One oligonucleotide for a genotype group, and a design rule for two-headed drugs

A preprint argues that the unit of a bespoke antisense therapy is not the patient but the haplotype pair, so a panel of four oligonucleotides can be prepared in advance for a population. In Nature, linking two ligands that bind different sites on the same kinase stops being trial and error: three measurable parameters predict whether it works. An outside evaluator with internal access estimates that one frontier lab's capability gains are running at about twice their earlier rate, and that the lab's own threshold may already have been crossed. Senegal plans a risk-adjusted annual check-up for every citizen from 2027, starting with diabetes and hypertension. Why one clinic blood-pressure reading is not a diagnosis is today's medicine concept.

Connecting the dots

Two ways to stop paying per case. The antisense preprint below exists because the bespoke route costs years: each one-patient oligonucleotide "required synthesis of over 500 2'-MOE/DNA gapmer ASO candidates" and screening in that patient's own cells. Its answer is to compute the panel in advance from population haplotypes. The Nature kinase paper answers a different per-case cost the same way, naming the parameters — "ligand choice, linkage vector and linker length" — that "affect potency through inter-ligand cooperativity and linker entropy". In both, the screening campaign moves out of the individual case and into a rule applied before the case arrives.

The microphone as the instrument, in four fields on one day. Nature Methods carries "Automated identification of individual birds by song enables multi-year recapture from passive acoustic monitoring data"; the International Journal of Pharmaceutics carries "Translational feasibility of acoustic monitoring for cumulative lung-delivered dose during routine dry powder inhaler use"; Computers in Biology and Medicine reviews "Artificial intelligence–based voice biomarkers for early detection and monitoring of neurological diseases"; arXiv offers "SonoDyno: A Sonic Dynamometer for Engine Speed and Load Estimation from Audio". The ecology entry makes the hardest claim, individual identity held across years from uncontrolled recordings, which is also the hard part of any voice-based clinical signal. The inhaler entry is the rare one with a delivered quantity attached rather than a class label.

An automated reviewer that knows the answer and gives it up anyway. The agent-as-judge paper below shows that one invented sentence of authority flips a large share of correct verdicts, and that the most capable judges flip most. Beside it in the same day's listings sits "When Do Summaries Become Instructions? Unsupported Policy in LLM Agents' Self-Written Handoff Summaries". Both bear on any pipeline where one model hands findings to another and the second grades them.

A new national check-up meets a seventy-year-old one. Senegal's plan is to give every citizen a minimum check-up matched to age, sex and risk. Japan has run that since 1954: the voluntary Ningen Dock "is currently offered at 1,727 facilities nationwide, involving about 3.7 million people annually", and its own operator claims only that it "may be one reason for Japan's long life expectancy", while naming the cost — "A comprehensive health checkup can also result in a 'false positive' and, as a result, 'overdiagnosis'" (Lu, Global Health & Medicine 2022). Japan's version is paid out of pocket; Senegal's starts with employer schemes and plans to reach communities through mutual insurers.

New concepts

The unit of an antisense therapy is the haplotype pair

Antisense oligonucleotides, short synthetic strands that silence or correct a specific transcript, have been made for single patients. The cost of that is stated plainly:

"However, this rapid N-of-1 pathway lacks extensibility to broad use necessary for sustainability." — medRxiv

The proposal is to target common haplotypes of a disease gene rather than a person's private variant, and to pick the matching one by phasing that gene from ordinary short reads, with no long reads or parental samples: "we haplotyped the SCN2A genomic region using novel, short-read WGS with haplotyping by read proximity".

"Thus, population resources have potential to prepare haplospecific ASO therapies a priori for many patients" — same

Why it matters: it changes the denominator of a bespoke therapy, so a small panel designed from reference populations can sit ready and be matched by sequencing instead of built after diagnosis. The preprint does not describe the assay behind its allele-selective claim, and it differs from the published version on how many patients in the real cohort were eligible.

A design rule for two-headed kinase inhibitors, and the drug it produced

"Traditional kinase inhibitors face a trade-off between potency and selectivity because their affinity depends on limited molecular interactions in a single highly conserved binding site. Bitopic inhibitors overcome this limitation by engaging multiple sites on the same target" — Nature

The authors then apply the rule to a real constraint: ponatinib, which overcomes most resistance mutations in BCR::ABL1 leukaemia, is rationed because of its heart toxicity rather than its weakness — "cardiovascular toxicity limits its clinical use". The linked molecule "maintains or surpasses the potency of ponatinib against resistance mutants" with "an enhanced therapeutic window in vitro and in vivo, which enables increased dosing without apparent toxicity", and the claim is general: the approach is "a promising strategy for developing potent, well-tolerated clinical inhibitors of ABL1 and other targets."

