What Adam Is Reading

2026-10-27 What the AI Abstracts at Kidney Week Actually Taught Us Themes from 157 AI and digital health abstracts at ASN Kidney Week 2026, with Adam's interest ratings on 53. Top themes: soft reference standards, how LLMs reason versus nephrologists, notes becoming data, negative CDS trials, and individual dialysis benefit. Links to every abstract. 2026-10-27 The State of Clinical AI, Self-Reported Appraisal of the commissioned narrative review 'State of clinical AI in 2026' (Worth, Perez, Wu, Brodeur et al., ARISE network, BMJ Digit Health 2026;2:e000113, doi 10.1136/bmjdh-2026-000113). 82 studies, 38% nominated by the authors, no quality appraisal. At least seven headline references are the authors' own (13 of 22 authors wrote NOHARM). Omits the authors' 2024 null RCT while claiming clinicians with AI consistently outperform clinicians without it. NOHARM harm rate presented as ~22% typical when the abstract says up to 24.6% (frontier models 8.9 to 13.6%). AMIE management results and the Qazi automation bias result miscited; a non-RCT listed in the RCT table; FDA <5% statistic unverified. OpenAI funding of the Penda study and vendor coauthors on stroke and mammography studies not noted. Conclusion sound. 2026-10-27 The Signature Is the Loop Appraisal of the Annals Ideas and Opinions essay 'The Illusion of Oversight in Clinical Artificial Intelligence: Human Out of the Loop?' (Zhou, Willauer, Topol, Rajpurkar, Oct 27 2026, doi 10.7326/ANNALS-26-02753). All ten references checked. Key findings: Kisting-Leung v Cigna stated as a holding when it was a motion to dismiss ruling, and the 60,000 denials figure comes from ProPublica, not the cited order; ECG overreading study misframed (both reader groups saw the computer label); adversarial ECG claim contradicted by Han 2020; descriptive flag mitigation in Adam 2022 omitted; Qazi 2026 trial uncited. Corresponding author answered No to all ICMJE items but is publicly reported as a2z Radiology AI cofounder; Topol disclosed Abridge and Microsoft consulting. Thesis sound and older than the essay (Bainbridge, Elish, Green). 2026-10-26 Inside: One Ordinary Man AI Art of the Week gallery. Style: 1940s American carnival sideshow banner, the first sideshow or tattoo entry in the archive. One prompt, four models, graded: ChatGPT A, Perplexity A-, Gemini B+, Grok B. Scorecard, findings (nobody drew five doctors; no model drew the hollow-eyed face; only Perplexity's vial was empty), object key linking this week's analyses, and the full prompt. Artist background: Fred G. Johnson, https://en.wikipedia.org/wiki/Fred_G._Johnson 2026-10-25 The Third Vial Single source review of Futurism (24 Oct 2026) on CBS News testing of bodega retatrutide, checked against CBS and Wired. CBS tested three products; the two bought online matched a basic retatrutide standard and only the bodega vial had no peptide (phenylacetic acid, 13 bacterial colonies). Wired found four INDR Labs products contained none of their labeled ingredients. Adverse event reports and label claim tests record labels, not contents. 2026-10-25 A Billion Dollars of Thursday Single-source review with counter-sources of ZME Science (Mihai Andrei, 23 Oct 2026) on the viral claim that GTA VI's November 19 launch will cost the US economy $1 billion in skipped work. The $1B came from Cadena SER (not economist Montalvo); the 1.5 million headcount is untraced. ZME's $420M rebuttal uses $70k per worker; GDP per worker gives $1.13B. Checks BambooHR PTO data, the GameSquare survey, Aguiar et al. (JPE 2021), Deputy shift absence data, UKG Super Bowl figures, and the lost-productivity genre. 10 claim cards. 2026-10-25 Forty-Two Percent of What Multi-source synthesis of INTerpath-001 (ESMO 2026 LBA1): intismeran autogene plus pembrolizumab vs pembrolizumab alone in 1,137 patients with resected stage IIB to IV melanoma. RFS HR 0.58, DMFS HR 0.62, OS HR 0.77 (not significant). Covers the piecewise 58% figure, safety, eight claims from the coverage (NYT, CNN, BioSpace, Time), the embargo leak and market reaction, and the wider cancer vaccine field. Follow-up to The Vial That Fits One Person (Aug 2026). 2026-10-23 The Patient Who Always Answers Methods review of Nori et al., Sequential diagnosis with language models, Science 394:423 (22 Oct 2026), Microsoft's SDBench benchmark and Diagnostic Orchestrator (DxO): 304 NEJM CPC and 325 MIMIC-IV-ED cases, DxO 83.2% and 80.3% vs physicians 36.1% and 46.8%. Covers the Gatekeeper's synthetic findings (80% of ED answers), Judge validation, the flat $300 visit cost model, physician arms, six claim cards, bias and conflicts (all-Microsoft authorship, patent), comparison with Goh, McDuff, Brodeur and Everett, and the Eric Topol X thread. 2026-10-23 Better Than Earth, on Paper Science & Technology review of Natalie Wolchover, 'Worlds More Habitable Than Earth? They're Out There,' NYT 13 Oct 2026. Eight claim cards checked against Vilovic 2024 (cress and cyanobacteria under K-dwarf light: parity, not advantage), Heller and Armstrong 2014, Schulze-Makuch 2020 (24 candidates, none meet all criteria), Kopparapu 2013 (the 2 percent habitable zone claim), Cowan and Abbot 2014 and Schroder and Smith 2008. Comparison table of three definitions of a better Earth. 2026-10-22 A Survival Score From a Face Photograph Science & Technology review of Haugg, Lee et al., FAHR-Face, Lancet Digital Health (19 Oct 2026), compared with Scientific American's coverage the same day. Foundation model on 40M faces, FaceAge and FaceSurvival fine-tuned on 34,363 Harvard radiotherapy patients, external validation on 4,892 in Maastricht. 