← The canon · AItopiaOrAImageddon?

The Human Genome Project declared complete — the capability that came in early and the medicine that did not come

moment · The International Human Genome Sequencing Consortium — twenty sequencing centres across the United States, the United Kingdom, Japan, France, Germany and China. Begun 1 October 1990 under James Watson at the NIH National Center for Human Genome Research and co-funded by the U.S. Department of Energy; led from 1993 by Francis Collins; the largest producing centres run by Eric Lander (Whitehead Institute/MIT), John Sulston (Sanger Centre, funded by the Wellcome Trust), Robert Waterston (Washington University in St. Louis) and Richard Gibbs (Baylor College of Medicine). Raced from May 1998 by Celera Genomics under J. Craig Venter · 2003

Something that happened and changed what people expected next.

Descends from The Asilomar Conference on Recombinant DNA Molecules.

The findings, stated up front so the rest of the file can be checked against them:

The engineering promise was kept early and under budget. The medical promise attached to it was missed by margins that are still being missed in 2026 — and the two facts have one cause, which is that sequencing a genome was the precondition and everybody including the people running the project talked about it as the outcome. "Complete" in April 2003 meant 99% of the gene-containing part; the remaining ~8% took until 2022, and John Sulston said in the room in February 1996 that it would. The cost curve is the most spectacular thing here and it is routinely quoted wrong, mixing a thirteen-year programme cost with a unit price. The second thread — the Bermuda Principles, daily public release of sequence, enforced not by a pledge but by a funder's renewal criteria — is a real precedent for the open-weights argument, the open side did plainly win it, and the analogy breaks in four places that the argument almost never names.

All of those are true at once, and a file carrying fewer than all of them is worth less than this one.

---

1. The kind, the id, and the descent

Filed under idea in canon/proposals.md, and idea is wrong. The proposal sits in the ## idea block (line 522, between the ## idea heading at 195 and ## limit at 722), and the filing does not survive its own citation sentence.

build.py defines idea as "A framing later work is built out of rather than argued about. Cited when the thing being watched descends from it and the descent explains its shape." That test asks for descent. Nothing a reading watches descends from the Human Genome Project. An AI capability announcement does not inherit from a DNA sequencing programme; it rhymes with one. google-index-1998 earns idea because the retrieval layer under this project's own searches is literally built out of PageRank. There is no equivalent sentence to write here, and writing one would be the kind of decorative ancestry this canon exists to avoid.

moment is the right kind on the house definition and the house standard. The definition is "Something that happened and changed what people expected next." The standard eliza-1966 argued its way out of, and deep-blue-1997, watson-jeopardy-2011, expert-systems-collapse-1987, asilomar-1975 and alphafold2-2020 have used since, is a date on which something visibly happened in public. This is that date. On 14 April 2003 the consortium announced completion; on 12 April 2003 the heads of government of the six sequencing countries issued a joint proclamation. And the expectation it set is the entire reason the entry is citable: it was that medicine was about to change, and medicine did not change on that schedule.

Why not prediction, which is the tempting one. More than half of what follows is graded claims with dates — Collins in 1999, Collins and McKusick in 2001, Clinton in 2000. But the kind requires "A dated claim", and the id names a project, not a claim. The honest resolution is that collins-mckusick-2001 is a missing prediction entry and somebody should write it; the table of milestones for 2010, 2020, 2030 and 2040 is one of the cleanest dated forecast sets in twentieth-century biomedicine and it deserves its own file, graded properly. §4 below grades what it can, because §4 is welcome in any kind, but it is not a substitute for that entry.

Why not interpretation. The frame — a capability milestone treated as the outcome rather than the precondition — is a real reading of events and it is the thing a reading would actually be reaching for. But alphafold2-2020 settled this: when the frame is still moving and the event underneath is fixed, anchor to the event. The frame here is moving (it is being applied to model benchmarks in 2026 by people who have never heard of CPS1 deficiency); the event is 14 April 2003 and it is not going anywhere.