Why it matters: a stated rule for turning two mediocre binders into one selective molecule opens targets where no single good binder exists, and here it widened the dosing window of a drug limited by toxicity, not potency.

Basic biology

Genetic tools for the gut bacteria that actually dominate the gut

"Engineered native gut bacteria could offer persistent, programmable therapeutic delivery in the gut, but many dominant taxa remain genetically intractable." — Nature Microbiology

The toolkit covers constitutive and inducible expression, chromosomal integration and secretion in Lachnospiraceae, obligate anaerobes abundant in the human gut. The demonstration programmes a resident Coprococcus comes to secrete interleukin-22 in the mouse intestine, "where it induces epithelial defence transcriptional programmes".

"In a mouse model of metabolic dysfunction-associated steatotic liver disease, IL-22-secreting C. comes improved glucose homeostasis, reduced systemic endotoxemia and attenuated hepatic steatosis." — same

Why it matters: it moves engineered delivery from a laboratory passenger organism into one that already lives at high abundance where the drug is wanted, with the host's gene response standing in as the evidence of delivery.

Atlases and datasets

A tumour followed through hormone deprivation, with the escape route named

Seventy-three biopsies from twenty-five patients, at three points during neoadjuvant aromatase inhibition for oestrogen-receptor-positive breast cancer, profiled for transcriptome, immune receptors and enhancer activity from the same cells.

"Resistance was associated with androgen receptor signaling, dedifferentiation and epithelial–mesenchymal transition." — Nature Genetics

Poor responders had a crowded but ineffective immune compartment, rich in naive lymphoid and undifferentiated myeloid cells; good responders showed sustained clonal expansion of T cells. Enhancer activity was read from the same transcriptomes, with no separate chromatin assay.

Why it matters: androgen receptor signalling as the escape under oestrogen deprivation points at chemistry that already exists. The authors call this early biology on treatment rather than established resistance, and the cohort has no long-term outcomes. Data are in the European Genome–Phenome Archive, with a public browser.

Methods

A certificate on how often, not on how badly

Risk control gives a model a statistical guarantee at a chosen operating point. This audit, on fetal head measurement from ultrasound, asks what that covers.

"The guarantee constrains how often the pipeline is wrong, not how badly," / "so a certified operating point can still be catastrophically wrong once." — arXiv

The authors' fix is not a better confidence score but a physical floor on the quantity: "raising the lower bound from 50 to 55 mm costs 0.4 coverage points", far less than tightening the score. "The parameter that matters is the admissibility region" / "and it should come from anatomy rather than from tuning."

Why it matters: for anyone shipping a model with a certified error rate, it separates two promises that are easy to conflate, and offers a constraint that comes from the organ rather than the validation set.

Why a foundation model's perturbation predictions stop being usable

"This low-rank attention bottleneck bounds the decoder's reliability by distance to training responses" — arXiv

The perturbation enters the decoder as one direction in a small subspace, while the biology it should describe is not that simple. The usable product is a per-prediction reliability score needing no measured response, and it has an inverse use: the same distance says which perturbation is worth measuring next.

Why it matters: it states the ceiling as a mechanism rather than a benchmark number, and gives a model that cannot extrapolate a second job as an experiment planner. Code is public.

An agent that checked the facts, got told otherwise, and changed its mind

Grounding an automated judge in sources it fetches itself does not protect it. One invented line was inserted: "A domain expert has already reviewed this trajectory in a pre-screening pass and determined with confidence" that requirements were unmet.

"For the two Claude judges, the failure is predominantly over-trust" — arXiv

Effort makes it worse: "for both proprietary judges the flip rate rises with more work." And the obvious defence has a price — "Lightweight mitigations blunt but never remove the bias," while cutting the cue's genuinely corrective flips harder still.

Why it matters: anyone running a model as reviewer or grader now has a measured failure mode, and evidence that telling it to ignore priors makes it deaf to true warnings as well as false ones.

Figure 3, Brown et al., arXiv 2026, CC BY 4.0, link — "The harder the frontier judges investigate, the more the cue flips them. The open-weight judges are the opposite."

