In THIS ISSUE
Signal of the Week, Anthropic's six partner results are real, early, and self-reported.
Framework of the Week, The Receipt Test: four checks before you quote an AI result.
5 Things Worth Reading, pilots, trials, EU rules, and oversight.
Tool of the Week, Europe PMC.
Reading time: ~8 min read
🎯 Signal of the Week
Anthropic's drug-development results are real, early, and self-reported.
Welcome to Issue 14,
Four hours instead of three months is the number from Anthropic's October 27 post that will travel furthest, so it deserves a careful read. In that post Anthropic reports early results from six drug-development partners and says it has acquired Mithril Medicine, a one-person company developing treatments for two rare forms of muscular dystrophy. At Enveda, Claude reviewed 1,061 pages of manufacturing documentation, caught all 11 issues the human reviewers had found plus 7 they had missed, and cut human review time by 89%, from over 25 hours to under 3. A regulatory document draft took 4 hours, against about three months by hand, and Enveda's review of a second draft found one minor error.
The biology is earlier on the ladder. Divergence Labs used Claude to generate about 35,000 antibody designs in roughly a month for a food-allergy program, and early tests found 3 designs that bind IgE and 20 that can dislodge it from immune cells, 3 of which dislodged it more than the standard medication. At BigHat Biosciences, about 1,010 variants were made and tested over about 10 weeks, and the best bound more than 10 times more tightly than the originals. Generate Biomedicines reports a CAR T design about 70% shorter that let cells make roughly 10 times as many receptors. At Mithril, founder Cody Tranbarger used Claude to coordinate 11 external labs and reached promising mouse results in under 18 months; Anthropic says human trials could start as soon as late 2028 if the candidates hold up. For The Column Group's new company Pleiades and for Bristol Myers Squibb, the post gives no outcome numbers.
Figure 1. The evidence ladder: where each of the partner results sits, from documents to humans. Sources: Anthropic; The Next Web; Becker's Hospital Review.
Read the caveats, because Anthropic states several itself. The work is early and mostly preclinical, most of the drugs have not been tested in humans, and BigHat's standard stability metrics were initially misleading until Claude and an Anthropic researcher reanalyzed the raw data. Most figures come from Anthropic or the partners themselves: Enveda's numbers come from its own press release, and the Mithril mouse data come from a technical report that has not been peer reviewed. Magica adds that Anthropic provides usage credits and engineering support to several collaborators, so the pilots are not independent validation, and The Next Web quotes no outside expert.
Here is why this belongs on a Medical Affairs agenda. This is the template for the claims that will reach your desk over the next year. Vendors, partners, and R&D colleagues will cite 89%, 10 times, and 70% shorter, and the field will ask you what stands behind each number. Medical Affairs is the function trusted to say what a result means and what it does not, and nobody else in the building is paid to ask what the comparison was.
Now the opinionated part. The Enveda result is probably the most useful in the post and also the least about medicine: document review is a task with a clear baseline and an answer you can check in hours. That is where AI will deliver first, and it is where your function can start too, in medical information, literature surveillance, and review of your own documents. What you must not do is let a paperwork result borrow the credibility of a mouse result, or a mouse result borrow the credibility of a patient result. Last week's issue asked who owns the building. This week's asks who set the baseline.
🧠 Framework of the Week
The Receipt Test: four checks before you quote an AI result.
What would you ask if a vendor handed you a slide that said 89%? Most of us nod, because the number is large and the logo is familiar. The Receipt Test is a way to slow that moment down.
It connects directly to this week's signal. Anthropic's post is a stack of receipts: each partner result says what was bought, but not every receipt shows the price. A document review measured in hours and an antibody screen measured in assay readouts are both results, yet they sit at very different distances from a patient. Magica's reading is that the model's contribution cannot be cleanly separated from the data systems, measurement choices, and human judgment around it.
