Everything a newsroom needs.

1. One line

DeepSensi is the Cognitive Medical OS — a verified clinical-intelligence system that gives medical AI the one thing it has always lacked: a quantitative, engineered bound on its most dangerous flaw.

2. Boilerplate — "About DeepSensi" (copy-paste ready)

Short (35 words). DeepSensi is a verified clinical-intelligence system — the Cognitive Medical OS — built around a quantitative, fault-tree-derived safety bound on AI hallucination. It is developed by DeepSensi PBC, a public benefit corporation in Dover, Delaware.

Long (~110 words). DeepSensi is a verified clinical-intelligence system — a "Cognitive Medical OS" — that treats AI hallucination as a reliability-engineering problem rather than a prompt problem. Its 23-barrier verification architecture carries a published worst-case bound on undetected clinical error of 3.23 × 10⁻⁶ per assertion, derived with the fault-tree methods used to certify aircraft and nuclear plants, and validated on 301 of the hardest published diagnostic cases in medicine. It is the first Cognitive Medical OS — vendor-agnostic and fully auditable — built to empower clinicians rather than replace them, compressing time-to-diagnosis and giving physicians a defensible, court-grade record of every decision. DeepSensi is developed by DeepSensi PBC, a Delaware public benefit corporation founded by Tomasz Jan Gomoła, author of the open, royalty-free DeepSensi Standard.

3. Fact sheet — canonical numbers

Safety (analytically derived). - Worst-case bound: 3.23 × 10⁻⁶ undetected hallucinations per clinical assertion (≈ 1 in 309,600). - Nominal bound: 1.69 × 10⁻⁷ per assertion under common-cause (β-factor) adjustment. - Method: Fault Tree Analysis per IEC 61025, over 23 independent verification barriers in three tiers. - The worst-case bound meets the numerical demand-mode SIL-4 target with an ≈31× margin (numerical-target equivalence only; the bound is an analytical quantity).

Accuracy (empirical, development-cohort). - Validation set: 301 published NEJM Clinicopathological Conference cases (2014–2023) — among the hardest diagnostic cases in medicine. - 86.0% top-1 (95% CI 82.1–89.9) · 93.7% top-3 (95% CI 91.0–96.4) after safety-first calibration. - Frozen baseline before calibration: 80.0% top-1 / 88.0% top-3. - Missed-critical rate: 4.6% → 0.0%, at a cost of just 2.3 additional seconds of median deliberation. - Median deliberation time: 14.3 s. Reported as a development-cohort result; a confirmatory study is designated.

Speed. - A full multi-specialist deliberation returns in 12 seconds to minutes for complex cases — versus the days it takes to convene an equivalent human panel by hand.

The standard. - The DeepSensi Standard (DSS), also called the Gomola Framework: the first quantitative safety-certification standard for clinical AI. Five auditable pillars, four certification levels with explicit per-assertion hallucination bounds. Open and royalty-free (attribution required). The reference implementation is proprietary.

Company & compliance. - DeepSensi PBC, a public benefit corporation. 8 The Green STE A, Dover, DE 19901, USA. - Zero PII by architecture · HIPAA · GDPR · EU AI Act (architected to Articles 9–17) · FDA Q-Submission dialogue engaged · HL7 · FHIR · DICOM. - Intellectual property: 12 patent applications filed, with a further 24 prepared — the majority designated for open publication as technical whitepapers. - Runs in the cloud or entirely offline, air-gapped if required, from a $500 edge node to a departmental server.

4. The story in one paragraph

Every safety-critical industry measures reliability before deployment: aircraft engines, nuclear reactors, and new drugs must all clear a quantitative bar. Clinical AI — a technology whose errors are counted in human lives — has been required to clear none. DeepSensi closes that gap. It applies aviation's fault-tree mathematics to a 23-barrier verification architecture, producing the first published, quantitative bound on medical-AI hallucination; validates it on the hardest public diagnostic cases in medicine, cutting the missed-critical rate to zero; and publishes the certification standard behind it, open and royalty-free, for the whole industry to adopt or attack.