On the id. human-genome-project-2003 names the project by its completion year, and the year is defensible but is not where the science happened. The draft was announced 26 June 2000; the analysis papers appeared 15–16 February 2001; the finished euchromatic sequence was published 21 October 2004. April 2003 was chosen because it fell near the fiftieth anniversary of Watson and Crick's 25 April 1953 paper. The date is a ceremony, and the ceremony is the point — this is an entry about what a declaration of completion does to expectations, so the ceremonial date is the right anchor as long as the file says plainly that it is one. It does, here.

The descent

asilomar-1975 is the ancestor, and it is load-bearing rather than decorative. Three distinct lines run from it:

1. The technical precondition. The recombinant-DNA techniques the HGP was built on — cloning into vectors, libraries, the whole apparatus of getting human DNA into a bacterium in pieces you can handle — were under a self-imposed moratorium from July 1974, and Asilomar is the meeting that lifted it under a rulebook. No lifting, no cloning programme, no HGP. 2. The governance move, repeated exactly. Bermuda, February 1996, is Asilomar's shape run again: the field's leadership convenes, agrees a rule that no law requires, and gets the funders to enforce it. §4.5 shows the enforcement sentence in the primary document. 3. ELSI. Watson allocated 3% of the HGP budget to ethical, legal and social implications at the project's start, later raised to 5%. A research programme pre-committing a fixed budget share to studying its own consequences is an Asilomar descendant and has no precedent outside that lineage.

gain-of-function-2011 is not an ancestor — it is the sibling this entry needs, and it runs the other way. It is the same scientific community, fifteen years later, facing a publication-openness question where a misuse story could be constructed, and choosing differently. §4.5 argues that the pair — Bermuda and H5N1 — is worth more than either alone, because together they say what the field's record actually is, which is not "open always wins."

Ancestors with no id, named rather than invented around.

---

2. What it is

2.1 The project

The Human Genome Project formally began on 1 October 1990, jointly funded by the NIH and the U.S. Department of Energy, with the Wellcome Trust funding the Sanger Centre in the UK and national programmes in Japan, France, Germany and China contributing. Congress was told it would take fifteen years and cost about $3 billion in FY1991 dollars, with completion by the end of 2005. The goal was a reference sequence of the human euchromatic genome — the gene-containing part — not of every base a cell carries; the heterochromatic regions were understood from the start to be out of reach of the technology.

In May 1998 Celera Genomics was founded by J. Craig Venter with PE Biosystems and announced it would sequence the genome faster and cheaper by whole-genome shotgun, which turned a public programme into a race. On 26 June 2000 the two sides announced a "working draft" together at the White House; Clinton's line was "Today, we are learning the language in which God created life." The analysis papers appeared on 15–16 February 2001, the consortium's in Nature and Celera's in Science.

On 14 April 2003 the International Human Genome Sequencing Consortium announced completion. NHGRI's release states the terms plainly:

> more than two years ahead of schedule

> at $2.7 billion in FY 1991 dollars, significantly under original spending projections

> The finished sequence produced by the Human Genome Project covers about 99 percent of > the human genome's gene-containing regions, and it has been sequenced to an accuracy of > 99.99 percent.

with "less than 400 defined gaps" remaining against "about 150,000 gaps" in the June 2000 draft, and the disease-gene count going from "fewer than 100 human disease genes" in 1990 to "more than 1,400" in 2003. Collins, in the same release: "The medical advances promised from the project can now be significantly accelerated." Watson: "The completion of the Human Genome Project is a truly momentous occasion for every human being around the globe."

2.2 What "complete" meant, and who said so in advance

It meant 99% of euchromatin at one error in ten thousand bases. It did not mean the whole molecule. About 8% of the genome — every centromere, the short arms of the five acrocentric chromosomes, the ribosomal DNA arrays, most of the Y — was never sequenced by the HGP and was not sequenced by anyone for another nineteen years.

The gap closed on 1 April 2022, when the Telomere-to-Telomere Consortium published T2T-CHM13 in Science: a 3,055 million–base-pair gapless assembly of every chromosome except Y, adding nearly 200 million base pairs of sequence and 1,956 gene predictions of which 99 are predicted protein-coding, and described as removing "a 20-year-old barrier that has hidden 8% of the genome from sequence-based analysis." The Y chromosome followed on 23 August 2023 (Rhie, Nurk, Cechová et al., Nature): 62,460,029 base pairs, adding over 30 million base pairs to the reference, more than half of the chromosome having been absent from GRCh38.