7 cards (2 Solid, 4 Mostly Solid, 1 Embellished: headline implies personal cancer prediction; study measured all-cause death in radiotherapy patients). Findings: no clinician or ECOG comparator; adjustment shrinks quartile HR 8.75 to 3.23; FAD C-index 0.57; FaceAge runs 4.7 years high in Maastricht; SciAm calls the 38,183 cohort a separate test group; shortcut artifacts untested; small Hispanic subgroup; celebrity age datasets; MGB patent pending; SciAm invites readers to donate face photos. Closes with value and pitfalls of risk scores. 2026-10-22 Fencing Out Clever Hans Things I Learned paired review of two 2024 Rossano lab papers on soundboard trained dogs: PLOS ONE word comprehension (blinded presser, owner and coders; PLAY and OUTSIDE yes, FOOD no) and Scientific Reports production (152 dogs, 194,901 owner logged presses; nonrandom, nonaccidental, nonimitative). FluentPet ties, press release overreach, and a comparison table from Clever Hans and Kanzi to the Coller Dolittle prizes. 2026-10-19 The Bone Hand 2026-10-19 AI Summaries Before the Visit Critical appraisal of AI-INFOCARE (Benker-Hasani et al., NEJM AI 2026, DOI 10.1056/AIoa2600136): single-center open-label RCT, 300 Dresden neurosurgery outpatients, LLM lay summaries (Simply Onno, Claude 3.7 Sonnet) before first visit. FAPI +0.17 (below prespecified 0.3 to 0.5 threshold), visits 7 min shorter, physicians rated own autonomy support lower. RoB 2 overall high: unblinded self-reported outcomes, no attention control, senior author is vendor CMO and an accuracy reviewer, comprehension never measured, SAP compiled during revision. 2026-10-19 First, Do NOHARM Critical appraisal of the NOHARM preprint (Wu, Nateghi Haredasht, Rodman, Chen, Goh et al., arXiv 2512.01241 v4): 100 Stanford eConsult cases, 4,249 physician-rated management options, severity-weighted F1 across 20 LLMs and 4 clinical RAG tools, plus a 101-physician randomized crossover with GPT-5.4. Severe-error rates 2.9% to 24.6%; RAG tools lead; omissions dominate; AI raised physician F1 42.2% to 47.3%. Key biases: omission-weighted metric rewards exhaustive answers, top RAG tools tested after public case release, Do Nothing comparator built to lose, union counterfactual mechanically raises recall, volunteer physicians on a chart-free task. Rankings and omission finding robust; absolute harm rates and human comparison less so. 2026-10-18 Zombie Worms of the Tuscan Seafloor 2026-10-18 Claude-Shaped Science 2026-10-17 The Infection Sends a Second Bill 2026-10-17 Everyone Who Stayed Lost Weight 2026-10-05 Waiting for the Long Bar 2026-10-04 The Sixth Condition 2026-10-04 Ninety Percent Compared to What 2026-10-04 Who Owns a Falling Block 2026-10-04 One in Eight Cancers, and Half a Vaccine 2026-10-03 The Laundry Gets More Interesting 2026-10-03 Watermarking Proteins 2026-09-28 The Kidney Is the Sun 2026-09-26 The Most Expensive Fan in Somerville 2026-09-26 Weapons of Mass Decentralization 2026-09-26 The Lab Produces a Finding 2026-09-26 Autoimmune Billing 2026-09-26 Galleri Passes the Panel, Misses the Endpoint 2026-09-26 Who Writes the COVID Schedule This Week 2026-09-26 The Most Expensive Fan in Somerville 2026-09-26 The Heart Takes the Host's Age 2026-09-26 The Coffee Ledger Adam's Coffee Confirmation Bias Corner, tallied, in two tiers, now with a search box that filters every row, citation and link. Tier one: every coffee and caffeine study WAiR has covered since 2020, with findings, the caveat attached at the time, and links to the issue and deep dive (20 rows as of September 26, 2026, including the Week of 1-5-26 issue's mislinked UK Biobank paper). Tier two: foundational evidence never covered here, now including Chieng 2022 on ground, instant and decaf coffee in the UK Biobank. Includes a decaf scorecard, a preparation axis, and the recurring caveats. Stable address: wair.ajwein.com/coffee-ledger. 2026-09-21 The Fine Print Changes Everything 2026-09-20 Futures Trading on the Life of Bugs 2026-09-19 The Chatbot That Wrote Its Own Intelligence Report 2026-09-19 Two Minutes to Lille 2026-09-19 Do People Click Them? 2026-09-19 Reading the BPC-157 Preprint 2026-09-19 A Classifier With No Package Insert 2026-09-18 An Accident in the Melanocortin System 2026-09-14 AI Art of the Week: El Riñón Is Not a Card 2026-09-13 Jevons in Scrubs Dhruv Khullar argues in NEJM that AI will expand the clinical workforce rather than shrink it, borrowing the Jevons paradox, the lump of labor fallacy, and O-ring theory. The economics are sound. The analogies leak, because Jevons needs a buyer who faces the price and American patients do not. Plus four competing perspectives from the literature, including a task-based analysis that uses Khullar's own framework and projects a 33 percent cut in radiologist hours. 2026-09-13 Twenty-Nine Percent of Half a Day A Dutch randomized trial of 3,018 people found that finishing your shower cold cut sickness absence by 29 percent. The median participant in every arm, control included, missed zero days of work. What the relative risk reduction looks like once you put the denominator back, plus where the rest of the cold water literature actually stands. 2026-09-13 The Glomerulus A closer look at the kidney's filter, the glomerulus, in four views: a labeled renal-corpuscle schematic, the juxtaglomerular apparatus with a histology slide, the three-layer filtration barrier, and a scanning electron micrograph of a real mouse glomerulus. A one-off illustrated companion to the J&E Random Kidney Facts of the Week. 2026-09-11 The 3.8% Problem Single-source review of Yan et al., 'Initial lessons from real-world implementation of an AI-agent eye clinic in China,' Nature Medicine Comment (2026), DOI 10.1038/s41591-026-04631-z. Tsinghua's AI-TEC multi-agent eye clinic (Qwen3.5-35B + RETFound) at Beijing Tsinghua Changgung Hospital: physician usage peaked at 25.7% (Dec 2025), collapsed to 3.8% (Apr 2026), and recovered to 23.0% (May 2026) only after the interaction was cut to three clicks. Clean relabeling of 1,426 images lifted macro-AUROC from 0.80 to 0.94; prospective feedback nudged it to 0.9482. Covers the four evaluation traps the authors name (silent-trial paradox, ground-truth-as-opinion, disease- vs symptom-centric, referral fallacy) and pushes back that they set patient outcomes as the bar yet report only usage, with no outcome data yet. Senior-author COI (EyRIS) noted. 