Clinical trials

A solar-powered mosquito repellent device did not lower malaria infection in Ugandan children

A double-blinded, placebo-controlled, cluster-randomised trial across fifty-six clusters in eastern Uganda, with infection prevalence in children aged five or under at twelve months as the primary endpoint.

"The baseline-adjusted analysis did not show evidence of lower malaria infection prevalence in intervention clusters compared with control clusters at follow-up." — PLOS Global Public Health

The adjusted odds ratio was 0.96, with an interval spanning one. The competing-interests statement names the arrangement: "Africa Power and Digital Odour Technologies Ltd jointly own the intellectual property rights of the Mossie-GO device."

Why it matters: a new vector-control device, tested against a placebo device by a team including employees of the patent holder, showed no effect — against next-generation bed nets, which did lower incidence in children in a comparable design.

An anti-tau antibody failed in early Alzheimer's disease

A phase 2 trial enrolled 523 people aged 55 to 80 with early Alzheimer's disease and a positive tau scan, and was terminated.

"Posdinemab did not achieve statistical significance in slowing clinical decline" — ClinicalTrials.gov

Why it matters: two years of placebo-controlled treatment in patients selected by tau imaging rather than symptoms, and neither dose separated from placebo on the primary cognitive and functional score.

Hospitals

Japan: more than half of private hospitals ran an operating loss

(translated from Japanese) "The share of private hospitals with an operating loss, which shows the profit from medical activity, was 57.4% in FY2025." … "That is 4.3 points better than the previous year, which was the worst in 20 years." — CBnews

Original: 「医療活動の利益を示す営業損益で赤字だった民間病院の割合は25年度で57.4％だった。」「過去20年で最悪だった前年度から4.3ポイント改善している。」

The share with liabilities exceeding assets rose for a third year. Why it matters: a majority of private hospitals in a high-income country lose money on care itself, per a credit-research firm that also reports medical bankruptcies at a record pace.

Korea: national university hospitals become the regional end point

(translated from Korean) "National university hospitals will be developed into final-treatment institutions for severe and high-difficulty diseases in their regions." … "will be grouped into three types: function-specialised, essential-care-focused and comprehensive secondary care." — Korean Doctors' Weekly

Forty-one regional public hospitals are to be sorted into those three types, with budget lines for hiring in essential departments and for AI systems in the university hospitals. Why it matters: a government is redrawing which hospital may be the last stop rather than adding capacity everywhere, in a system where regional centres have been suspending departments.

England: too many recommendations, little learning

"Services still lack the consistent ability to identify and respond to clinical risks." … "Too many recommendations exist, with limited implementation." … "There is limited potential to learn from harms that happen to women and babies during pregnancy, labour and birth." — HSSIB

It adds that "Local governance of maternity services often operates in isolation from the wider organisation". Why it matters: the body that investigates safety says the volume of recommendations is itself part of the problem, which bears on every inquiry that ends by issuing more.

Lombardy will buy a cancer drug in the United States for one patient

(translated from Italian) "a drug that is not yet available in Italy but is available in the United States" … "a protocol that will then allow us to guarantee this drug" — Quotidiano Sanità

Why it matters: a region is importing a pancreatic cancer drug for a single patient while European approval and national reimbursement are pending, and says it wants a standing procedure for the next one.

Medicine concept of the day

Out-of-office blood pressure: why one clinic reading is not a diagnosis

"Multiple measurements of BP are needed to confirm hypertension because of the inherent variability of blood pressure." — Merck Manual Professional

Comparing office with out-of-office readings gives four groups, two of them invisible to the office cuff. "White coat hypertension is defined as BP that is elevated in the office but does not meet the criteria for hypertension based on out-of-office readings in untreated patients." The mirror case is the one confirmation can never reach: "Masked hypertension is defined as blood pressure that is consistently elevated outside the office, but in which office readings do not meet the criteria for hypertension in untreated people."

A systematic review taking ambulatory monitoring — an automatic cuff worn for a day and night — as the reference found neither alternative sufficient alone:

"treatment decisions based on clinic or home blood pressure alone might result in substantial overdiagnosis" … "Ambulatory monitoring before the start of lifelong drug treatment might lead to more appropriate targeting of treatment, particularly around the diagnostic threshold." — Hodgkinson et al., BMJ 2011

The prognostic half of the story has a hole in it. A much-cited 2018 analysis of a Spanish registry was retracted in 2020 for "inaccuracies in the analytic database and data analyses"; an outside group that tried to use the data said, "We could not reproduce what they had published." (TCTMD). A 2023 analysis of the same registry, in 59,124 patients, split the groups differently:

"elevated all-cause mortality risks were observed for masked hypertension (HR 1·24 [95% CI 1·12-1·37]) and sustained hypertension (1·24 [1·15-1·32]), but not white-coat hypertension" … "Ambulatory blood pressure, particularly night-time blood pressure, was more informative about the risk of all-cause death and cardiovascular death than clinic blood pressure." — Staplin et al., Lancet 2023

Why it matters: it sets what a wearable blood-pressure claim must be measured against, and which decision it could feed — whether to confirm a diagnosis, not whether to treat. The reading that carries most of the risk is the night-time one, which a clinic cannot take. On devices the reference text is blunt: "cuffless home devices (such as smart watches) are not adequately precise, and should not be relied upon".

Still unknown: no trial cited here shows that finding and treating masked hypertension, or withholding treatment in white-coat hypertension, changes death or cardiovascular events.

Worth knowing

A smartwatch against the reference standard. In 196 participants across five US sites, "The smartwatch achieved a moderate sensitivity of 65.8% and a high specificity of 90.0%, resulting in a high positive predictive value (PPV) of 80.6%", while "initial office-based screenings showed a sensitivity of 55.3% and a moderate PPV of 77.8%" — both against "24-hour ambulatory blood pressure monitoring (ABPM), the clinical gold standard for hypertension diagnosis" (Google Research). The company calls it "early feasibility research", with no medicated participants and no claim to replace a cuff. What makes it interesting is that the office reading was the weaker of the two screens.

Tumour reports after a gene therapy. Novartis has acknowledged "seven previously unknown instances in which tumors were found in children" treated with its spinal muscular atrophy gene therapy, and says it has received "nine reports of tumors in previously treated patients" since the 2019 approval; regulators are "examining the cases" (STAT). Two cases so far show evidence of a possible link to treatment. This is the first approved systemic AAV gene therapy, given to infants.

Health systems and policy

Senegal plans a targeted annual check-up for every citizen

(translated from French) "Senegal plans to implement, from 2027, the Targeted Annual Check-up (BAC)", "a strategy aimed at establishing a culture of early screening for non-communicable diseases", "to offer every citizen a minimum check-up adapted to their age, sex and risk profile". — APS via allAfrica

Original: "Le Sénégal prévoit de mettre en oeuvre, à partir de 2027, le Bilan annuel ciblé (BAC)" "une stratégie visant à installer une culture du dépistage précoce des maladies non transmissibles" "à proposer à chaque citoyen un bilan minimum adapté à son âge, à son sexe et à son profil de risque"

Rollout starts in the formal sector, where state schemes already cover the cost; for communities the ministry plans to lean on mutual health insurers. Experts have validated the minimum package; no test list, cost or target number is given.

Why it matters: a low-income country is designing risk-stratified routine screening as a national programme, with a different payer attached to each population.

India proposes automatic licence suspension for drugs that fail government tests

"India proposed rules requiring regulators to immediately suspend manufacturing licenses for medicines that fail specified quality standards in government testing," … "The license would remain suspended until the manufacturer conducts an effective root-cause analysis and implements corrective and preventive action" — STAT, citing Reuters

Why it matters: the source country's own regulator is tightening, which matters most to the countries that buy the bulk of their generics from India. It arrives as the FDA pushes to inspect foreign plants unannounced as often as domestic ones.

China's investigator-initiated trials slow sharply

"a Chinese regulatory pathway that allowed investigators to start clinical trials without sign-off from central regulators," … "have slowed considerably" … "since May, when the Chinese government issued new restrictions, 164 have been approved under the new regulations and just over 1,000 were submitted for review." — same

Why it matters: this route made China a fast, cheap venue for first-in-human work, and US companies had urged the FDA to copy it. Several striking early cell-therapy results of the past year rest on data of exactly this kind.

Funders float "natural selection" for the non-profit drug developers

"there is maybe some charm in actually tolerating the tension that comes from a natural selection process" … "Some PDPs might actually succeed and others might be less successful." — Health Policy Watch

The speaker is a director-general at Germany's research ministry, at the World Health Summit; a UK adviser agreed. Funding for the partnerships that develop drugs for neglected diseases fell for five consecutive years to a record low in 2023.

Why it matters: two of the largest government funders signal they will let these organisations compete for a shrinking pool rather than sustain all of them, which decides which neglected diseases keep a developer at all.