This was foreseen, on the record, seven years before the announcement. The minutes of the Bermuda meeting of 25–28 February 1996 record, in the session on model organisms:

> John Sulston commented that completion of the human genome would be asymptotic and that > some features (such as centromeres) may be omitted by design.

That sentence is the entry's sharpest single fact. The most accurate prediction anyone made about what "complete" would mean was made in the planning meeting, by the person running the largest non-US centre, and it did not survive contact with a press release.

2.3 The cost curve, stated so it can't be quoted wrong

This is the most-cited number in the whole story and it is almost always cited as a comparison between two things that are not the same kind of thing. NHGRI itself warns: "It is difficult to cite precise genome-sequencing cost figures that mean the same thing to all people because … different researchers, research institutions, and companies typically track and account for such costs in different fashions."

There are three separate accountings and they should never be mixed:

The inflection is dated and NHGRI names it: the switch from Sanger to next-generation sequencing in January 2008, after which both of NHGRI's graphs show "the sudden and profound out-pacing of Moore's Law."

As of 2026 the retail floor for a 30× consumer whole genome runs roughly $169–$430, and three instrument vendors claim $100 or below — Ultima has announced an $80 genome and Element Biosciences described VITARI in February 2026 as "the first ever high-throughput benchtop sequencing system capable of delivering a high-quality whole genome at $100." Those are vendor claims about instruments, not audited prices anyone paid, and this file does not promote them.

On NHGRI's own unit series the fall is roughly 180,000-fold (2001 → 2022). Divide the programme cost by a 2022 unit price and you get about 5 million-fold, which is the number that circulates. Both are defensible; only one of them is a like-for-like, and it is not the big one.

---

3. Why a reading would cite it

3.1 The occasion the owner named

From proposals.md: "Cite when: a capability milestone is being treated as the outcome rather than the precondition — sequencing a genome turned out to be the easy half, and the gap between 'we can read it' and 'we can do something with it' is the best-documented instance of that gap there is."

That is right, and the documentation is why it is worth an entry rather than an analogy. Very few technological gaps have this: a dated capability announcement, an itemised list of what the announcers said would follow and by when, independent grading published at the ten-year and twenty-year marks, and a still-running record of what actually shipped. The prediction kind exists in this canon to build a base rate for how wrong forecasting runs. This entry is the closest thing the canon has to a base rate for a different and more useful quantity: how long the second half takes, after the first half is declared finished, in a field with money, talent and public attention all pointed at it.

The answer, for the single best case in genomic medicine, is: twenty years and seven months from the completion announcement to the first approved CRISPR medicine (8 December 2023), and delivery had not begun in earnest two years after that — about 60 patients treated worldwide as of 5 February 2026 (§4.4).

3.2 The concrete case for this project

The AI as accelerant lens in LENSES.md requires that a finding show AI actually used in the work and say what became possible that wasn't. That discipline is exactly the discipline the genomic-medicine press failed for twenty years, and this entry is what a reading cites when a finding fails it. Three shapes recur, and all three have a genome precedent with a date attached:

And there is an in-repo hook that makes this entry load-bearing rather than ornamental. METHODOLOGY.md §1e — "Thresholds are the scoring events — cost-and-team to capability" — proposes exactly this instrument and uses this exact example — "Genome sequencing from about $3bn to a few hundred dollars" — as its illustration of the abundance direction. That sentence divides a programme cost by a unit price (§2.3). The redesign is right about the instrument and the illustration is off by more than an order of magnitude in a way bin/ethos-check.py would not catch, because the check looks for fake precision and this is fake comparability. A session working on METHODOLOGY.md should take the 180,000-fold figure and the three accountings from §2.3. This file may not edit that document, and does not.

3.3 The second thread: Bermuda as an open-weights precedent

The owner's other citation occasion: "the Bermuda Principles required daily public release of sequence data, which is a direct precedent for the open weights argument and one where the open side plainly won."

It is a direct precedent, the open side did win it, and §4.5 grades the claim in both directions — including the four places the analogy breaks, one of which (the enforcement mechanism) makes the precedent more useful than the loose version, not less.

