Ask us the hard ones.

Today: self-assessed against the open DSS methodology, published in full in WP-001 and DSS-001, with the fault-tree derivation, assumptions, and supplementary data public so anyone can re-run the mathematics. And here is the part we designed on purpose: our own standard requires third-party audit at its Gold and Platinum tiers. We wrote a bar we must publicly clear, and independent audit is formally invited. A vendor grading itself against a secret rubric is marketing; a vendor publishing the rubric and inviting the examiner is a standard.

No, and the distinction is printed on the certificate. It is deliberately two-dimensional: architecture · probability. DeepSensi implements the complete Platinum architecture, all five pillars at the highest tier. Probabilistically, its conservatively derived worst-case bound of 3.23×10⁻⁶ meets the IEC 61508 SIL-4 demand-mode target (<10⁻⁴) with an approximately 31-times margin; the nominal, common-cause-adjusted figure is 1.6875×10⁻⁷. The DSS probability tiers are intentionally stricter than IEC’s, which means we grade ourselves harder than the industry does. We could have published a table our worst case tops comfortably. We published one it does not, yet.

We claim the numerical demand-mode target only, and we say so explicitly in the papers: it is an analogy of measurement, not a claimed IEC certification. The methodology underneath, fault-tree analysis per IEC 61025 over 23 independent barriers with a justified common-cause factor, is domain-neutral mathematics. Our position is simple: if aviation and nuclear engineering earned quantitative reliability targets fifty years ago, the discipline that decides whether your chest pain is a pulled muscle or a dissection deserves the same grammar.

In motion, in public, and falsifiable at every step. All nine papers are public; the benchmark supplement (S1) publishes case-level inputs, outputs, and adjudications; submission to NEJM AI is underway alongside medRxiv preprints; a prospective confirmatory study on a frozen configuration is designated; and third-party audit is invited. Every figure we publish carries its development-cohort label in the same sentence, not in a footnote. One more thing: the system is still being calibrated, so we regard the current numbers as a floor, not a ceiling.

It would be, for a system that always answers. The published arc is 4.6% → 1.3% → 0.0% across 301 cases, and the mechanism is not omniscience, it is discipline: safety-first calibration plus LIMBO, the structured refusal to guess. In 4.0% of cases (12 of 301) the system declined to issue a diagnostic verdict and named what would settle the case instead. The price of zero was 2.3 additional seconds of median deliberation. Development-cohort result, confidence intervals published, confirmatory study designated.

On top-1 alone, at these sample sizes, we would not argue, and we publish our confidence intervals precisely so you can check. But the claim was never the half point. The claim is the combination: accuracy at the top of the field plus a certified worst-case safety bound plus a 0.0% missed-critical rate plus an audit trail to the last assertion plus an open standard anyone can verify against. Remove any one of those from a competitor and the comparison collapses. Nobody else brings all of them to the same table.

Public benchmark data already prices the alternatives. In Microsoft’s own SDBench study (Nori et al. 2025, the same NEJM CPC case family), reaching a diagnosis carried thousands of dollars of ordered tests per case: o3 averaged $7,850, Microsoft’s best MAI-DxO configuration reached 85.5% at $7,184, and physicians averaged $2,963. The protocol differs, SDBench orders tests sequentially, so we do not claim a same-protocol comparison. But the economics are the point: DeepSensi reaches 86.0% top-1 from the case record itself, in 12 seconds to a few minutes, and when evidence is genuinely missing it does not order a shotgun battery, LIMBO names the single test that settles it. The reasoning itself runs on commodity infrastructure at a computational cost that is a rounding error against one lab panel. Deployment pricing is set per pilot.

That is the first question we asked ourselves, and the defense is architectural, not chronological. The anti-contamination gate strips identifiers, authors, dates, and citations from every case before reasoning begins, forcing inference from pathophysiology rather than recall. The measured gap tells the story: naked single-model baselines on the same cases score roughly 39–49% top-1; the orchestrated consilium reaches 93.7% top-3. If memorization explained the result, the naked baselines would not be forty points behind. An input-grounding watchdog additionally verifies that conclusions trace to the presented case, and the full case-level audit is published in the S1 supplement.

Version 90 is in production on infrastructure spanning four continents, and access is deliberately gated: pilots and partnerships first, with LOIs in active discussion across health systems, payers, and public-health bodies. We chose to publish the proof, the standard, and the audit before the customer logos, not instead of them, because in medicine that order is the honest one. Announcements follow signatures, not the other way around.

The objection has history exactly backwards. Medicine’s breakthroughs have repeatedly come from one person standing against the institutional consensus: Semmelweis against childbed fever, Fleming noticing a single contaminated dish, Marshall drinking Helicobacter to prove the establishment wrong, and the web itself was one architect’s protocol before it became everyone’s industry. Committees ship consensus. Architects ship inventions. Being one mind is not this project’s handicap; it is the reason it exists. The audit trail here: one architect wrote the operating system, twelve pending patent applications with twenty-four more in preparation, nine papers, and an open certification standard, and shipped them operational, not as slideware. There is also a structural advantage a concern cannot buy: one mind holding the entire pipeline end to end can optimize it without committee, legacy, or vendor politics. The public benchmark economics show what unoptimized architectures burn: in Microsoft’s own study, thousands of dollars of cost per diagnosis (o3: $7,850; the best MAI-DxO configuration: $7,184 per case; Nori et al. 2025). DeepSensi’s architecture, governed by the Gomola Framework, the open certification standard that carries its architect’s name, was engineered by one mind that knew exactly where the waste lived: the same class of result, at a computational cost that rounds to zero against a single lab panel, without relaxing a single safety threshold. And the expansion underway is his design as well: clinical partnership leads and regional representatives joining on five continents, and a platform that turns thousands of physicians into paid co-authors. That is not a one-person company patching a weakness. That is an architect scaling his architecture, on purpose.

No. NeuroTokens are a transparent unit of account for cognitive work inside the platform: they meter computational cost and settle physician compensation to a thousandth of a cent. There is no public coin, no speculation, nothing to trade. And the line that matters most: economics never touch clinical judgment. Compensation is outcome-gated, audit-sealed, and engineered inside anti-kickback limits. Hypothesis staking lives in the research layer, not in the care of any patient.

The physician holds the signature, and with it the accountability, on every case. But we refuse to romanticize the loop: a signature is authority, not infallibility, and on the hardest public cases unaided physicians score 19.9%. So DeepSensi is built on a more honest contract: both minds are measured, both are auditable. The consilium can challenge the clinician on the record; the clinician can override the consilium on the record; GCPS scores the human’s performance as objectively as the fault tree bounds the machine’s; and liability is partitioned per decision, AI, physician, or rule, cryptographically sealed. The patient gets the benefit of the disagreement, and the business model built on that measurement, expertise as a measured, compensated asset rather than an assumption, is deliberately unlike anything else in medicine. Regulatory posture: HIPAA and GDPR compliant, EU AI Act aligned by architecture, FDA pre-submission (Q-Sub) engaged; deployment follows each jurisdiction’s clinical-software pathway.

Because a safety standard behind a paywall is not a standard, it is a product. The catch, if you want one, is strategic and we state it openly: adoption is the moat. Every hospital that writes a DSS level into procurement, every insurer that prices against a certified bound, every regulator that references the Gomola Framework makes the measuring stick more real, and DeepSensi is the reference implementation of that stick. Rivals are welcome to certify against it. We built the bar and we are pleased to compete on it.