1. :books: _OpenEvidence split its answer engine into three speeds — 5 seconds, 30 seconds, 5 minutes — and locked the fourth model away over bioweapons risk
September 8, 2026 · 4 items
1. :books: _OpenEvidence split its answer engine into three speeds — 5 seconds, 30 seconds, 5 minutes — and locked the fourth model away over bioweapons risk
Healthcare IT News · Nathan Eddy · September 8, 2026Practice operations
Four models, three of them live now and free to verified clinicians on the OpenEvidence site and its iOS/Android apps. Osler answers in ~5 seconds and is built for questions that come up during an encounter — it replaces the model that previously generated all OpenEvidence answers and becomes the default. Sackett takes ~30 seconds, is for questions needing a deeper read of the evidence, and can ask you for more clinical context before it weighs the research. Snow (successor to Deep Consult) takes ~5 minutes and examines multiple lines of inquiry across the literature before producing a report.
The company's stated position is that all three production models are held to the same clinical-accuracy standards — the difference is response time and depth of literature search, not correctness. Worth reading that claim carefully rather than assuming the 5-minute answer is the "better" one.
Darwin, the fourth, is in limited research preview for institutional partners, research collaborators and accredited academic AI researchers only. Reported scores: 100% on MedQA (the licensing-exam benchmark), 72.8% MedXpertQA, 82.7% HealthBench Professional, 87.2% NOHARM — all self-reported by OpenEvidence, and it says it is publishing Darwin's responses and explanations for each answer.
:triangularflagonpost: The restriction is the story, not the score. OpenEvidence says access to Darwin is limited because of dual-use risk from advanced medical and scientific reasoning, naming virology, bioweapons-related research and human germline editing. First user named is the National Organization for Rare Disorders, for complex cases. Context: the company raised $250M at a $12B valuation in January, ~$700M total.
Why it matters for an independent practice: Your physicians are already using this — it's free, and it's the tool a doctor opens between patients. What changed today is that the answer they get now depends on which model they happened to be in, and the fast default is the one that reads least. That is a new, invisible source of variation inside your own practice: two clinicians, same question, different depth of evidence, no indication in the output which mode produced it. Two things worth doing. One: if anyone in the practice cites an AI answer in a chart note or a patient conversation, the standard should be that they can name the underlying study — the 5-minute model exists precisely because the 5-second one isn't doing that work. Two, the marketing read: patients are running the same questions through consumer chatbots with no accuracy standard at all. The differentiator isn't having AI. It's being the practice that can say "here's the paper, here's why it doesn't quite apply to you." Surfaced on Healthcare IT News (Trusted); the article's canonical home is MobiHealthNews (Candidate, 0/1) — same HIMSS Media desk. Vote reads for both. · Read it
2. :dna: _Six aging clocks say an AI-designed drug made patients look ~3 years younger. Only 21 of 54 comparisons cleared the study's own threshold — and the authors say it plainly.
The Rundown AI · Zach Mink · September 8, 2026Practice operations
Insilico Medicine published an exploratory analysis on rentosertib, a drug whose target protein (TNIK) was identified with AI and whose molecule was AI-designed. Researchers applied six independently developed proteomic aging clocks to blood from 42 participants in an earlier Phase IIa trial. All six predicted a younger biological profile in treated patients. At week four, four clocks estimated reductions of 2.71 to 3.46 years at 60 mg once daily; the 30 mg twice-daily arm produced broader agreement across clocks. Age-associated protein patterns were compared against 55,319 UK Biobank participants as an external reference.
The caveats are in the source, not added by me. Those are changes in model estimates — they do not demonstrate patients gained years of healthy life. A fibrosis-associated protein, LTBP2, influences all six clocks, so the agreement between them "could partly reflect a shared response to the lung disease." Statistical support was uneven: 21 of 54 comparisons met the reported threshold, with weaker significance by week 12. Insilico's founder and employees are among the coauthors.
The underlying trial is the real medical anchor: 71 patients, 21 sites in China, July 2023–June 2024, 12 weeks of treatment, published in Nature Medicine June 2025. At 60 mg once daily mean forced vital capacity rose 98.4 mL while placebo fell 20.3 mL. Sixteen participants discontinued, with liver toxicity and diarrhea among the leading reasons.
Base rate worth keeping: a 2024 analysis of AI-native biotech pipelines found ~40% phase II success — comparable to historical industry averages, on a small sample. AI-designed does not yet mean likelier to work.
Why it matters for an independent practice: This is the cleanest version of the argument a regen practice has to win every week. A "biological age" number is a model's estimate, and the model can move for reasons that have nothing to do with the patient getting younger — here, quite possibly because their lung disease responded. Every longevity panel, every epigenetic-age test, every "your biological age dropped 4 years" report sold in this space has exactly this problem, and almost nobody selling one says so. Two applications. One: if you offer or refer biological-age testing, the honest framing is "a proxy that moved," not "you got younger" — and the practice that says that out loud is the one that survives the first skeptical patient who reads the paper. Two: note what the authors propose — building aging measurements into disease trials from the start. That's the direction the evidence base for this field will actually come from, and it's worth watching more than any single clinic's outcome data. The Rundown AI (Trusted) · Read it · :microscope: Primary: Integration of proteomic aging clocks in a phase 2a clinical trial supports simultaneous geroprotective assessment, Nature Biotechnology, Sep 7, 2026
3. :round_pushpin: _A recovered Google binary shows your practice isn't a listing — it's an entity Google assembles from 793 data providers, and your Business Profile edit is only one vote in that
Search Engine Land · Olivier de Segonzac (RESONEO) · September 8, 2026Patient acquisition
RESONEO obtained a binary exposing a non-public scope of Geostore, Google's internal system for representing geographic entities, and cross-referenced it with Maps protocols, network traffic, the web index, on-device components and the 2024 Google leak. Recovered: 72 Geostore ranking signals, 793 data source providers, 446 local search intent types, 50,998 Mapcore styles, 10,936 searchable Geostore declarations. The full study and a searchable archive are public.
