Autonomous, Fail-Closed Design of a Clinical Trial Enrollment Network from Live Registry Evidence: Phase 2 Atopic Dermatitis
An AI system designed the site network for a Phase 2 trial from live public evidence, refused to guess where the evidence was missing, and proved by simulation that the plan beats the industry-standard estimate. Objective: enroll 240 participants with moderate-to-severe atopic dermatitis in the United States within 18 months of the first site opening, with 90 % confidence. GENESIS queried ClinicalTrials.gov live and admitted only the 112 of 305 completed trials whose enrollment and primary completion were recorded as actual, quarantining the rest with the reason kept, never imputing. It learned real site-start-up times from PubMed itself (five verified statements from three papers, median 6.3 months), tested protocol strictness against the registered eligibility criteria of 112 trials, fitted a robust Poisson-lognormal recruitment model and searched the network by trial and error. The naive mean-rate plan of 33 sites finishes on time in only 48 % of simulated trials; the final network of 40 named institutions, each with a registry track record in the condition and ranked by current competing trials, finishes on time in 91 %. 876 generic sponsor placeholders were scrapped: every site is a real, contactable institution. JURIS re-derived every number by an independent route and re-found every literature quote verbatim (12 of 12 checks). No efficacy, safety or regulatory claim is made. What GENESIS is GENESIS is an autonomous R&D system. A plain-English objective goes in; out comes a verified design, real drawings, 3D models, a 13-section build packet, an official report and an academic paper, in minutes, on a laptop. One intelligence interprets the objective, plans the research, selects and sizes architectures across fields, learns from its own failures and closes the objective; JURIS, an independent verifier, re-derives every relation and every number and decides nothing. Every outside number is pulled live from a public source and carries its retrieval time and SHA-256 fingerprint. Architectures that need a part nobody sells are scrapped before design: GENESIS builds only from parts on sale today. Every report shows the trial and error that led to the final product. The companion record contains the full reports and build packets (site plan, enrollment forecast, timeline, risk register, regulatory checklist, full admission ledger, data provenance and integrity hashes) for this and a Phase 3 heart-failure objective. Professional review: not yet performed. Values are computed predictions from cited live data and declared assumptions, independently re-derived by JURIS; they are not measurements. Proprietary components Everything in this record was produced autonomously by GENESIS, software held privately by the author. The code is not distributed with this record; the GENESIS release that generated and stamped these documents is fixed by the SHA-256 fingerprints below. A fingerprint identifies a file but cannot be reversed into it. GENESIS campaign engine (objective to closure, trial-and-error search, scrap rule)c0df355cf20516dade7a37a1be0991fc13a3839d7176a7983a77c4fe14f7c05d GENESIS decision core (architecture selection, memory, learned lessons)2238214c83de212eaa1db3dee9cb38b7c6802ff3a0a845837f94f8b5f4d45f30 JURIS independent verifiere1ba5881e3a31d32e23d36f35cde3b95f399aa39ac046592a586fc7c21a33197 Coupled physics solver and optimiser1c98cde1eee37aa85b121ca9e6830374c72e106c82a6c72c08b0a28ac9b503ae Cooling domain (heat transfer, hydraulics, rack catalogue, annual operation)a0a5d4e6d29fb629784820fd7191cb9b2d3c6de62ca20701350ecaf967b30812 Clinical domain (fail-closed registry admission, recruitment model, site network)6f830f064f86d17cdedbdd6db22431a73aa5cbee602146a9a68edc38cb48a798 Live data feeds with provenance88456b2f03cb8a2d3de0444c921224ee85464b6cf52d1f513b17f11428f12296 Evidence engine (PubMed, datasheets, verbatim-quote verification)9d1312a0fc1465ef904852ac86f7da6fde2153a05a500f0b882668460cab98dd Real-parts selection and transientsd4278ad8e1136b00663cf012c186796fd4905b13afa87fc0a9c4fae6fdaf1733 Build-packet generator592281e3a1029203bcc5b535117fe2a7233f5a08ffe9cca2c73a90a109d96297 Technical-data-package writer7077c6df49503e371dc49297bc9d44d0c11db72074d13d2c54f6debf8083e8aa Official report generatora397259611a7d04fb2e6259609d33fe4fd02d8e826cd4fc0a17cfd88bde9e04e Academic paper generator4d3179eb1b80a950c9a3cc086db6989eb638089b5101b83213efd36123ceeaaf CAD export (general arrangement, DWG via BricsCAD)c3b3f5919188f8f056279579e815e66d5605def0aedded3a259ebfc046e13112 CAD schematics (P&ID, electrical one-line)3be2cad0cdf74cc87a3aa5a622ea377c451c4d53838cc2090c56d1c9beacc2cc Proprietary notice. The GENESIS software (source code, solver and verifier) is proprietary to Jarid Shaub and is available only by licence or acquisition. The documents in this record are licensed CC BY-NC-ND 4.0. Commercial use, implementation, evaluation and licensing inquiries: [email protected] Acknowledgement The governance architecture of the complete GENESIS model was designed by C. J. (Chloe Jane) Tully (ORCID 0009-0007-5661-7332, TTG-41/42 Cryptosystems): the Queen protocol (the final authority gate), Lattice (structural intelligence and admission authority) and TIT GIT (the governance layer). These components are not active in this demonstration release; they are designated for implementation in the complete, final version of GENESIS.
Authors
- Jarid Shaub Shaub
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24379519
- Primary Topic
- Statistical Methods in Clinical Trials
- Type
- preprint