3.4 The admission test, answered

Can the "why cite it" paragraph be written honestly? Yes, and it does not depend on the analogy holding. Even if every AI parallel here were wrong, this entry would still be the canon's best-documented instance of a research programme meeting its stated engineering goal ahead of time and under budget while the benefit that justified the spending arrived two decades late and, so far, in units of dozens of patients. The reading occasion is not "AI is like the genome." It is: when someone tells you what a capability will produce and by when, this is what the record of that class of claim looks like when you go back and check.

---

4. What it got right, and what it got wrong

Not required for moment. Mandatory here, because the entry is worthless ungraded.

4.1 The project's own promise — a hit, with two asterisks

Claim (1990, to Congress): fifteen years, about $3bn in FY1991 dollars, done by end of 2005. Due: 31 December 2005. Verdict: hit, early and under. Declared complete 14 April 2003 at $2.7bn FY1991. On the terms the project set itself, in a class of programme where overrun is the norm, this is one of the cleaner large-science deliveries on record and the canon should say so as flatly as it says the rest.

Asterisk one: the mechanism was not the plan. The acceleration came from a competitor that did not exist in 1990 (Celera, May 1998) and a machine that did not exist in 1990 (capillary sequencers, from 1998). A plan that beats its schedule because the technology underneath it was replaced mid-flight is a different object from a plan that was well estimated — this is the same finding spacex-booster-reuse carries about schedules, from the other side.

Asterisk two: the sentence that reached the public was not the sentence the project delivered. "99% of the gene-containing regions" is scoped and true. "The human genome is complete" is unscoped and was false by about 8% for another nineteen years (§2.2).

4.2 The gene count — the field's central quantitative expectation, wrong by ~5×

The standing pre-project figure was around 100,000 protein-coding genes. The February 2001 papers cut it to 30,000–40,000; the 2004 finished-sequence paper to 20,000–25,000; the settled modern figure is about 20,000.

The best artefact of this is GeneSweep, the betting pool Ewan Birney ran at the EBI from 2000 to 2003. Over 460 entries, some above 200,000. It closed at the 2003 Cold Spring Harbor genome meeting and was won by Lee Rowen of the Institute for Systems Biology with 25,947, against a then-assessed 24,847; the $1,200 pot was split with Paul Dear (MRC) and Olivier Jaillon (Genoscope). The winning bet was still about 30% high against today's number.

Verdict: the field was wrong by a factor of about five about the thing it was counting, and it was wrong in the direction of assuming that a complex organism needs proportionally many parts. That is a general failure mode and it is worth carrying: a community can be excellent at building the instrument and badly wrong about what the instrument will show, without either fact contaminating the other.

4.3 The medical predictions — the graded set

Claim, 8 May 1999, Francis Collins, 109th Shattuck Lecture, published as "Medical and Societal Consequences of the Human Genome Project," NEJM 341:28–37, 1 July 1999. The lecture's device is a hypothetical patient, "John," a 23-year-old smoker, imagined in 2010, whose genetic testing reveals elevated risk of coronary artery disease, colon cancer and lung cancer. Quoted by Joyner and Paneth from the original:

> Confronted with the reality of his own genetic data, he arrives at that crucial > "teachable moment" when a lifelong change in health-related behavior…is possible

and the finding "provides the key motivation for him to join a support group of persons at genetically high risk for serious complications of smoking, and he successfully kicks the habit."

Claim, 7 February 2001, Collins and McKusick, "Implications of the Human Genome Project for Medical Science," JAMA 285(5):540–544, with a milestone table by decade. From NHGRI's own adaptation of it:

> By the year 2010, predictive genetic tests are likely to be available for many common > conditions

> By 2020 the impact of genetics on medicine will be even more widespread.

> The pharmacogenomics approach for predicting drug responsiveness will be standard > practice for many drugs.

> New gene-based "designer drugs" will be coming on the market for diabetes, hypertension, > mental illness, and a long list of other conditions.

> By 2020, it is likely that every tumor will undergo precise molecular fingerprinting

Claim, 2000, Collins, as quoted by Joyner and Paneth: "Over the longer term, perhaps in another 15 or 20 years, you will see a complete transformation in therapeutic medicine." Due: 2015–2020.