4a. The benchmark, in context

Exact-diagnosis accuracy on the publicly available NEJM Clinicopathological Conference (CPC) cases — among the hardest diagnostic cases in medicine, the same benchmark used across the field:

System	Exact-diagnosis accuracy	Source
DeepSensi — Cognitive Medical OS	86.0% (93.7% top-3)	301 cases, 2014–2023, development-cohort
Microsoft MAI-DxO (o3)	85.5%	Nori et al. 2025 (SDBench, 304 cases)
OpenAI o3	78.6%	Nori et al. 2025
OpenAI GPT-4o	49.3%	Nori et al. 2025
OpenAI GPT-4	39%	Kanjee et al. 2023 (70 cases)
Practicing physicians (unaided)	19.9%	Nori et al. 2025 (21 US/UK physicians)

Others guess the answer. DeepSensi deduces it — a patent-pending, SIL-4-grade anti-hallucination architecture, auditable to every step. A score you cannot audit is still a guess.

4b. What only DeepSensi brings

Capability	DeepSensi	Frontier LLMs*	Orchestrated AI	Ambient scribe
SIL-4-grade anti-hallucination architecture	3.23×10⁻⁶	✗	✗	✗
Open safety-certification standard (DSS)	✓	✗	✗	✗
Deterministic multi-layer verification	✓	✗	partial	✗
Verified deductive chain — auditable, not a guess	✓	✗	✗	✗
Structured "I don't know" (LIMBO)	✓	✗	✗	—
Anti-bias adversarial design	✓	✗	✗	✗
Court-grade audit + liability split	✓	✗	✗	partial
Physician score + outcome-gated royalties (GCPS)	✓	✗	✗	✗
NEJM-CPC exact-diagnosis accuracy	86.0%	39–79%	80–85.5%	n/a

*Frontier LLMs = general-purpose and medical models from OpenAI, Anthropic, Google, xAI, DeepSeek and others. Categories, not a rating of any single product; "✗" = none published as of July 2026. DeepSensi's DSS certificate reads two-dimensionally: Platinum architecture · certified worst-case bound 3.23×10⁻⁶ (SIL-4 grade).

4c. The human errors that vanish — a panel with no ego

No fatigue, no hierarchy, no memory of a "difficult patient," and a mandatory adversary in every case. The classic human diagnostic biases DeepSensi's architecture is designed to eliminate: anchoring, confirmation, premature closure, availability, base-rate neglect, search satisficing, framing, ascertainment/stereotyping, overconfidence, authority gradient, groupthink, and fatigue. Machine bias is engineered against through cross-vendor independence and deterministic evidence checks.

Benchmark sources: Nori et al., Sequential Diagnosis with Language Models, arXiv:2506.22405 (2025); Kanjee et al., JAMA (2023). DeepSensi figures are development-cohort on 301 NEJM CPC cases; a confirmatory study is designated.

5. Key messages

1. A number, not a vibe. DeepSensi gives clinical AI a quantitative safety bound — the way aviation and nuclear engineering have measured reliability for decades.
2. Empowerment, not replacement. DeepSensi makes one clinician stronger: it compresses time-to-diagnosis from days to minutes and hands the physician a defensible, court-grade record of exactly what was checked. The clinician signs, and stays in command.
3. Safety and accuracy are not in tension. Safety-first calibration cut the missed-critical rate from 4.6% to 0.0% at a cost of 2.3 additional seconds.
4. The proof is public; the blueprint is not. The papers, the standard, and the audit trail are open. The reference implementation stays proprietary.
5. Honesty is a feature. The system is built to say "I don't know" — and to name the single test that would resolve the doubt.
6. A standard, not just a product. The DeepSensi Standard is open and royalty-free, so any vendor, regulator, or health system can use it.

5a. The physician, empowered — not replaced

DeepSensi is built to make one clinician stronger, not to stand in for a department.

