This week: the UK government accepts every recommendation of its doctors-led AI Commission and opens AI Airlock phase 3, a neurosurgeon ships an aneurysm clipping simulator in 24 hours, and we count what clinicians actually built.
Top signal
The government said yes to all 44. On 6 October the UK government confirmed it will accept all 44 recommendations of the National Commission into the Regulation of AI in Healthcare (the blueprint this briefing covered at launch on 10 September), and set out how they will be delivered across the UK. Alongside the response, the MHRA opened applications for the third phase of AI Airlock, its regulatory sandbox for AI-enabled medical devices, focused on post-market surveillance and lifecycle regulation: the Commission's central conclusion, that AI devices should be assessed and monitored throughout their working life rather than relying on one-time assessments, put into practice immediately, with three years of funding secured and a prospective-applicant webinar on 22 October. Draft guidance on managing changes to AI-enabled devices is promised by December 2026, a consultation on qualifying and classifying them begins next year, and the response commits to exploring staged authorisation pathways that let promising AI tools into the NHS earlier under close supervision. A full implementation roadmap, with timelines and responsibilities across all 44 recommendations, follows by Spring 2027. GOV.UK
Shipped this week
- IC-PCom Clipping Simulator: Richard Dalyai, MD, a cerebrovascular and endovascular neurosurgeon and surgical director of stroke at ECU Health in Greenville, North Carolina, built a browser-based microsurgery simulator for internal carotid and posterior communicating artery aneurysm clipping with Claude Opus 5.5 in about 24 hours, live at his own site: look at the aneurysm, dissect the Sylvian fissure, deploy the clip, and see what happens. "I clipped the real aneurysm first," he posted on 1 October, "then I used Claude Opus 5.5 to build a prototype in 24 hours that lets you try it yourself." It is a personal teaching prototype, not an evaluated training tool, and we have seen no validation of it. X
- Digital Logbook: Karthik Deegutla, an anaesthesiologist and perioperative physician in India, built a free web app for logging cases, live at his own domain: "Log any case in 20 seconds. Your data stays on your device." "I'm a doctor, I built a Digital Logbook app for doctors", he posted on 24 September, "used Claude opus 5 model, 8 hours, 2 M output token. The app is free." It targets the paid logbook software juniors otherwise buy; it stores data locally, and we have seen no external review of its accuracy. X
Build safely
- Can vibe coding do the statistics?: a team led by Jeeva Sam, writing in Studies in Health Technology and Informatics (published 17 September), benchmarked five current LLMs (GPT 5.3, Claude Sonnet 4.6, Gemini 2.5 Flash, Perplexity and Grok) on reproducing expert-validated statistical analyses from a published pulmonary hypertension workflow across three tasks: descriptive tables, Kaplan-Meier survival analysis and Cox modelling. All five produced correct descriptive statistics once the dataset variables were specified explicitly; all reproduced the correct Kaplan-Meier p-value, though figure completeness differed; four reproduced every required hazard-ratio term while one omitted the interaction results. Their conclusion is the one to keep: outputs still require careful validation before use. Diego Forero, MD PhD, a clinician-researcher editor, surfaced the paper on X; the study itself is the source. IOS Press
- "Refined exclusion": when the aggregate hides the harm: Lee and colleagues, writing in the Journal of Medical Internet Research on 1 October, name a governance condition they call refined exclusion: an AI system that appears successful in aggregate while uncertainty, error and reduced reliability concentrate in populations that are under-represented, under-measured or under-monitored. The viewpoint connects model-level disparity, patient safety signals and documented patient harm, and proposes lifecycle decision gates, each linking minimum evidence to one of four actions: proceed, enrich or validate, restrict use, or stop. It is a viewpoint rather than a study, but the gates read as a usable checklist for anyone deploying locally, and it lands the same week the UK committed to lifecycle regulation of AI devices (see the top signal). JMIR
Events
- 6 Oct (London): Patient Safety Learning with Silver Buck convene a roundtable on what AI means for patient safety, aiming to produce practical guidance, a procurement checklist and an AI performance review checklist, with quarterly roundtables planned to follow. Eventbrite
New in the directory
- Arman Yalcin, MD (United States): an infectious disease physician in Boston whose bio reads "ID Physician | AI nerd", verified through his US News doctor listing, with ScrubBoard and the Salt Shaker Studio clinical study tools live on their own sites. ScrubBoard featured in Issue 8. GitHub
- The directory now counts 87 clinician-builders with verified, evidence-graded ships, its biggest week yet: 17 joined, no removals, 17 more than last week. Several are already familiar from these pages (Khalid Shamiyah, Shruti Malik and Sohrab Arora among them) formally joining now their entries are verified, alongside longer-standing builders such as Serdar Balci, whose ClinicoPath jamovi module has been maintained since 2019. Browse it, forward this to a colleague who ships, and reply if you know someone who belongs in it. cliniciansthatcode.com
From us
- What clinicians actually built, counted: we went back over the window from 1 August to 5 October and counted 108 artefacts clinicians shipped, 46 of them the Shipped items of Issues 1 to 9 of this briefing, each classified by hand on 32 columns. The shape of it: 23 teaching tools, 26 pieces of plumbing (developer tools, standards toolchains, research code) and 18 practice and ward systems; 55 were built for the builder's own colleagues; 59 have no AI inside them at all; 97 of the 108 were built solo. Of the 20 that handle identifiable patient data, 19 are silent on assurance. The essay carries its caveats (English-language sources, self-reported ships, one editor's selection) and closes with an invitation: reply with anything shipped in the window that the 108 missed. It opens the new Essays hub on the site, and its own social posts follow tomorrow. cliniciansthatcode.com