Independent grade, 2019. Michael J. Joyner (Mayo Clinic) and Nigel Paneth (Michigan State), "Promises, promises, and precision medicine," J Clin Invest 129(3):946–948, March 2019 — the sharpest published grading and the one this file leans on:

> We have passed the 2010 deadline and are rapidly approaching 2020, yet the "complete > transformation in therapeutic medicine" has not occurred.

> However, nearly two decades after the first predictions of dramatic success, we find no > impact of the human genome project on the population's life expectancy or any other > public health measure, notwithstanding the vast resources that have been directed at > genomics.

On "John" specifically: "We must presume that smoking itself, which increases the risk of lung cancer by 20- to 100-fold, depending on cell type, has somehow failed to provide this young man a teachable moment," and "no gene has been discovered since 1999 that conveys a relative risk of six for lung cancer."

Independent grade, 2018, from inside the government genomics establishment. The CDC Office of Public Health Genomics blog, 7 March 2018, "Dear John…": "There are still NO validated genetic tests that can help smokers kick the habit." Smoking fell anyway — 23.5% of U.S. adults in 1999, 15.5% in 2016 — by public health campaigns, not by genotyping.

Independent grade, 2021, by panel. Julianna LeMieux, Genetic Engineering & Biotechnology News, 3 February 2021, put Collins's milestones to specialists. On genetic literacy among clinicians, Laura Hercher: "In truth, we remain well short of the goal." On pre-prescription drug testing, Mary Relling: "Such testing remains relatively rare in the United States." On tumour fingerprinting, Jesse Salk: it "was actually more on the mark than he probably realized." Mixed, and mixed in the right places — the oncology prediction is the one that came closest.

The self-grade, and it differs — in kind, not only in degree. The prediction discipline requires the predictor's own grade beside an independent one and requires saying they differ. Here the difference is structural and worth naming precisely: Collins has never published a scored self-assessment. What he published instead, in Nature 464:674–675 on 1 April 2010 under the title "Has the revolution arrived?", is a reframe:

> genomics obeys the First Law of Technology: we invariably overestimate the short-term > impacts of new technologies and underestimate their longer-term effects

Two things about that. It is substantially correct — §4.4 is the evidence, and by 2026 the long-term half is visibly cashing in. And it is not a grade. It converts a specific missed date into a general property of technology, which is a move that cannot be scored and cannot be wrong. It is also Amara's Law (Roy Amara, Institute for the Future), restated without attribution and promoted to a "First Law."

This file's grade, from the dated record: the direction was right, every date was wrong, and the errors are not symmetric. The 2010 predictions failed almost entirely. The 2020 predictions failed on designer drugs and on pharmacogenomics-as-standard-practice, and substantially succeeded on tumour molecular profiling. The 2030 and 2040 predictions are not yet due and are currently trending better than the 2010 set did.

4.4 What actually shipped, 2003–2026 — the other half of the ledger

Rule 3 requires both cases before a needle. There is no needle here, but the same discipline applies to a grade, and a file that carried only §4.3 would be dishonest. The delivered record is substantial and it is mostly not what was promised.

Casgevy is the one clean, traceable chain from the genome to an approved medicine, and it should be cited as such rather than as a general claim. Its target, the erythroid enhancer of BCL11A, was found by genome-wide association studies of fetal-haemoglobin levels in 2007–2008 — studies that were only possible because there was a reference genome and a variation map — and validated as a causal repressor by Stuart Orkin's laboratory. The editing tool that acts on it came from somewhere else entirely (bacterial CRISPR, 2012). The genome supplied the target; it did not supply the drug.

And the delivery record is the finding. As of 5 February 2026, more than two years after approval, about 60 patients across the U.S., Middle East and Europe had been treated, with stem-cell collection the reported bottleneck; Vertex booked about $43 million over the first two years (reported December 2025), with over 500 people reported as having initiated the treatment journey by Q1 2026. A one-time functional cure for a disease affecting millions worldwide, approved, reimbursed in several countries, and reaching dozens. The gap between "we can read it" and "we can do something with it" does not close at approval. It has a third stage nobody forecast at all.

