The Science
A safety bound you can audit,
not a benchmark you must believe.
DeepSensi's safety case is built the way aviation and nuclear engineering build theirs: with Fault Tree Analysis (IEC 61025) over a cascade of independent verification barriers — not with leaderboard scores. The full mathematics, per-barrier failure estimates, and independence justification are published in WP-001.
The verification cascade
Twenty-three independent barriers in three tiers. A hallucination reaches the physician only if all of them fail at once.
| Tier | Barriers | Nature |
|---|---|---|
| I — Input sanitization | 7 | Deterministic, pre-LLM: physiological coherence, document forensics, adversarial input, bias correction, complexity routing |
| II — Evidence verification | 9 | Provenance, citation resolution, retraction screening, citation-network analysis, cross-source triangulation, bias audit |
| III — Adversarial validation | 7 | Multi-round convergence, dual pathways, dedicated adversaries, red-team swarm, cross-domain checks, uncertainty failsafe |
Tier III validators are deliberately distributed across four independent model vendors — a systematic failure in any one vendor's family cannot compromise more than one barrier. This is why DeepSensi is not, and cannot be, a wrapper around any single LLM: models are interchangeable components under permanent cross-examination.
The results
| Quantity | Value | Meaning |
|---|---|---|
| Raw multiplicative bound | 7.7 × 10⁻²⁰ | Theoretical only — reported, never claimed |
| Nominal bound (CCF-adjusted, β = 0.1) | 1.69 × 10⁻⁷ | Common-cause failures conservatively included |
| Worst-case operational bound | 3.23 × 10⁻⁶ | APIs down, novel attacks, implementation defects — ≈1 per 309,600 assertions |
| IEC 61508 comparison | SIL-4 target, ≈31× margin | Demand-mode numerical target only¹ |
| Degradation floor | SIL-1 grade | Even fully offline, on the weakest quantized open-source model |
| vs. raw clinical LLMs (15–25%) | 46,000–1.48M× | Improvement range, worst-case to nominal |
¹ SIL classification under IEC 61508 additionally requires systematic-capability and lifecycle evidence; DeepSensi claims equivalence to the numerical probability target, and says so explicitly in WP-001. Values below the SIL-4 band floor exceed the scale's resolution.
The empirical program
Analytical bounds and clinical accuracy answer different questions, so we measure both. On N = 301 published NEJM Clinicopathological Conference cases (2014–2023) — medicine's hardest public diagnostic exam — the configuration-frozen system achieved 80.0% top-1 and 88.0% top-3 accuracy. After safety-first calibration it reached 86.0% top-1 (95% CI 82.1–89.9) and 93.7% top-3 (95% CI 91.0–96.4) — reported, deliberately, as a development-cohort result, with a frozen held-out replication registered as the confirmatory study. The missed-critical rate fell to 0.0% at a median deliberation time of 14.3 seconds. The case-level protocol is available to auditors, editors, and qualified press.
During calibration, the same evidence-integrity layers were pointed at public benchmarks themselves — and identified defective items in a widely used medical benchmark (HealthBench Hard), where deterministic safety gates are graded as failures: the Safe Triage Paradox. The full audit and the proposed replacement protocol (Auto-CSA) are published as WP-002, "The Flawed Yardstick". A verification architecture audits evaluations, not just outputs.
What failure looks like
When verification cannot establish confidence, DeepSensi does not guess. The LIMBO protocol outputs the core disagreement, a working hypothesis, and the specific tests that would resolve it — with a reliability score of zero. The failure mode is an explicit declaration, never a silent hallucination.