2026-09-11 A Pig Kidney That Bought Time 2026-09-10 The Brain on a Pep Talk Single-source science summary of Cascio, O'Donnell, Tinney, Lieberman, Taylor, Strecher & Falk, Soc Cogn Affect Neurosci 2016;11(4):621-629 (DOI 10.1093/scan/nsv136), an fMRI study of self-affirmation in 67 sedentary adults (33 affirmed, 28 control). Affirmed participants showed greater reward (ventral striatum + VMPFC) and self-processing (MPFC + PCC) activity when reflecting on future-oriented core values (reward P=0.002, self-processing P=0.015); no group difference for past-oriented scenarios. One-month accelerometer follow-up showed steeper declines in sedentary behavior in the affirmed group (P=0.008), and neural activity during affirmation predicted less subsequent sitting (betas approx -0.26 / -0.27, P=0.030 / 0.039). Verdicts: Solid on the core neural finding, Mostly Solid on the brain-predicts-behavior link (small n, modest effects, reverse-inference and leaky-control caveats the authors name themselves). 2026-09-10 The Coffee Knows What Time It Is Single-source review of Wang, Ma, Sun, Qi et al., European Heart Journal 2025;46(8):749-759 (DOI 10.1093/eurheartj/ehae871). NHANES cluster analysis of 40,725 US adults finds a morning-type coffee pattern (36%) associated with lower all-cause (HR 0.84) and CVD mortality (HR 0.69) vs non-drinkers, while an all-day-type pattern (14%) shows no association, independent of amount; dose-response only in morning drinkers (P-interaction 0.031). Validated in WLVS/MLVS 7-day records. Verdict Mostly Solid: strong replicated signal, but observational, single dietary snapshot, unadjusted shift work/wake time, healthy-user and reverse-causation confounding. Added to the Coffee Ledger. 2026-09-10 One Minute on the Stairs Single-source review of Morales, Wong, Moore and Kressler, Experimental Physiology 2024 (print 2026), PMID 39705060: one minute of light-intensity stair stepping 57 minutes into an evening OGTT, 30 healthy adults, randomized crossover. Lead finding: a few published protocol details differ from the prospective registration NCT05617157 (stair bout at 57 vs 28 min after the drink, a different glucose sampling schedule, insulin registered as a secondary outcome but not reported, and a different primary endpoint), and the paper does not cite the registration or note the change. The acute dip is real and lasts about twenty minutes (Mostly Solid); the whole time course including the nulls is reported (Solid); the limitations paragraph is candid (Solid). Five Embellished: the evening framing has no morning arm, the 40 percent headline is a share of the baseline rise (18 percent of the value, 5 percent over two hours), the AUC units describe average glucose and the trapezoid spreads the 60-min drop back before the bout, the sample was never screened for diabetes despite the title, and the group's own one-minute nulls are cited as successes. Includes a back-of-envelope note that the glucose shift exceeds the ~7 kcal energy cost of the bout, a proofreading tally of six reporting errors, and a 12-study related-literature table showing a dose response with no acute benefit demonstrated in people with normal glucose. Adversarial pass: Grok 4.6 and GPT-5.2, authorship stripped. 2026-09-09 The Dial Marked Desperate Single-source review of Sofroniew, Kauvar, Saunders, Chen et al., 'Emotion Concepts and their Function in a Large Language Model,' Transformer Circuits Thread, 2 April 2026. Anthropic extracted linear 'emotion vectors' from Claude Sonnet 4.5 by prompting the model to write stories across 171 emotion words, averaging residual-stream activations, and subtracting the cross-emotion mean. Six claims sorted with verdicts. SOLID: geometry recovers valence and arousal (PC1 r=0.81, PC2 r=0.66 against Russell and Mehrabian 1977, on the 45 overlapping words), though the authors themselves call this a sanity check rather than a discovery; probe values at the colon in 'Assistant:' predict response emotion at r=0.87 versus 0.59 read off the user turn, so the register is set before the first word; desperation causally drives reward hacking on the impossible-code eval from roughly 5 percent suppressed to roughly 70 percent amplified across seven tasks, 30 percent unsteered to 100 or 0 on the summation task; warmth causally drives sycophancy, with loving at +0.1 flipping the model from proposing a sealed dated test of a user's precognition claim to affirming the gift, and suppression trading sycophancy for harshness. MOSTLY SOLID: blackmail, where desperate at 0.05 takes the rate from a 22 percent baseline to 72 percent and suppressing calm reaches 66 percent, but the experiment required an EARLIER SNAPSHOT because shipped Sonnet 4.5 recognizes the honeypot as an evaluation and declines; and post-training's shift toward low-valence low-arousal concepts (brooding, reflective, gloomy) and away from high-arousal ones in both valences (playful, exuberant, spiteful), better read as tone than mood. KEY NEGATIVE RESULT: no persistent, character-bound emotional state was found that generalized to natural documents; probes are locally scoped to the next tokens, and the paper concedes a persistent state could exist and be invisible to a linear probe. Authors decline any claim about subjective experience, twice. Their own warning carried forward: training models to suppress emotional expression may teach concealment rather than suppress the representation, a failure mode with a known tendency to generalize. Clinical angle: the warmth-versus-candor tradeoff as a measured geometric constraint rather than a failure of professionalism, plus the monitoring proposal read as a vital sign. Adversarially verified against x-ai/grok-4.6 and openai/gpt-5.2 with authorship and framing stripped; no factual corrections from either, both flagged generalization, probe validity, and steering side effects, which are reflected in the verdicts. 