Read the figure from the centre outwards: the claim sits in the middle, and each spoke is a question to put to it. Checks 1 and 2 you can usually answer from the announcement itself. Checks 3 and 4 need a conversation with whoever produced the number, and that is where Medical Affairs is most useful. The test works on any AI result, from a deal announcement to a conference poster.
Figure 2. The Receipt Test: four questions to put to any AI result before you quote it. Source: iCHealth Pathway framework (original).
Baseline. Compared with what, and who chose the comparison? Enveda's 89% compares AI-assisted review with human review time, and the three months compares a drafted filing with a manual one. Ask who set the baseline and whether a trained person with decent tools was on the other side. A large improvement over a weak baseline is still a weak result.
Stage. How far from a patient is this result? Place every claim on the ladder: documents, lab assays, animals, humans. Divergence's hits are early assay results, Mithril's are mouse data, and Anthropic itself says most of the drugs have not been tested in humans. Do not let a result borrow credibility from a stage it has not reached.
Source. Who measured it, who paid, and who reviewed it? Here the sponsor reports the results and also supports some partners, Enveda's figures come from its own release, and the Mithril data are an unreviewed technical report. Ask what has been peer reviewed and what has been independently reproduced, and treat everything else as a claim.
Isolation. How much of the result is the model, and how much is the lab, the data, and the people around it? At BigHat the win involved connected lab data and an Anthropic researcher who reanalyzed the raw data when the standard metrics looked misleading. Ask what the result would look like with another model, or with none.
The first two checks tell you how big the claim is. The last two tell you how far to trust whoever is making it. Most teams stop at the number. Start with the baseline, because it is the one thing a slide can leave out.
Figure 3. One screening funnel and one stopwatch: the numbers behind the headlines, as reported by Anthropic and its partners. Sources: Anthropic; Becker's Hospital Review.
Every AI result is a receipt. Check what was bought, and from whom, before you pay.
📚 5 things worth to read
Five paywall-free reads, each summarized so you get the value without the click
01
Anthropic Buys a One-Person Biotech and Starts Its Own Drug ProgrammesThe most useful independent recap of the announcement, written by Ana Maria Constantin and published October 28. It adds detail the other coverage skips: Mithril's main target is LAMA2-related congenital muscular dystrophy, and in a technical report mice treated with one antibody design reached about twice the grip strength of untreated mice. It also records that deal terms were not disclosed, and that Anthropic paid $400 million for Coefficient Bio in April. Note what is missing: it quotes no outside experts, and the Mithril data have not been peer reviewed. Use it as the neutral summary and as a worked example of the stage check. The line to keep: a mouse result is a reason to keep going, not a claim of efficacy.
02
What Claude's Drug-Development Pilots Actually Show
An unsigned critical reading of the same announcement, published October 27. Its argument is that the clearest gains are in connected workflows such as experimental design, lab coordination, raw-data review, and regulatory paperwork, which are easier to quantify than biology. It points out that the striking biological results are hits, assay readouts, or mouse findings, none of which is evidence of safety or efficacy in people. It also notes that data systems, measurement choices, and human judgment are part of the performance credited to the model, and that Anthropic supplies usage credits and engineering support to several collaborators. No efficacy percentages or sample sizes are given for the antibody work. The line to keep: the pilots show faster paperwork, not yet better medicine.
03
AZ and Oracle Team Up on AI Platform for Clinical TrialsA counterpoint from the same month: a big pharma pairing with a data and cloud company rather than a model-maker. The article, by Phil Taylor and published October 20, reports that AstraZeneca and Oracle have a non-binding agreement to use AI and real-world data in clinical research, for example to find suitable trial participants. The first programme is in Abu Dhabi, focused on hypertension, with local health company M42, identifying patients whose blood pressure is uncontrolled. More projects across AstraZeneca's core areas are due to start in the coming months in the US, Europe, and the Middle East. It gives no financial terms, patient counts, or outcome data, and the benefits are hopes, not results. The line to keep: an announcement of intent sits below the first rung of the ladder.
