Work

Fail-Closed Autonomous Research for Artificial Intelligence and Clinical Medicine

Year: 2026
Type: preprint
Abstract: Background. Autonomous AI can generate candidate designs and clinical hypotheses faster than physical measurements, evidentiary review and regulatory decisions can verify them. A common failure is to treat an absent observation as zero, an inferred state as measured, or an impressive model score as operational authorization. This study formulates a single research architecture for AI systems and clinical medicine without equating their domain-specific safety requirements. Methods. We define a typed physical-and-evidentiary state ledger, an internal three-valued proof logic with binary release, a history-dependent consequence state, and falsification criteria specified before independent deployment. We derive a fail-closed admissibility result and a capacity upper bound. Retrospective public-data analyses examine 463 labeled robot execution episodes from five UCI tasks and 400 ClinicalTrials.gov-derived trial-history anchors (6,029 recorded versions). In robotics, a fixed three-nearest-neighbor diagnostic benchmark compares final-sample measurements with 15-sample temporal summaries under task-stratified fivefold cross-validation. Results. Temporal summary features yield pooled balanced accuracy 0.9563 versus 0.7551 for final-sample features, yet miss 24 of 334 labeled failures (7.19%; Wilson 95% interval 4.88%–10.47%). A proposed independent-deployment rule requiring a one-sided 95% upper miss-rate bound below 1% is therefore not met. In clinical registry records, only 338 of 400 historical pre-results primary-completion entries are both marked ACTUAL and day-precise; 62 require quarantine for an exact-date historical audit. Median time from recorded primary completion to first results submission is 388 days in this selected corpus, not a treatment-effect or compliance estimate. Conclusions. The framework supplies an implementable separation between AI proposal, deterministic evidence verification and authorized release. The executed analyses demonstrate consequential temporal information and registry-evidence incompleteness; neither constitutes clinical efficacy, medical-device clearance, or live autonomous-system safety validation. Keywords: autonomous R&D; clinical trial provenance; evidence gating; robotics; formal verification; temporal memory; falsifiability. Proprietary components The analyses were produced by software held privately by the author. The code is not distributed with this record; its existence and exact content on the publication date are fixed by the SHA-256 fingerprints below. A fingerprint identifies a file but cannot be reversed into it. Source-pinned replication engine for the robotic-diagnostic and clinical-registry analyses (content-addressed inputs) Typed-state fail-closed verifier: physical-kind units, three-valued proofs, binary release, participant reconciliation Verifier test suite SHA-256 fingerprints: replicate_study.pyc01d33f165c0d4a1002179f8afb0a12a80149021967185f653e663c772bb01bd fail_closed_verifier.pya7131f8824f7a3dc89256047bdab98b6f7ebb1d6398011e11efbe6561c0ad650 test_verifier.py6c3e620668590d7d229abb73df7909199b4e074caefbec51fefd5cbdbfa03666 Proprietary notice. The code is proprietary to Jarid Shaub. Licensed parties receive it with these fingerprints for independent verification. Commercial use, implementation, evaluation and licensing inquiries: [email protected]
Author: Shaub, Jarid
Language: English
Open Access status: green
OpenAlex ID: W7236677345
Cites: 4
Cited by: 2
Related to: 10