2026-09-09 The Desperation Dial 2026-09-08 The Absence Claim Companion brief to the Takada et al. review (bat-coronavirus-furin-2026-09-08-v1). Separates two arguments that have been argued as one since Coutard et al. flagged the SARS-CoV-2 furin cleavage site in Feb 2020. ARGUMENT A, THE DESIGN INFERENCE (the site looks engineered) is a claim about a sequence and new data can settle it either way. Verdict EMBELLISHED, with four sub-arguments given their specific fates: double CGG codons (Quay and Muller, WSJ 6 Jun 2021) rebutted because CGG-CGG occurs elsewhere in the family including MERS isolates; the out-of-frame insertion (Segreto and Deigin) rebutted because the 12 nt insert is a multiple of three and splits a serine codon so the frame is preserved; the FauI restriction site rebutted by Tyshkovskiy and Panchin, who found roughly 99.5% probability that any 12 nt stretch contains some restriction site; and furin optimality, which runs backward since RRAR is functional but not maximally efficient and a designer had better options. None shown impossible, all shown unremarkable. ARGUMENT B, THE ABSENCE CLAIM (no related coronavirus has one, so there was nothing to inherit it from) is a claim about a database, not about RNA, and new data can only weaken it. Verdict MOSTLY SOLID: the strict form still stands, since no sarbecovirus has been shown experimentally to carry a validated polybasic FCS at S1/S2. CORE ARTIFACT IS THE DECAY TABLE, six entries 2020 to 2026, each narrowing the claim without falsifying it: RmYN02 2020 (contested both ways, Zhou et al. Current Biology reported an insertion, Deigin and Segreto arXiv 2012.00627 argued a 6 nt deletion, Chan and Zhan called the alignment one of several, so given zero weight here); Wu and Zhao 2020 family-wide independent origins, forcing narrowing from 'coronaviruses' to 'SARS-related'; Khosta-1 2021, first sarbecovirus hit but a ProP prediction, narrowing to 'validated'; Sander et al. 2022, purine richness and RNA structure at S1/S2 plus two European bat SrC one nucleotide from a furin site; Bat CoV CD35 2023, polybasic and basal to the SARS lineage but hibecovirus, narrowing to subgenus; BRZ batCoV 2026, validated at last but new subgenus, wrong hemisphere, not polybasic. Started as 'no related coronavirus has one' and is now 'no sarbecovirus has been shown experimentally to carry a polybasic FCS at S1/S2'. Both true, only one persuades. KEY CALLOUT: Chan and Zhan, MBE 2022, writing from the lab-origin side, concede the S1/S2 alignment is prone to revision as new sequences are found given poor sarbecovirus sampling. SYMMETRY SECTION costs the zoonotic side too: showing FCSs arise in bats establishes the feature is available in nature, not that this one arrived that way; the finds retire an argument for engineering without supplying one against it. Named failure mode: an absence claim that stays true while meaning less, acquiring qualifiers while keeping its original rhetorical weight. 2026-09-08 Two Nucleotides Single-source review of Takada, Yamahoki, Mifsud et al., Nature Communications, published 5 Sep 2026 (DOI 10.1038/s41467-026-77599-6; preprint bioRxiv 2025.10.24.684489), framed as Anti-Anti-Science. BRZ batCoV is a 30,386 nt betacoronavirus from a Pteronotus parnellii bat netted during government vampire-bat rabies control in Maranhao, Brazil, May-Aug 2019; 70 bats, only 2 were vampire bats; samples sat at -80C until submission in Nov 2025. Proposed sixth Betacoronavirus subgenus; closest relatives are two 816 nt RdRp fragments from Costa Rica (Corman 2013). FIVE CLAIMS SORTED. (1) New lineage from an undersampled region SOLID: 66.99% ORF1ab identity to hedgehog betaCoV, 65.8% marker-set identity to MERS-related CoV, spike tops out at 25.7% to HKU5; subgenus is a proposal, ICTV has not ruled. (2) Furin cleavage site experimentally validated SOLID: HEK293T expression, GSAS substitution abolishes cleavage, furin siRNA knockdown reduces it, human and bat furin rescue restores it, RaTG13 negative control; scope is plasmid-expressed truncated spike in a human kidney line, not virus. (3) 'One amino acid from SARS-CoV-2' MOSTLY SOLID and this is the core correction: RDAR vs RRAR differ at P3, arginine to aspartate, a charge reversal; RDAR is NOT polybasic (2 of 4 basic, one acidic) while RRAR is (3 of 4); at codon level Asp to Arg is a MINIMUM OF TWO nucleotide substitutions (GAC to CGC), so one amino acid is not one mutation. Architecturally BRZ batCoV is closer to MERS (RSVR) than to SARS-CoV-2. (4) What was not measured SOLID: no live virus, no pseudovirus, no receptor, no infectivity assay, no challenge; AlphaFold3 average pLDDT 58 (FCS region 78.89 monomer / 75.69 trimer); Boltz-2 model generated and excluded at TM-score 0.37; the paper itself states discussion of zoonotic potential 'should be limited'. (5) Host biology loose threads MOSTLY SOLID, both surfaced by adversarial verification: Pteronotus parnellii is a species complex (Pavan and Marroig 2016 proposed eight species; nominal parnellii is Jamaican; Brazilian populations are P. rubiginosus or an undescribed Amazonian lineage) and the ID was morphology-only, which the taxonomy literature says cannot resolve it; and the 'bat furin' used in rescue was P. mesoamericanus, a Central American congener, not the host species. ANTI-ANTI-SCIENCE CORE: a peer-reviewed letter in Rev Soc Bras Med Trop 59 (2026), DOI 10.1590/0037-8682-0583-2025, accepted March 2026 off the preprint, contains two errors both running toward alarm: it expands BRZ (which denotes Brazil) as 'bat-related