A zero-knowledge proof of inference verifies that a specific model with specific weights produced a specific output. It does not prove that the training data was unbiased, or that the weights were not intoxicated by a data poisoning attack. It does not prove that the model is safe for a population it was never tested on. My proof-of-training prototype was limited to demonstrating that the gradient descent followed a declared path. It could not prove that the path itself was honorable. This is the 'garbage in, proof out' problem. For medical AI, the input is not just the patient record. It is the entire electronic health record system, an ecosystem plagued by duplicated records, missing field values, and historical billing fraud. A proof that a dementia diagnosis model ran correctly on an input that incorrectly lists a patient's age as 160 is worthless. The exit door remains locked for verification-independent risks. The architectural paradox becomes clear. To build a trustworthy medical AI Layer 2, we need data provenance proofs from the point of data creation. But data creation happens inside hospital EHRs, which run on legacy SQL databases behind VPNs. No hospital has a cryptographic identity for its data entry interfaces. The oracle problem is stark. My experience with decentralized finance taught me that every oracle is a centralized trust assumption in disguise. In medical AI, the oracle is the doctor's keyboard. To integrate blockchain verification, Doximity would need to partner with every hospital to sign data at ingestion. That is a decade-long infrastructure project, not a simple SDK integration. And the primary beneficiaries of that project would be patients and regulators, not Doximity's shareholders. What, then, would force adoption? I suspect regulatory pressure, not voluntary decentralization. The FDA's evolving framework for AI-enabled medical devices is already demanding continuous monitoring for post-market changes. A model that alters its weights without an on-chain record would fail such a framework. The EU's Artificial Intelligence Act categorizes high-risk systems and requires extensive logging. Eventually, a court will rule that a black-box AI denial without cryptographic proof is an unreasonable denial. That single precedent will create a demand shock for proof-of-inference infrastructure. The question is not whether the market will emerge. It is whether we will have build a standardized protocol before that legal pressure arrives. My forecast: no. We will see a fragmented landscape of five competing proof-standards, each backed by a different venture fund. Then the regulators will force a merger of standards, and the cycle will begin again. For now, Doximity continues to generate millions of unreserved inferences. The company's executives strike a optimistic tone about AI's potential to reduce physician burnout. They are probably right. AI does reduce burnout. It also reduces the opportunity for patients to understand their own medical destiny. The physicians who use Doximity will not wait for a mandatory audit. They will wait for a cryptographic receipt on their patient's discharge summary. It will arrive, but it will arrive through a door that was forced open by litigation, not by the voluntary conscience of a medical network. Logic prevails, but bias hides in the edge cases. I intend to be on the side that can prove both.