Writing

Every article, case study and report here is an analysis built from primary documents -- filings, trial registries, patent offices, a plan's own characters -- with each figure endnoted to the document it came from, so a reader can check any number rather than take it.

The pieces below -- biotech industry analysis articles, case studies, one translation and five company reports -- are grouped by the line of work each demonstrates, newest first within each group. If you want the size of an industry, start with the biomanufacturing pieces, where eight analysts' estimates for one year run $19-40.1 billion and the filings tell a tighter story than the forecasts; if you want how the industry works, the review covers sixty years, every production stage and every product class.

Diligence and landscapes

  1. China's pharmaceutical industry plan for 2026-2030, side by sidetranslationThe full text of 医药工业发展“十五五”规划 (工信部联规〔2026〕210号), the Chinese quoted above every English paragraph. Unofficial. Every number, threshold, and negation machine-matched to the characters; one cross-family adversarial review; two source-internal inconsistencies left visible.
  2. The consolidation nobody priced incase studyEight analysts disagree by 2.1x on how big biomanufacturing is. None of them disagree that Samsung Biologics' margin hit 45% last year, that Catalent stopped filing with the SEC, or that nearly 3,700 cell and gene trials are already committed to future capacity. The filings tell a tighter story than the forecasts do.
  3. How biomanufacturing actually works, and what it is wortharticleA living cell that has been talked into making a medicine is a strange kind of factory: variable where a chemical reaction is not, fragile where a pill is not, and licensed as a process rather than as a product. This is the long version -- sixty years of history, every production stage, every product class, the frontier graded honestly against a maturity scale, and what the filings, registries and patent office say the whole thing is worth.
  4. Biomanufacturing's blurry numberarticleEight analysts sized the same industry this year and landed between $19 billion and $40.1 billion. The gap is not sloppiness. It is a category whose edges are still being drawn, live, by workforce shortages, an onshoring boom and a regulator that has started promoting the technology it once merely tolerated.
  5. What a verified read looks likecase studyA hazard ratio of 0.16 is the kind of number that ends an argument. The work is not finding it -- it is establishing that it means what the press release says it means.
  6. Corbus Pharmaceuticals HoldingsreportCRB-701 ORR 47.6% in 2L head and neck. Zero peer-reviewed data. Pliant's αvβ8 lead boxes out CRB-913.
  7. Immunome, Inc.reportPhase 3 hazard ratio 0.16. NDA filed April 29. ASCO oral on May 30 fills in subgroup detail.
  8. Wave Life SciencesreportGalNAc cleared 11 µM where one LNP didn't. Beam's LNP cleared it harder.
  9. Compass TherapeuticsreportPrimary hit. Secondary hit. OS missed because 54% of placebo crossed over.
  10. Absci CorporationreportThe AI designs antibodies. The clinic opened eleven months ago.

Commercialization assessments

  1. Rounding errorsarticleRapid tests now tell doctors within hours which antibiotic will work against a bloodstream infection, yet six trials have found they save no lives. Two of those trials show where the time goes. Tests keep getting faster, but paradoxically, in today's hospitals the delay that slows the cure is no longer the technology. It is the operations around it.

The research engine

  1. The agreeable machinearticleA quarter of a million chats show an AI that argues with hackers and compliments copywriters. The people drafting compliance filings get neither.
  2. From pilot to protagonistarticlePharma is handing AI the decisions, not just the slides. Revisited seven weeks on, with a finding that moves the argument: the machine's characteristic error is no longer inventing sources but quietly leaving things out -- which is harder to catch, and strengthens the case rather than softening it.

How the work is checked

  1. Lost in summarisationcase studyChina's new plan for its drug industry has no full English text, so the English-language debate runs on Xinhua's summary instead. We measured 128 English articles about it. Six link to the plan; seven quote a Xinhua reporter's sentence as the plan's own words; and 84 of the 86 that cite its first-in-class target leave out what that target counts.
  2. Read the footnotecase studyYou may have read that 95% of companies get zero return from AI. MIT's report measured something narrower -- AI pilots, six months in, as reported by the users and executives involved. We traced four widely shared AI-returns figures back to the surveys behind them, and list five things to check before you repeat one.
  3. Chapter, verse and checksumcase studyAn unofficial English text of China's 2026-2030 pharmaceutical plan, 工信部联规〔2026〕210号, with the ministry's Chinese above every paragraph and the original file fingerprinted.
  4. Pipe dreamsarticleA pitch deck said two patents stood issued. The public record said both had expired. What changed is not that somebody finally checked -- it is what checking now costs.
  5. Every sentence truearticleA trial write-up can be accurate in every particular and still leave an expert reader with the wrong impression. This is not a theory: someone ran it as a randomised trial, on 300 clinicians, using real abstracts and their own de-spun rewrites. It is the one failure a citation check cannot catch.
  6. The shape agreed in advancearticleA schema is written for a reader who cannot ask a follow-up question. That constraint, not tidiness, is why schemas exist -- and it explains why a false sentence in a document is not merely hard to find but structurally invisible.

Also on the record