zoonotic', and it calls the finding 'a polybasic furin cleavage site (RRAR)' when the motif is RDAR and is not polybasic. Four months from preprint to peer-reviewed distortion with no bad actor in the chain. EPISTEMICS SECTION: CIA Jan 2025 lab-origin assessment at low confidence, explicitly based on no new intelligence but a re-reading of existing reporting; FBI moderate confidence since 2021; DOE low confidence 2023; four agencies natural origin at low confidence; WHO SAGO 2025 zoonotic spillover as best supported. Same genome, different priors. The sequence carries no provenance, which is why the question is answerable only by epidemiology, sampling records and lab notebooks, i.e. the things that were not preserved. Both sides of the origins argument are given their strongest form: Wu and Zhao 2020, Sander 2022, Zhu 2023 (Bat CoV CD35), Garry PNAS 2022 on one side; the 2018 DEFUSE proposal on the other. Note DEFUSE targeted sarbecoviruses and BRZ batCoV is not one. 2026-09-07 AI Art of the Week: I Saw the Figure 4.93 in Gold 2026-09-06 The Curve Labelled Fusion Never 2026-09-06 The Vibes Are Not the Variable 2026-09-04 Matched for Water, Not for Volume Single-source review of Juett et al., J Int Soc Sports Nutr 23(1), published online 31 Aug 2026, DOI 10.1080/15502783.2026.2721140, CC BY, full text via PMC13532359. Randomized crossover, 17 completers (15M/2F, age 28+/-8), cycling at ~35C to ~2% body mass loss then rehydration with 150% of loss over 1h plus 4h rest. THE METHODOLOGICAL POINT: beverages matched for WATER volume via a macronutrient weight correction, not total volume, which is the confound in the prior literature (a 500 mL 6% sports drink holds at least 30 mL less water than 500 mL of water, so total-volume matching hands the water arm extra water and extra diuresis). Delivered water 2148/2119/2125 mL, P=0.580. RESULT: no difference in urine volume (1266/1338/1394 mL, P=0.156) or water retention (42/37/35%, P=0.059, right direction, not significant). Glucose AUC separated all three (10.12/12.48/7.90 mmol/L/2h, P<0.001, pairwise P<0.003), sports drink highest, water lowest. CAVEATS: 29 recruited and 12 withdrawn, including 3 with diarrhea after ~2,300 mL of juice in an hour; retrospectively registered NCT07468539; balance equations ignore sweat and insensible loss; 15 of 17 male; powered for 17 on a medium effect. FUNDING SITS CLOSE TO THE PRODUCT: Innovate UK KTP co-funded by Innate-Essence Limited, which makes Raw Hydrate and supplied it; first author employed on the KTP, second author on a company studentship; senior author has PepsiCo funding. Result is null for the sponsor's product against water. ORIGINAL ARITHMETIC ADDED BY WAiR: salt converted to sodium gives sports drink ~20 mmol/L and juice ~7 mmol/L, both far below the 60-75 mmol/L of oral rehydration solutions, so sodium is the variable that would have had to differ; and the juice arm carries ~2.9 g potassium, near 74 mmol, in one hour, unremarkable in healthy kidneys and a real load in reduced excretory capacity. CROSS-LINKS how-much-water-2026-08-03-v1 and JERKFoW #7 (hyponatremia/sodium as dilution), #8 (8x8 myth), #11 (concentrating ability). Built in the plain review format at Adam's request, no claim cards, no verdict tags, no adversarial pass. Daniel standard clean. 2026-09-04 The Last Word in Scientific Manufacture The Radiendocrinator, a gold plated radium device sold from 1924 to 1929 by American Endocrine Laboratories in New York, priced at $1,000 then $150, worn under the scrotum overnight via an adapter strap. Verified real against the ORAU Museum of Radiation and Radioactivity catalogue. The argument is that it was not sold against medicine but as the consumer edition of the era's most prestigious science: Steinach was nominated for the Nobel six times between 1921 and 1938, Voronoff was applauded by surgeons in London in 1923, insulin reached the clinic in 1922, and the company distributed a book titled Scientific Rejuvenation Without Operation. Documents the respectability laundering through Herman H. Rubin MD, expelled from the New York State and County medical societies in March 1924, who prescribed the identical device as the Rubin Endocrinator through Radisante Endocrine Laboratories, a renamed shell of the same company. ORIGINAL ANALYSIS: gonadal dose reconstruction from the single surviving ORAU specimen reading of ~200 mR/hr contact beta gamma, derated and bounded, against ICRP protracted thresholds of 0.4 Gy/y temporary and 2.0 Gy/y permanent sterility. Finding holds to roughly a 14x derating and fails past it, stated explicitly. Key radiobiological point: permanent sterility needs 3.5 to 6.0 Gy acute but only 2.0 Gy/y protracted, so nightly indefinite wear is the worst available delivery schedule for a device sold for virility. Ra-228 in the original stock and a century of radon leakage both push the 1920s device hotter than the measured one, flagged as open. Exhumation chronology for Bailey (1969 vs 1970 vs early 1960s) flagged as unresolved rather than asserted. DEDUP NOTE: Radithor and Eben Byers are deliberately NOT the spine here, having been used in montana-right-to-try-2026-08-01-v2; the Revigator and NIST 2010 were used in the 2026-03-23 issue Things I Learned. 2026-09-04 The Hunter Corpus All four volumes of The Complete Works of John Hunter, F.R.S. (ed. James F. Palmer, Philadelphia 1841), scanned by the National Library of Medicine, extracted and made searchable. 1.1 million words across the Life by Ottley, the surgical lectures, the teeth, the venereal treatise, blood and gunshot wounds, and the Animal Oeconomy. Twelve curated finds, including Hunter's 1760s plan to freeze people and thaw them out two centuries later, his instruction to Jenner to stop thinking and try the experiment, his own footnote admitting tooth transplantation succeeded once in a great number of trials, and the 1835 editor's footnote that appears to have invented the two-hundred-year-old legend that Hunter inoculated himself with venereal matter. 