DeepSensi publishes the first quantitative safety standard for clinical AI — free to the entire industry

Dover, Delaware — July 2026. DeepSensi PBC today released the Gomola Framework, the first quantitative safety-certification standard for clinical artificial intelligence, together with six technical papers headed to the medRxiv preprint server and peer review at NEJM AI. The framework is open and royalty-free.

Clinical AI is the only safety-critical technology deployed at the bedside without a quantitative reliability requirement. Aviation certifies against DO-178C; nuclear engineering against IEC 61508; medicines against Phase III trials. Generative clinical AI — whose dominant failure mode, hallucination, is fluent and clinically plausible — has had no equivalent. The Gomola Framework defines four graded certification levels by explicit per-assertion hallucination-probability bounds, organized around five independently auditable pillars.

Key results, from the published papers:

• A quantified safety bound. Fault tree analysis (the method used to certify power plants and avionics) of DeepSensi's 23-barrier verification architecture yields a worst-case bound of 3.23 × 10⁻⁶ undetected hallucinations per assertion — meeting the numerical target of safety-integrity level SIL-4 with an ≈31-fold margin.
• Elite diagnostic accuracy with zero missed criticals. On 301 NEJM Clinicopathological Conference cases (2014–2023), a full-architecture reference implementation reached 86.0% top-1 and 93.7% top-3 accuracy after safety-first calibration, with the missed-critical rate reduced from 4.6% to 0.0% at a cost of just 2.3 additional seconds of median deliberation (total median: 14.3 s; a full multi-specialist deliberation completes in 12 seconds to minutes for complex cases — a panel that takes days to convene by hand).
• Vendor independence. DeepSensi is not a wrapper on any single model; it cross-examines multiple models from independent vendors, so a better base model yields a better verified answer — the architecture consumes the progress of the model race rather than competing in it.
• A structured "I don't know." The LIMBO protocol produces an explicit uncertainty declaration — the disagreement, a working hypothesis, and the resolving tests — instead of a confident guess.

A full cognitive infrastructure. The newly published papers describe a system that extends beyond verification: the Hyper Consilium integrates physicians as mathematically scored cognitive partners (with royalties for verified knowledge, released only after patient outcomes confirm value); AutoResearcher autonomously generates and adversarially validates research hypotheses with a formal Bayesian convergence guarantee; and Golden Horizon provides borderless, free-by-architecture clinical-trial matching and autonomous patient advocacy for those who have exhausted standard-of-care.

A challenge to the industry's benchmarks. A companion paper, "The Flawed Yardstick," documents how popular static benchmarks systematically penalize safe clinical behavior — the "Safe Triage Paradox" — and proposes a dynamic, safety-first evaluation protocol.

The system is operational on edge nodes across four continents, including fully offline and air-gapped deployments. DeepSensi is engaging the U.S. FDA through the Q-Submission pathway and is architected for the EU AI Act ahead of its December 2027 deadline. DeepSensi PBC is a Public Benefit Corporation whose charter commits it to free global trial matching, open research into autism and refractory epilepsy, and the open safety standard.

"A safety standard behind a paywall is not a standard. It is a product," said founder and chief architect Tomasz Jan Gomoła. "We publish the proof, the standard, and the audit. We invite the world to try to refute them."

Technical documentation: www.deepsensi.com/papers Press and auditor access to research protocols: press@deepsensi.com

About DeepSensi — DeepSensi PBC is a Public Benefit Corporation building verified clinical intelligence: a Cognitive Medical OS that surrounds any language model with an auditable verification architecture. The Gomola Framework and DeepSensi Standard are open and royalty-free; the reference implementation is proprietary. Founder and chief architect: Tomasz Jan Gomoła (ORCID 0009-0001-5222-6154).