• Faster to a verified answer. A deliberation that takes days to convene as a human panel returns in 12 seconds to minutes for complex cases — time-to-diagnosis collapses from days to minutes.
• A defensible record — a first. For the first time, the physician holds a court-grade, cryptographically sealed record of exactly what was checked, by whom, and why, with liability partitioned per decision (AI, physician, or rule). The decision they sign is documented to a standard no chart has offered before.
• The clinician stays in command. DeepSensi is clinical decision support; the treating physician signs every decision. It augments judgment — it does not override it.
• The first Cognitive Medical OS. Diagnosis, documentation, longitudinal patient memory, and a physician reputation-and-royalty economy on one verified substrate.
• Vendor-agnostic and auditable. Not a wrapper on any single model — a better base model yields a better verified answer — and every step is independently auditable, including against the operator.

The net is empowerment with a safety net: the clinician gains the depth of a full specialist consilium and a record that protects the decision, while remaining, unambiguously, the person in charge.

6. Approved quotes — attributable to Tomasz Jan Gomoła, Founder

"Aviation doesn't ask a jet engine to feel safe. It asks for a number. We gave clinical AI its number."

"We published the standard and kept the implementation. A category belongs to whoever writes its rules — so we wrote them, and gave them away."

"The most dangerous sentence in medicine is a confident mistake. We built a system that is allowed to say 'I don't know' — and rewarded for it."

"Every knowledge profession pays its best minds for the reuse of their best ideas. Medicine was simply last, because its attribution problem was the hardest. It isn't anymore."

"We report development-cohort results and name the confirmatory study we still owe. The proof travels with the claim. That is the whole point."

"We didn't build an AI to overrule the physician. We built one to give a single clinician the depth of a full specialist panel in minutes — and, for the first time, a record that proves exactly what was checked."

7. Founder bio

Short. Tomasz Jan Gomoła is the founder and chief architect of DeepSensi PBC and the author of the DeepSensi Standard (the Gomola Framework), an open safety-certification standard for clinical AI. He designed DeepSensi's 23-barrier verification architecture and its quantitative, fault-tree-derived safety bound. ORCID 0009-0001-5222-6154.

Long. Tomasz Jan Gomoła is the founder and chief architect of DeepSensi PBC, a Delaware public benefit corporation building verified clinical intelligence. He is the author of the DeepSensi Standard — also known as the Gomola Framework — the first quantitative safety-certification standard for clinical AI, which he has published open and royalty-free. He architected the system's 23-barrier deterministic verification cascade and the fault-tree reliability analysis that produced its published per-assertion safety bound, and is the corresponding author on the associated body of technical papers (preprints on medRxiv; manuscripts in submission to peer-reviewed venues). His work is protected by twelve filed patent applications, with a further twenty-four prepared, most of them designated for open publication. ORCID 0009-0001-5222-6154.

8. Capabilities glossary (plain-language, public names)

• Hyper Consilium — a multi-specialist AI deliberation that reasons like a hospital tumour board, in seconds, with physicians participating as scored contributors.
• LIMBO — the structured "I don't know": when evidence is insufficient, the system declares uncertainty instead of guessing.
• Active Sensing — recommends the single examination or test that would most decisively resolve an ambiguous case.
• GCPS (Global Clinical Performance Score) — an objective, bias-resistant measure of physician performance; validated insight earns outcome-gated royalties.
• Golden Horizon — clinical-trial matching and patient advocacy, free for patients by architectural mandate.
• GERN — a privacy-preserving syndromic-surveillance layer: an early-warning signal derived from de-identified clinical documentation (methods only; no detection results are claimed).
• Retroactive alerting / digital twin — when new medical knowledge changes a past patient's picture, the treating physician is alerted automatically, even years later.

9. The papers (nine)

Preprints headed to medRxiv; manuscripts in submission to peer-reviewed venues (including NEJM AI). Full texts: www.deepsensi.com/papers.