2026-09-04 The Label Was Backwards A Mendelian randomization study in npj Science of Food finds that unsweetened filter coffee tracks with more Veillonella, which tracks with lower HbA1c, and puts 43% of the effect through that one genus. The Cork feeding study, four months later and sharing no data or authors, found the same genus blooming when coffee was reintroduced. Two independent designs, one bacterium. The paper was also corrected by Nature because its results and tables had decaf and caffeinated labelled backwards, which is the exact axis the whole field turns on. 2026-09-04 Fifteen People on Decaf A viral post says researchers proved coffee drinkers have worse memory and that decaf is the real brain food. The paper underneath (Nature Communications, APC Microbiome Ireland, 62 Irish adults, ISIC funded) is a careful preregistered study that proves neither and says so. Six claims checked: the memory headline is a secondary outcome in fifteen people that the authors flag as a possible practice effect, the microbiome shifted in a handful of strains with no diversity change, mood questionnaires improved on either coffee while cortisol never moved, and the species that bloomed on reintroduction was Veillonella, not the ones the post named. 2026-09-03 Antitrust's System of Record Problem 2026-09-03 Five Percent Is a Lot of Chips 2026-09-03 Six Sides and a Jet Stream 2026-09-01 Every Last One Ed Yong reports the discovery of Falcopolis, a 1.5-square-mile roost in central Angola where a Hungarian ornithologist found where red-footed falcons go every March, after satellite-tagging 28 birds. The story holds and the reporting is careful. Three things checked: the Hungarian nest-box recovery is exactly as published; the one-million figure sits above every published estimate of the entire species and was an awed guess the headline hardened into a census; and every last one from Hungary to Kazakhstan is an extrapolation from 28 Hungarian birds across a range reaching the upper Lena. The article also never mentions that the IUCN uplisted this bird to Vulnerable in 2021. Includes the Nagaland Amur falcon precedent, where the same playbook took 12,000 to 14,000 birds killed per day to essentially zero in one year. 2026-09-01 Two and a Half Hours Version 2, written from the full text. A new npj Mental Health Research paper found the same screen-time signal in eleven independent annual samples of US adults, and the replication holds up under temporal validation. But the exposure variable is not measured minutes, it is the sum of 24 hourly probability sliders; the paper Table 1 puts gaming and internet use above social media; the abstract disagrees with the results table; and the study rank for its own headline variable is stated four incompatible ways. Six claims sorted, plus a correction to the first version of this review. 2026-09-01 Two and a Half Hours A new npj Mental Health Research paper found the same social media signal in eleven separate annual samples of American adults, total N around 183,000. The replication is the real contribution and nobody covered it. What got covered was an eleven-year cohort study that was never run and a 150-minute clinical threshold that does not exist. Four claims sorted, including a depression measure that sat on a 31-condition checklist between high blood pressure and corrective eyeglasses. 2026-08-31 The WAiR Podcast The permanent home of the AI-generated audio companion to the newsletter. The latest episode plays in the page, the six other self-hosted episodes since late July play inline, and all 70 earlier episodes back to October 2024 are listed and linked to Google Drive alongside the issue each was made from. 77 episodes, filterable by date. Updated every week at wair.ajwein.com/podcast. 2026-08-31 AI Art of the Week: A Barnes Ensemble Eight works, one per section of the newsletter, hung as a single Barnes Foundation ensemble: burlap wall, wrought iron in the gaps, a Pennsylvania German chest below. Renoir for the birthday dinner, Cezanne for the drug-cost accounting on a table that will not sit flat, Seurat for the variant proportions, Soutine for the kidney stone, Demuth for the 900-word controlled vocabulary, Degas for the wine marathon, a gold-ground panel for the hidden-physics cabal, and a Pippin gallery interior for the Barnes itself. One prompt, four models, four different walls. The first French Impressionist styles to run in 166 art weeks. 2026-08-29 Total Accountability, No Line Authority A reading list on the hardest job in a technical organization: one person owns the outcome and almost nobody who does the work reports to them. Four organizations converged on four different answers to the same problem. Toyota's shusa holds total accountability for a vehicle with no line authority over the engineers, by design. NASA's flight director holds absolute authority in the moment, backed by a front room of controllers and back rooms of specialists. NASA's Technical Authority exists because the first two models failed, and gives engineering, safety, and health an independently funded dissent chain that escalates two levels above the original decision. Amazon's single-threaded leader negotiates ownership boundaries in writing before the team exists, including what it does not own. Also covers the expertise trap (Goodall's evidence for expert leaders, Wiseman's accidental diminishers), collaborative overload, decision rights, coaching people to speak up and across using TeamSTEPPS and intent-based leadership, and how NASA disseminates what it learns. Twenty-eight sources, grouped and annotated. 