The finding that changes how you work. A business's name may come from one provider, its phone from another, its category from a third, its geometry from a fourth. Google runs conflation — mechanisms to pick, merge or combine values when sources disagree — with trust levels ranging from blocked/untrusted to trusted/super-trusted. In the author's words: changing a field in Google Business Profile "doesn't guarantee that Google's canonical representation immediately becomes that new value. The edit becomes another piece of evidence entering a system that may already have competing evidence." That is the explanation for every attribute that keeps reverting.
The ranking layer is called Oyster Rank. Named signals include Google reviews, web query volume, listing impressions, listing opens, direction requests, website clicks, chain membership, Wikipedia signals, popularity, prominence. 25 of the 72 are explicitly marked deprecated, and the author is emphatic that they recovered signal names, not weights — and that Oyster Rank is not "the Maps algorithm," since a query still passes through query understanding, semantic matching, candidate generation, geography and reranking. There is also a separate scorer running offline on the phone with 8 signals across 13 tiers.
Two measured results worth the whole piece. Radius is not fixed: from the same Paris origin, a dense query (`pharmacie`) produced a far smaller search area than a brand query (`Carrefour`), and the same pharmacy query expanded dramatically in a rural area. And geography changes what gets considered, not just the order: across 5,083 calls and 86,584 results, removing geographic weighting moved median distance from 6.87 km to over 4,000 km — while the non-geographic ordering stayed extremely stable. Separately, web pages carry Knowledge Graph machine IDs through a layer called webref, storing topicality and confidence — meaning your location page is evidence about the entity, not just a page competing for a query.
Why it matters for an independent practice: Local search is patient acquisition for an independent practice, and this reframes the whole job. Three things. One: stop treating a stubborn wrong address, phone number or category as a GBP bug and start treating it as a sourcing problem — find the third-party directories, data aggregators and old profiles still asserting the wrong value, because those are the competing evidence. Two: multi-location practices have a named failure here — the author observed stores of the same brand in the same metro carrying different primary concepts despite a canonical chain model. Consistency across your locations is not automatic, and nobody checks it. Three, the one that matters most: `direction requests` and `listing opens` sit in the ranking vocabulary. Behavior on the listing is a signal, which means the photos, hours and answered phone that convert a patient are the same things that earn the ranking. There is no separate "SEO task" — and with Gemini now sitting on top of this stack, the practice Google can describe confidently is the one it will recommend in a conversation. Search Engine Land (Candidate — eleventh supplied item, still zero votes) · Read it
4. :movie_camera: _Video — "AI agents need to be kept in check in healthcare settings"
HIMSS TV via Healthcare IT News · published Sep 8, 2026, 9:29 AM (verified from the page's own `published_time`) · September 8, 2026Practice operations
Not watched or transcribed — stated up front, as always. Verified: the source (Trusted), the publish timestamp, the specific watch URL, the speaker, and the topics HIMSS filed it under — Artificial Intelligence, Clinical, Privacy & Security, Quality and Safety, Workflow, Workforce. Everything below the bullets is my read, not the video's claim.
The publisher's own framing: "Organizations need safeguards in place as agents become more autonomous. Ankit Jain, cofounder and CEO of Infinitus, discusses ways to prevent unintended or unauthorized AI actions from affecting patient care."
:warning: Honest limitation: this is a short HIMSS TV interview clip and the page carries no runtime, no transcript and no abstract beyond that one sentence. It is thinner on stated specifics than any item above it, and I'd rather say that than pad it. It runs because it is a date-verified Trusted-source video on the exact question the rest of today's digest circles, and because the alternative was another day with an empty video slot.
Why this one and not a builder channel. No automation/builder video could be date-verified as days-old this morning through the tools available — and the standing rule is that an unverifiable upload date is a hold, not a guess. So the slot went to the fresh Trusted video that exists rather than a stale one that doesn't. Nate Herk sits at (0/1) and Nate B Jones at (0/0); neither ran today.
Why it matters for an independent practice: "Unintended or unauthorized AI actions" is an enterprise phrase for a small-practice problem. In a health system there's an IT function that notices when an agent starts doing something odd; in a nine-person clinic the agent is the process. The concrete version: an AI receptionist or follow-up agent that can reschedule, cancel, message a patient or touch the chart is taking actions nobody re-reads. Three questions worth asking any vendor pitching one this month — what actions can it take without a human clicking approve; where is the log of what it actually did, and can staff read it; and what is the kill switch, who can pull it, and has anyone rehearsed pulling it. Notice this is the same question as items 1 and 3 in different clothes: in each case the system holds a representation and takes action on it, and the practice's only real protection is being able to see what it held. Healthcare IT News / HIMSS TV (Trusted) · Watch it