• WP-001 — Fault Tree Analysis of multi-layer verification (the safety case)
• WP-002 — The Flawed Yardstick (why static benchmarks penalize safety; the Safe Triage Paradox)
• WP-003 — The Hyper Consilium (physicians as scored nodes; LIMBO; outcome-gated royalties)
• WP-004 — The Global Clinical Performance Score (objective physician reputation)
• WP-005 — AutoResearcher-HYPO (autonomous hypothesis generation; no biomedical findings claimed)
• WP-006 — Golden Horizon (compassionate access, free for patients)
• WP-007 — GERN (privacy-preserving syndromic surveillance; methods only)
• DSS-001 — The DeepSensi Standard (the open, royalty-free certification standard)
• NEJM AI Perspective — why medical AI needs quantitative safety certification

10. What we do NOT claim (please hold us to this)

• We do not claim FDA clearance. DeepSensi is decision support; the treating physician signs every decision. A pre-submission (Q-Submission) dialogue with the FDA is engaged.
• We do not claim the accuracy figures are peer-reviewed or final. They are development-cohort results; a confirmatory study is designated; the case-level protocol is open to auditors.
• We do not claim to have empirically verified a probability of order 10⁻⁷–10⁻⁹. The bound is analytical; accuracy studies validate deployability, not the probability itself. Certification is deliberately two-dimensional (architecture · probability).
• We do not claim GERN detects outbreaks today. WP-007 is a methods paper with a pre-registered, designated prospective evaluation.
• We do not replace physicians. DeepSensi is clinical decision support: the treating clinician signs every decision and stays in command. It empowers judgment — faster answers, a defensible record — it does not override it.

11. FAQ

Is DeepSensi FDA-approved? No. It is clinical decision support; the physician signs. A Q-Submission dialogue with the FDA on quantitative certification of verification architectures is engaged.

Are the accuracy numbers peer-reviewed? Preprints are being deposited on medRxiv and manuscripts are in submission to peer-reviewed venues, including NEJM AI. The N = 301 figures are development-cohort estimates with a designated confirmatory study; auditors may request the case-level protocol.

Does "SIL-4" mean it is certified like a nuclear plant? It means the worst-case per-assertion bound meets the numerical demand-mode SIL-4 target with roughly a 31× margin. It is a numerical-target equivalence, not a third-party SIL certificate.

Does it replace doctors? No — it empowers them. DeepSensi compresses time-to-diagnosis and gives the clinician a court-grade, auditable record — with liability partitioned per decision — of exactly what was checked and why. It is clinical decision support: the physician signs and remains in command.

Which AI models does it use? DeepSensi is model-agnostic and cross-vendor by design; it is not a wrapper on any single model, and a better base model yields a better verified answer. Specific vendors are not disclosed.

Is the standard really free? Yes — the DeepSensi Standard is open and royalty-free (attribution required). Only the reference implementation is proprietary.

How is patient privacy handled? Zero PII by architecture; k ≥ 5 anonymity for cohort work; HIPAA- and GDPR-aligned. Identity is never part of the analysis.

Where is the company based? DeepSensi PBC, a Delaware public benefit corporation, 8 The Green STE A, Dover, DE 19901, USA.

12. Assets in this kit

See images/ for logo, icon, and social-card files, and images/README_IMAGES.txt for exact filenames, dimensions, alt text, and usage. A founder photograph is provided separately on request ([email protected]). Feature articles and the press release are in articles/ (Markdown) and in ready-to-send/ as Word (.docx) and PDF, ready to attach.

13. Usage, trademark & licensing

• Text in this kit (boilerplate, quotes, fact sheet, bios) may be reproduced freely in editorial coverage.
• The DeepSensi Standard / Gomola Framework is open and royalty-free; attribution required. The reference implementation (DeepSensi Medical OS) is proprietary.
• "DeepSensi" and the DeepSensi mark are trademarks of DeepSensi PBC; please do not alter the mark's proportions or colours, or imply endorsement or partnership without written agreement.
• Numbers are canonical — please reproduce them as written, with the development-cohort qualifier where relevant.

14. Contact & embargo policy

• Press, interviews, auditor access: [email protected]
• Partnerships & institutions: [email protected]
• Author / corresponding (papers): [email protected]
• Materials shared "under embargo" may not be published before the stated date. Once preprints are live on medRxiv, DOIs will be circulated and the embargo lifts.

---

DeepSensi PBC is a Public Benefit Corporation. Free clinical-trial matching for patients, open research, and the open DSS safety standard are written into its charter. We publish the proof, the standard, and the audit. We do not publish the blueprint.