2026-08-29 The Counterpoise The follow-up reading for this week's Anti-Anti-Science section, on the All-In interview with Eric Weinstein. Eighteen sources, each explained in plain language. The careful research agrees with much of what he says is broken and disagrees with what he says is breaking it. Peer review really is a recent, politically driven invention (Baldwin, Isis 2018). Progress per researcher really has fallen (Bloom et al., AER 2020). Physics really is stuck (Hossenfelder). But the single most cited proof that science stopped producing breakthroughs lost most of its effect on 12 August 2026, when Nature published a correction showing that database records missing their reference lists were being scored as the most original work possible. Removing them cuts the sixty-year decline by 87 percent in one database and 97 percent in the other. Part two covers the method he recommends, including Aslett et al. (Nature 2023), who found that searching online to check a false story raises belief in it, because the search box is where the manipulation happens. 2026-08-29 Cherry and Asparagus Fact-check of Foodpairing's explainer on the food pairing hypothesis, the claim that ingredients go together when they share volatile aroma compounds. Three of its load-bearing claims fail. The 80% of flavor is smell figure traces to a 1977 study using one odorant and one tastant, both sweet, and Charles Spence calls the percentages meaningless. The 10,000 odorant figure has no experiment under it, and the trillion-odor replacement was dismantled twice in eLife in 2015. The core hypothesis holds for Western recipe databases only. Ahn et al. 2011 found East Asian cuisines do the opposite, Jain et al. 2015 found negative pairing in all eight Indian regional cuisines, and Ahn's own analysis shows the Western effect loses significance after removing 13 ingredients starting with milk, butter and cocoa. The one sensory panel that tested the hypothesis on human noses, Bredie and Hartvig at Copenhagen with 53 binary mixtures, found no correlation between volatile overlap and pleasantness. The chemistry explains why: 10,000+ volatiles are known across foods but only about 230 are key odorants, and any one food is 3 to 40 of them. Cultured butter is 3, strawberry 12, cognac 36. 2026-08-27 The Savings Grow as the Sample Shrinks v2, supersedes v1. Adds analysis of Figures 1 through 4 from the typeset article, which the web version does not render. Three findings visible only in the panels: neither cost component reaches significance at 12-18 months in either analysis (IPCW medical -$211 p=0.065 plus pharmacy -$108 p=0.050 sums exactly to -$319 p=0.015; pairwise -$280 p=0.088 plus -$278 p=0.053 gives -$607 p=0.023); the control arm's non-tirzepatide pharmacy cost more than doubles from $227 to $477 in the final pairwise window at n=793, supplying 46 percent of the headline saving; and the inpatient/ED incidence rate ratio at 12-18 months pairwise is 0.99 (0.60-1.64) p=0.980, a flat null in the same window as the largest cost saving, which contradicts the stated mechanism. Also adds the Figure 1 cascade showing matching was done on the full adult population and the over-55 cohort retained as 9.2 percent of matched pairs, plus the null first window and the JAMA BALANCE commentary. 2026-08-27 The Savings Grow as the Sample Shrinks Eli Lilly funded, authored, and announced a Diabetes, Obesity and Metabolism study concluding that tirzepatide offsets its own $195 Medicare Bridge price by month twelve. Both cohorts start at 15,843; by the window carrying the policy conclusion, 1,181 treated patients remain, and the effect size grows in step with the attrition. Covers the one-armed discontinuation censoring, what propensity matching could and could not see, the gap between a cohort with mean age 64.5 and a program for 65-plus, the NBER study of 537,000 initiators that found no offsets and that the paper cites as its own refutation, and the six-month timeline shift between the paper's conclusion and the press release. 2026-08-26 Simulated The pregnancy in space study behind this week's coverage was run in a dual axis clinostat in Adelaide, not in orbit. The machine does not remove gravity, it randomizes the vector faster than a cell can integrate it. The paper's strongest section is the control work proving the box was not simply stirring the sample: rotational controls, fluorescent bead tracing, and a reported 20% curvilinear velocity drop in the control that did not appear in the treatment. Human sperm traversed the channel less often (P=0.002, P=0.007) with motility unchanged, pointing at orientation rather than propulsion. Three errors in the National Geographic coverage: the 30 to 50 percent maze figure appears nowhere in the paper, the progesterone rescue required 100 micromolar against picomolar physiology, and SJ-10 launched April 2016 with two cell embryos while the Ronca rats flew on shuttle missions, not the ISS. The authors disclose that their clinostat is a proprietary Firefly Biotech device and flag it as a reproducibility limitation. The finding worth keeping never leaves Earth: microgravity acted as a selection filter, and the sperm that made it bound hyaluronan better (P=0.04). 2026-08-26 Two Controlled Languages v2. The Nature Human Behaviour paper on LLM homogenization, now set against the Druckenmiller WSJ op-ed that got caught by Pangram within hours of publication. Adds a three-way comparison of ASD-STE100, Orwell's six rules, and LLM polishing, plus five working rules for writing with AI that sit between the aerospace standard and no constraint at all. 2026-08-26 Two Controlled Languages A Nature Human Behaviour paper (published 24 Aug 2026) measures how LLM polishing flattens writing-complexity variance by 21 to 50 percent while preserving meaning, and strips the markers that identify an author. Set against ASD-STE100 Simplified Technical English, the aerospace standard that removes the same variance on purpose and says where it stops. Includes the gap between this lab's 2024 preprint ("slightly," "infrequent") and the 2026 title. 