<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>Precision Insights</title><description>Sourced analysis of healthcare AI evaluation: landscape, evidence and gaps.</description><link>https://precision.example/</link><language>en</language><item><title>Who evaluates healthcare AI in 2026, and what do they miss?</title><link>https://precision.example/blog/who-evaluates-healthcare-ai-2026/</link><guid isPermaLink="true">https://precision.example/blog/who-evaluates-healthcare-ai-2026/</guid><description>Healthcare AI evaluation is split across governance platforms (Qualified Health), benchmark publishers (Vals AI, HealthBench), simulation environments (Alterra Labs), horizontal voice-testing tools and free validation software for predictive models. We found no dedicated, funded evaluator of deployed AI scribes, and no one measuring a product on a health system&apos;s own encounters with calibrated error bars.</description><pubDate>Sun, 15 Nov 2026 00:00:00 GMT</pubDate><category>Landscape and gaps</category><author>hello@precision.example (Precision research team)</author></item><item><title>Why do AI graders miss omissions in clinical notes?</title><link>https://precision.example/blog/why-llm-judges-miss-omissions-in-clinical-notes/</link><guid isPermaLink="true">https://precision.example/blog/why-llm-judges-miss-omissions-in-clinical-notes/</guid><description>In published tests, LLM judges discriminated complete from incomplete clinical notes at AUC 0.49 to 0.66, close to chance, and on omissions at 0.50 to 0.63 against 0.79 to 0.94 for added or altered content. A single call detected 36.9% of omissions at 6.2% false alarms. Both are unreviewed preprints, one with probable commercial ties.</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><category>Evaluation science</category><author>hello@precision.example (Precision research team)</author></item><item><title>Why can&apos;t you compare AI scribe error rates between studies?</title><link>https://precision.example/blog/scribe-error-rates-cannot-be-compared/</link><guid isPermaLink="true">https://precision.example/blog/scribe-error-rates-cannot-be-compared/</guid><description>Published scribe error rates range from 1.47% of sentences (CREOLA) to 18% of notes with omissions (UC Davis) to 31% of notes with hallucinations (Palm) to 31.3% of notes with any verified failure (a 565-note census). Changing only the review instruction moved the verified-error share of candidate errors from 9.3% to 79.0%, so no single scribe error rate is defensible.</description><pubDate>Fri, 13 Nov 2026 00:00:00 GMT</pubDate><category>Evaluation science</category><author>hello@precision.example (Precision research team)</author></item><item><title>Abridge, GAO and Ontario&apos;s auditor on evaluating AI scribes</title><link>https://precision.example/blog/what-abridge-gao-and-ontario-say-about-scribe-evaluation/</link><guid isPermaLink="true">https://precision.example/blog/what-abridge-gao-and-ontario-say-about-scribe-evaluation/</guid><description>Abridge publishes the most detailed vendor method: claims are labelled supported, unmentioned or contradicted and rated by severity, and omissions are out of scope. GAO&apos;s July 2026 report found few independent studies of AI note and coding tools. Ontario&apos;s auditor found 11 of 20 approved scribe vendors submitted no third-party audit.</description><pubDate>Thu, 12 Nov 2026 00:00:00 GMT</pubDate><category>Landscape and gaps</category><author>hello@precision.example (Precision research team)</author></item><item><title>Voice AI testing for healthcare: Coval, Hamming, Cekura, Bluejay</title><link>https://precision.example/blog/voice-ai-testing-for-healthcare-compared/</link><guid isPermaLink="true">https://precision.example/blog/voice-ai-testing-for-healthcare-compared/</guid><description>Coval, Hamming, Cekura and Bluejay test voice agents with simulated calls and monitoring. They are horizontal tools; healthcare is a use case. Coval raised a $28M Series A in June 2026 and lists plans from $100 a month to $4,500 and up. Published failure rates, such as Cekura&apos;s single Medicare scenario, are vendor-run.</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>Landscape and gaps</category><author>hello@precision.example (Precision research team)</author></item><item><title>Why chatbot benchmark scores miss real-user medical outcomes</title><link>https://precision.example/blog/benchmarks-do-not-predict-real-user-outcomes-oxford-trial/</link><guid isPermaLink="true">https://precision.example/blog/benchmarks-do-not-predict-real-user-outcomes-oxford-trial/</guid><description>In a Nature Medicine randomised trial of 1,298 people (published 9 February 2026), users of GPT-4o, Llama 3 and Command R+ identified relevant conditions in under 34.5% of cases and the correct action in under 44.2%, no better than a control group, even though the models scored well on medical benchmarks. The models were older, not current frontier.</description><pubDate>Tue, 10 Nov 2026 00:00:00 GMT</pubDate><category>Evaluation science</category><author>hello@precision.example (Precision research team)</author></item><item><title>Five gaps in healthcare AI evaluation that nobody fills well</title><link>https://precision.example/blog/healthcare-ai-evaluation-gaps-nobody-fills/</link><guid isPermaLink="true">https://precision.example/blog/healthcare-ai-evaluation-gaps-nobody-fills/</guid><description>Across vendors, benchmarks and health-system programmes, five gaps recur: graders that report their own error rates, causal diagnosis of which component failed, independent testing on a buyer&apos;s own encounters, continuous monitoring after deployment, and evaluation in languages other than English. Existing tools cover tracing, generic scores and benchmarks well.</description><pubDate>Mon, 09 Nov 2026 00:00:00 GMT</pubDate><category>Landscape and gaps</category><author>hello@precision.example (Precision research team)</author></item><item><title>What do health-system AI governance committees need from vendors?</title><link>https://precision.example/blog/what-health-system-ai-governance-committees-need/</link><guid isPermaLink="true">https://precision.example/blog/what-health-system-ai-governance-committees-need/</guid><description>Health systems say they want local validation, bias testing, PHI safeguards and a monitoring plan before onboarding a vendor, yet in one survey fewer than half have a dedicated testing environment and 63% call their approach developing or ad hoc. In an NHS poll, 83% treat clinical safety assurance as a procurement gate.</description><pubDate>Sun, 08 Nov 2026 00:00:00 GMT</pubDate><category>Building with evidence</category><author>hello@precision.example (Precision research team)</author></item></channel></rss>