Picks (5)
1. NVIDIA/skills — skills now have signatures, evaluation sets and benchmark reports
Repo: https://github.com/NVIDIA/skills
Overnight the repository regenerated metadata.json, skills.sh.json and the benchmarks, and unified its Physical AI components under CDF naming, shipping skills for CDF anomaly generation, data augmentation and auto-labelling. Skills stop being content and become artefacts with admission, provenance and regression requirements.
2. jcode — the real bottleneck in parallel agents is handoff, not start-up
Repo: https://github.com/1jehuang/jcode
Six commits between 00:53 and 01:17 all orbit /cloud: the TUI refuses to migrate an unsaved empty session to the cloud, all account-router credentials move with the migration, and the hosted-cloud preflight now requires router credentials. Shipping to the cloud is a chain of checks, not a button.
3. NotePilot v1.8 — the point of a clinical summary is being verifiable, not sounding right
Repo: https://github.com/calvinpatel/notepilot
A run of v1.8 commits, all about clinical semantic boundaries: "a dose carries its prn" (a dose must carry its as-needed condition), "an indication isn't a finding". These are the rules that decide whether a generated note survives review.
4. Neonatal imaging study — expert level decides whether AI helps or disturbs
Repo: https://github.com/junjslee/neonatal-ai-reliability ｜ npj Digital Medicine (2026), doi:10.1038/s41746-026-03399-3
Published open access. The key method is injecting errors into the AI's output to measure how the human reader reacts — a direct probe of automation bias rather than another accuracy number.
5. Open Health Sync v0.1.0 — the health data layer gets a self-hosted standard exit
Repo: https://github.com/soma-intelligence/open-health-sync
A sequence of commits moved it from "can sync" to "can be integrated": GET /v1/providers now states exactly what each data source can deliver, and all eight example recipes were rewritten. One dataset, four export paths, self-hosted.
Trends
- The skill ecosystem enters supply-chain governance. NVIDIA defines "verified skills" with signatures, evaluation sets and benchmark reports; AAS spent the same day fixing missing release identity, skipped source verification and false-ready admissions. Medical skill libraries will be the first asked to produce this evidence.
- Coding agents compete on resources and handoff, not capability. jcode runs dozens of agents in little memory and turns cloud migration into a checklist; t3code moves the control plane to a phone; magpie puts model routing in the menu bar. All three answer the same question: once there are many agents, where is the human-machine boundary.
- Clinical AI starts falsifying itself with reader studies and regression baselines. The neonatal paper injects errors to measure human response; NotePilot tests its own checking layer on a 24×3 synthetic corpus. Both state their limits openly — closer to shipping than leaderboard chasing.
- The health data layer converges on self-hosting plus a standard exit. open-health-sync offers one dataset with four exits, HealtOS got import and rollback right first, quantified-self-mcp first fixed "listens on all interfaces by default". The order is always trust first, intelligence second.
- Three new landing spots in nutrition: specialist agents (carefold adds a dietitian navigator), explainable scoring (eat.me.at's recipe nutrition score), and supply-chain hygiene (SuperMealPlanner cleaning up a leaked intermediate certificate). "Photograph it and estimate calories" is no longer the only story.
- Medical infrastructure experiments with Rust. FHIR servers (hfs) are rebuilding query backends for memory footprint and concurrency rather than language preference — worth tracking quarterly.