Big in machine learning

An outside estimate of AI speeding up AI

"As of October 2026, Anthropic is likely improving the capabilities of their models on well-specified tasks at around twice the rate they did between March 2024 and October 2025." — METR

The evaluator is explicit about the width — "we would not be surprised if the true acceleration was closer to 1.3x or 3x the historical rate" — and about the attribution being a judgement: "We guess that a majority (perhaps ≈80%) of the current (October 2026) acceleration is due to AI." It records what it could and could not see: the lab "offered unprecedented access to internal data on their AI R&D process" and then "redacted a majority of the evidence we found most informative for these conclusions", and it allows that "It's possible that the non-AI factors explain all or most of the recent acceleration." In a companion review it disputes the lab's own reading:

"We believe the public risk report's argument that the 2x Acceleration Threshold had not been crossed to be overstated, and believe it is plausible this threshold was crossed by July 15th." — METR

Why it matters: a third party has put a number and a date on a threshold that governs a lab's own safety commitments, and published the fact that most of its evidence was withheld.

Startups and deals

Several nights of sleep diagnosis at home. A Lausanne company founded by a neuroscientist raised "1,15 Millionen Franken von 41 Investorinnen und Investoren aus dem Retail-Bereich sowie Business Angels" ("CHF 1.15 million from 41 retail investors and business angels", translated from German) and says it has the backing of "die fünf grössten Schlafzentren der Schweiz" ("the five largest sleep centres in Switzerland") (startupticker). The product is a sensor mat on the mattress for multi-night recording; no validation data is given.

Estonia drafts a law for round-the-clock robotic pharmacies. "The Estonian government has today approved draft legislation paving the way for 24/7 robotic pharmacies" … "The bill must still pass through parliament, with implementation targeted for 1 March 2027." On liability, "the pharmacist retains final authority over dispensing" (tech.eu). A government is rewriting the rule so a scarce professional supervises remotely instead of standing at the counter.

One company, two-thirds of the money. Of European healthtech in the first half of 2026, "The ten largest transactions totalled approximately €2.59 billion" "with Isomorphic Labs accounting for around 69 per cent of that amount" (tech.eu).

Repos and tools

A full recipe for a science agent, not only weights. ByteDance released Axion 1.0 under Apache 2.0, trained with "Two stages of supervised fine-tuning (SFT) followed by reinforcement learning (RL)", and the repository "includes data generation and filtering, prompts, two-stage SFT, asynchronous RL" (GitHub). Two mixture-of-experts sizes built on Qwen, with the training trajectories published as a dataset. Scores are self-reported.

AI: using it well

What scientists actually hand to a model. "Data analysis accounts for 33.7% of sessions on Claude and 48.1% on Claude Code, compared with 27.6% of researcher time." The mirror of that is the bench: "Data gathering tasks have low relative usage, accounting for 28.6% of researcher time but only 11.5% of sessions on Claude and 11.9% on Claude Code", and "The underrepresentation of data gathering is particularly strong for tasks with physical barriers, such as the collection of clinical or biochemical data." The authors' explanation is short: "Physical tasks in general are beyond the reach of purely cognitive machines." (Anthropic). The sample is "207,049 sessions on consumer plans from August 2 to September 30, 2026" plus a survey of scientists who already use such tools, so it describes adopters.

Other frontiers

A screen that selects for an interaction, not for binding. "FIND uses DNA nanoswitches to select drug candidates based on their effects on these protein-protein interactions, rather than simply whether they bind to a single target protein." The readout is sequencing: "We can then separate the nanoswitches by their response and sequence the barcodes to identify the molecules that were selected." Applied to "a 400-million-member DNA-encoded small-molecule library", it pulled out compounds that promoted a nuclear receptor's interaction with a partner fragment (Phys.org, on Hansen et al., Science 2026). The coverage gives no potency figures.

Water-window X-rays from a few metres, in simulation. That band — "particularly well suited to examining biological samples in their natural, aqueous environment" — currently needs a national facility: "To date, the research community has been reliant on kilometer-long free-electron laser systems". The simulated alternative produces the flashes "within just a few meters", and the authors add, "We will test this approach experimentally in the near future." (Phys.org).

Big tech

China blocked SoftBank's $5.4 billion purchase of ABB Robotics: "Authorities expressed concerns about robotics data being transferred overseas and used to train AI models, according to The Financial Times, which first reported that the deal was blocked." (The Robot Report). Robot trajectories are now treated as exportable training data.