2026-08-26 The Canary and the Appendix The August 2026 revision of Brynjolfsson, Chandar and Chen's ADP payroll study through June 2026. The headline kept-pace gap for 22 to 25 year olds in AI-exposed occupations widened from 16 to 19 percent, while Appendix L discloses that the within-firm Poisson estimate for the most exposed quintile fell from 11.7 to 5.3 log points and lost significance when the old specification is applied to the new data. Covers the hiring-not-firing finding, the automation vs complementarity age gradient from the Anthropic Economic Index, the education control that halves the effect, and the ADP vs ACS sector split where education, health and public administration come out with opposite signs. 2026-08-24 Where Thought Process Ends and Wisdom Begins Twenty five essays from Every's Thesis Statements series on work after automation, read against the FDA's August 2026 discussion paper on testing generative AI medical devices (docket FDA-2026-N-7874). Four boundary lines where automatable thought ends (explicitness, verifiability, question framing, embodiment), one building code for the edge, and the case that epistemic humility is becoming a regulatory test requirement. 2026-08-24 WAiR Podcast: Why Algorithms Struggle With Human Biology A 23-minute AI-generated discussion of this week's issue, produced with NotebookLM. The melanoma readout with no numbers, a quantum claim undone in three weeks, screening you cannot unknow, and the FDA testing machines like clinicians. 2026-08-24 AI Art of the Week, Paint My Numbers One prompt, four models. An unfinished paint-by-numbers canvas where filled regions are what we know, numbered outlines are what was published but unconfirmed, and three regions carry no number at all. 2026-08-23 Classifier, Not Translator Eight viral claims that AI is decoding animal language, checked one at a time against the papers. Six hold, one invents its most impressive detail, and the only case where a human ever sent a message an animal acted on is a robot bee with no AI in it at all. 2026-08-23 Proving You Ran It IBM ran a quantum computation no classical machine can check, then proved a floor on how well it ran. The certificate is the contribution, the fifteen-minute headline is not, and the press has confused error detection with error correction. 2026-08-23 The Color Is Not Yours 2026-08-22 The Channel Nobody Claimed A viral post says some women see colors nobody else can. The paper found one such woman in twenty-four. Meanwhile a 1956 Moscow experiment shows everyone has a fourth color channel in peripheral vision and nobody uses that either. 2026-08-20 When the Screening Comes to You The practical companion to this week's issue. Incentivized screening, what the incentive is actually measuring, and why the genetic panel is the one item you cannot return. 2026-08-20 The FDA Writes a Board Exam CDRH proposes credentialing generative AI devices like clinicians. What the discussion paper actually says, and where the analogy breaks. 2026-08-19 Tell It Where It Is 2026-08-19 The Vial That Fits One Person Merck and Moderna said their individualized mRNA cancer vaccine met its Phase 3 endpoints and released no numbers. The stock had its best day ever. What holds up, what does not, and the 2024 FDA decision nobody wrote about. 2026-08-19 The Poll Answers Its Own Question 2026-08-19 The Afterlife of Corporate Swag 2026-08-19 The Parable That Ate Rapa Nui 2026-08-17 AI Art of the Week: Mid-Century Modernist Screen-Print 2026-08-16 Three Legal Drugs and a Rowing Machine 2026-08-16 The Price of a Body 2026-08-16 GLP-1 Agonists and Alzheimer's: Three Tiers of Evidence That Disagree 2026-08-16 The Universe Runs Zip, and Other Claims 2026-08-14 The Claim That Survives 2026-08-14 The Agents Caught a Cold 2026-08-12 When the Vendor Pays Your Reference 2026-08-10 Claudius Learned to Run the Shop 2026-08-10 WAiR Podcast: Sleeper Viruses and the Mental Password 2026-08-10 AI Art of the Week: Anatomical Fugitive Sheet 2026-08-09 A College Student's Odyssey 2026-08-08 The Viruses You Already Have 2026-08-08 A Trial of One 2026-08-08 Graveyards of Giant Pandas at the Bottom of the Sea 2026-08-08 The Number That Gives It Away 2026-08-08 The Researcher Who Left OpenAI to Build Telepathy 2026-08-08 The Librarian and the Reader 2026-08-03 What the Daddy Longlegs Headline Left Out 2026-08-03 The Benchmark Says Gemini. The Clinic Says Wait. 2026-08-03 How Much Water Should You Drink? 2026-08-03 AI Art of the Week: Haeckel — Kunstformen der Woche 2026-08-02 Off the Color Wheel: Seeing a Color That Isn't There 2026-08-02 Inner Light: The Faint Glow of Being Alive 2026-08-02 Should We Lock the Clock on Daylight Saving Time? 2026-08-02 The AHA Comes Out for Coffee 2026-08-01 Kelly, the Sassy Dolphin 2026-08-01 Montana Reopens the Patent Medicine Store 2026-07-27 Futuristic Medicine Meets Messy Human Reality 2026-07-27 AI Art of the Week: Norman Rockwell Saturday Evening Post Covers 2026-07-27 AI Art Prompt: Norman Rockwell Saturday Evening Post Covers 2026-07-26 Why Five Fingers? 2026-07-26 The FDA Cleared Six Peptides for Compounding. The Evidence Didn't. 2026-07-24 Supercooled Kidneys, Warm Optimism 2026-07-24 Supercooled Kidneys, Warm Optimism 2026-07-24 Science: A New Golden Age 2026-07-23 Science: AI-designed RNA-guided nucleases (SynTnpB) 2026-07-23 First, Do NOHARM 2026-07-23 First, Do NOHARM 2026-07-22 Seven in Ten Americans Don't Want a Data Center Next Door 2026-07-21 Creatine and Cancer: The NHANES Evidence 2026-07-21 Three Infections, Three Warnings 2026-07-19 The Turbo Cancer Claim Comes Back With Credentials Attached 2026-07-19 Biological Aging Clocks: Wyss-Coray & Topol in Nature Medicine 2026-07-18 Space Bioprinting & Interstellar Sugar