On 28 April 2026, the U.S. Food and Drug Administration announced two real-time clinical trial proofs-of-concept and a pilot Request for Information, framing the conduct of trials in which key safety signals no longer take years to reach the agency. The prevailing oncology AI patient-prediction baseline remains a heterogeneous set of narrow supervised models with ceilings such as Manz 2020 AUC 0.89, the SHIELD-RT prospective randomized trial, SCORPIO, and PROGPATH, set against the Huang 2025 null result that machine learning provides no significant gain over Cox regression on real-world structured survival data. This paper uses four author Physical AI oncology trial simulations - Simulation 1 in hour-00 through hour-55, Simulation 2 ten patient-journey stages, Simulation 3 a 24-hour autonomous sponsor, and Simulation 4 a 168-hour 7-day sponsor extension verified locally on a Core i5-6200U laptop with 4 GB RAM - to demonstrate that Claude Code Opus 4.7 Max produces working agentic code applicable to patient prediction better than supervised models in current trial practice and with more utility than the FDA RTCT proof-of-concept. The computational signature - 1M token contextual code-and-text awareness, hourly commit cadence, repository-scale forecasting - is the source of the advantage, without claiming clinical deployment readiness.
Kevin Kawchak (Mon,) studied this question.