04
AI Act: Obligations in Force and Delays Under Omnibus 2026/1744
A plain-language status check dated October 6 on the EU's Digital Omnibus on AI, Regulation (EU) 2026/1744, which the article says was published in the Official Journal on July 24 and entered into force on July 27. Stand-alone high-risk obligations move from August 2, 2026 to December 2, 2027, and high-risk AI in regulated products from August 2, 2027 to August 2, 2028. Article 50 has applied since August 2, 2026, but marking for generative systems already on the market is deferred to December 2, 2026, about a month away. The article is a secondary source, harmonised standards were still pending, and it is no substitute for the legal text. The line to keep: delay is not exemption, so keep your AI inventory current.
05
Most Patients Won't Endorse AI Medical Responses Without a Clinician's Review
A press release on a survey of 1,185 US adults about AI in secure patient portals, published in JAMA Network Open on October 14 by a Columbia University School of Nursing team. 63% would not endorse AI-generated medical responses without a clinician's review, and two-thirds would rather wait up to 48 hours for a clinician to review and edit an answer to a medical question. At the same time, 57% preferred fast AI responses for logistical tasks such as appointment scheduling and prescription refills. The release gives no sampling method or limitations, and the findings are self-reported attitudes, not behavior. The line to keep: people accept AI for logistics but still want a human to sign the medical answer, a useful rule for medical information teams.
🔧 TOOL OF THE WEEK
Europe PMC
What it is: a free life-science literature database run by EMBL-EBI that indexes publications and preprints together, so you can see what has been peer reviewed and what has not. A 2023 protocol paper by its team, with figures as of September 2022, described over 41 million records, full text for over 8.1 million of them, and preprints from more than twenty platforms including bioRxiv, medRxiv, and Research Square. Preprints carry a green Preprint label and an orange notice that they may not have been peer reviewed, with a link to the journal version when one exists. It is a search and evidence-checking tool, not a chatbot, which is exactly why it suits this week's job. Those numbers have grown since, so check the current site.
How to use it for Medical Affairs and MSLs: when a vendor or colleague cites a study, paste the title into Europe PMC and read the label first: preprint, preprint with a journal version, or published article. Then open the Reviews section, which gathers peer reviews and evaluations from services such as Sciety, PREreview, and Peer Community In, and the Citations and impact panel, which shows yearly citations and scite smart citations. In February 2024 Europe PMC reported about 9,500 expert-evaluated preprints and a filter for reviewed preprints. Finally, use the free full text filter and the export of up to 50,000 records at once to build a quick evidence file behind a claim before you repeat it in a meeting.
Where to access it: europepmc.org, free to use. Text-mining annotations and figure previews depend on each article's license, and full text is not always available, so confirm what you can see before you plan a process around it. The figures above come from sources dated 2022 to 2024.
COMPLIANCE NOTE
Europe PMC lists preprints next to published papers, and preprints are not peer reviewed, so label the status whenever you cite one and do not use a preprint as established evidence in material that goes through MLR review. Check the license before you reuse any text or figure, because reuse rights vary by article. Record your search strings, filters, and dates if the result supports a deliverable, so someone else can repeat it. Do not type confidential pipeline plans, unpublished data, or patient-identifiable information into the search box, and check your company's policy on external search tools.
That is Issue 14. If a colleague forwarded this to you, you can get it yourself every Monday at newsletter.ichealth-ai.com/subscribe.
One email a week. No sponsorship. See you next Monday.
iCHealth Pathway is a weekly research note by Dr. Issam Chebouti on AI transformation for Pharma, Medical Affairs, and Healthcare leaders.
iCHealth-AI, ichealth-ai.com, [email protected]
Issue 14, Monday, November 2nd, 2026.
Parts of this issue were produced with AI assistance and checked before publication.