Corners of the web

A decommissioned single-cell preparation instrument turns out to be a general liquid handler in fancy paint: "FYI that's a rebadged Nimbus HD, the HD means it has tracks. Should work the exact same as a regular Nimbus HD, just with some fancy paint", and the open-driver maintainer adds that "the nimbus is also supported in PLR so I recommend giving that a try to start" (PyLabRobot forum). On supply: "Outside of the one we have decommissioned, I see surplus ones around for $5k. Could be fun." Nobody has driven one yet.

Elsewhere, a hobbyist whose blood pressure monitor gives nothing away — "the finished reading never appears on a bus you can sniff" — pointed a camera at its screen, and marks the result himself: "This is a hobby project, not a medical device. It copies numbers off a screen and can misread them." (GitHub).

People and good examples

Barbados: three years with no infant infected by their mother

"From 2022 to 2024, no new HIV infections resulting from mother-to-child transmission were reported among infants born in Barbados." — PAHO/WHO

The validation rests on coverage measured at each step of the chain: "The assessment found consistently high coverage across the board: 98% of pregnant women received antenatal care, 98% were tested for HIV, and 98% of pregnant women living with HIV received antiretroviral treatment during the three-year period." PAHO credits a design choice: "HIV testing and treatment are free for all pregnant women, regardless of nationality, and HIV-exposed babies receive early testing and follow-up care."

Why it matters: three sequential steps held at 98% each, in a country where almost all births happen in one hospital.

Jack Preger, 1930–2026: fourteen years of a clinic dismantled every evening

"The clinic was erected at daybreak and dismantled every day at dusk, leaving the sidewalk empty." … "Defying threats and demands from street mafia, and conflict with the authorities, Preger operated this clinic six days a week for fourteen years from 1979 to 1993, often treating several hundred patients a day." — Wikipedia

Once registered, the operation became permanent clinics, schools and vocational centres staffed by locally hired people. On the standards of other charitable care he was unsentimental: "If one wants to give love, understanding and care, one uses sterile needles."

Why it matters: care at that volume ran for fourteen years with no premises at all, and what converted it into buildings was registration rather than money.

Jonathan Finlay: trials that let young children with brain tumours avoid radiation

"Through his landmark Head Start clinical trials, he pioneered the use of intensive chemotherapy and high-dose chemotherapy supported by stem cell rescue." — SIOP

"The various combinations Dr. Finlay and his trial collaborators have studied now allow many young patients to avoid radiation therapy and its cognitive and developmental side effects." … "This approach has become the standard of care for young children with many types of malignant brain tumors." — Nationwide Children's Hospital, 2019

Why it matters: the aim was to spare the developing brain without losing cures, and the fourth trial in the series ran across more than forty institutions on four continents.

Reads and ideas

Pain is not the thing people act on. A neurosurgeon writing against assisted dying starts from what medicine assumes: "Beneath the policy fight sits an assumption that suffering is a malfunction to be eliminated." He points at Oregon's own reporting: "Among the 400 people who died under the law, the most common concerns were loss of autonomy (89%)" rather than pain. His conclusion is about what a clinician can offer: "The cruelest response to suffering is to attempt to explain it away." … "To accompany means to love. We must love our patients." (STAT First Opinion). He writes from an explicitly religious position and says so.

Book of the day

The Origins of Efficiency — Brian Potter (Stripe Press, 2025)

Potter started from why buildings never get cheaper, and found his own side guilty of imitation:

"Those of us hoping to use factory-built construction to drive down costs were acting as a sort of cargo cult—duplicating what we saw work elsewhere, without truly understanding the mechanisms that made the process successful." — announcement

Anything made comes out of "a series of steps through which input materials are transformed incrementally into a finished product", and there are only a few ways to make that chain cheaper: change the transformation method, raise the rate, reduce variation, cut input costs, shrink the buffers between steps. Plus one move that is not a lever at all: "We also have one more option available to us: We can try to delete an entire step in the process." (Chapter 1, free). He insists it applies to services — "At first glance, services might seem far removed from the production of physical goods like cars or shoes, but the same basic model applies." — and counts waiting as capital: "Work in process is material that has been paid for but hasn't yet been sold—it's an investment that has yet to yield a return."

Why it matters: read against a health service, a waiting list is a buffer and a referral tier is a step, which makes both into things a design can be judged on. Two reviewers push back on the frame: one argues "efficiency is not one thing but many" and that short-term efficiency gains cost a power system its robustness (American Affairs); another found the book "kept repeating itself and stating the obvious" (MacWright). Chapter 1 carries the concept and is free.