Net grade. The medicine was oversold on timeline by ten to twenty years and oversold in kind — the promise was common-disease prediction and designer drugs for hypertension and mental illness, and the delivery has been rare-disease diagnosis, oncology profiling, a better prior for drug-target selection, and a small number of extraordinarily expensive one-time therapies. The people who said it would take longer were right about the schedule. The people who said it would not work were wrong. Both halves are needed, and a reading that cites only the first is doing the thing this canon exists to prevent.

4.5 The Bermuda Principles, graded — and where the open-weights analogy breaks

What was agreed, from the primary document. Report of the International Strategy Meeting on Human Genome Sequencing held at the Princess Hotel, Southampton, Bermuda, on 25th–28th February 1996. The panel session was chaired by Jim Watson, Bob Waterston and John Sulston, and Sulston and Waterston proposed the release principles. The endorsed text:

> Primary genomic sequence should be in the public domain. > > It was agreed that all human genomic sequence information, generated by centres for > large-scale human sequencing, should be freely available and in the public domain in order > to encourage further research and development and to maximise its benefit to society. > > Primary genomic sequence should be rapidly released. > > • Sequence assemblies should be released as soon as possible; in some centres, assemblies > of greater than 1 kb would be released automatically on a daily basis. > > • Finished annotated sequence should be submitted immediately to the public databases.

And then the sentence that states the purpose, which is the one that matters most here:

> It was agreed that these principles should apply to all human genomic sequence generated > by large-scale sequencing centres, funded for the public good, in order to prevent such > centres establishing a privileged position in the exploitation and control of human > sequence information.

The stated goal was anti-concentration. Not safety, not scientific culture, not reproducibility — preventing the producers from acquiring a privileged position over the resource. That maps onto the Concentration lens in LENSES.md exactly, and onto almost nothing in the AI safety literature.

How it was enforced, and this is the part the analogy usually loses. Not by a pledge. The same document records:

> It was noted that some centres may find it difficult to implement these principles because > of legal constraints and it was, therefore, important that funding agencies were urged to > foster these policies.

and, in the funders' session, the NIH National Center for Human Genome Research stating that grants awarded under its request-for-applications would be reviewed in December 1997, and

> The assessment criteria for renewal would include the amount of finished sequence > submitted to public databases. Early release of data and a high level of accuracy would > also be required.

Release became a renewal criterion. The Wellcome Trust, which had committed £84 million over seven years to the Sanger Centre (£60M of it for human sequencing and mapping), applied the same condition. The principles were in formal funder policy by early 1998. This is the transferable finding: the norm held because whoever wrote the cheque wrote the release policy. The AI analogue is therefore not a community statement or a company's values page; it is a condition attached to funding, compute allocation, or procurement.

Did the open side win? Yes, and the record is unambiguous.

Where the analogy to open weights breaks — four places, all of them load-bearing.

1. A sequence is not a model. The reference genome is a map. Weights are a working capability that anyone who downloads them can run. The Bermuda debate had no capability-transfer axis at all, and every argument imported from it inherits that absence. 2. The objection was commercial, never safety. Nobody at the Princess Hotel argued the reference sequence was dangerous, because it isn't. And when biology did meet a genuine capability-transfer argument, the same community chose differently — the 2011–12 H5N1 transmissibility manuscripts, which is gain-of-function-2011. So the field's actual record is not "open always wins." It is: open won where no serious misuse story could be constructed, and did not win where one could. Citing Bermuda for open weights without citing H5N1 alongside it is citing half a record. 3. Enforcement was a funder's condition, not a norm. See above. Analogies that reach for "the scientists agreed to share" have the mechanism wrong, and the wrong mechanism gives the wrong policy advice. 4. The economics were free, and in AI they are not. Sequence is non-rival, does not improve by being hoarded, and no producer could build a durable moat from it — releasing it cost the centres approximately nothing they could otherwise have captured. Weights are the product. The condition that made openness cheap at Bermuda is exactly the condition that does not hold at a frontier lab, and it is why a Bermuda-shaped agreement is harder now, not easier.

And the erosion, which the clean version of the story omits. The pure Bermuda rule was renegotiated twice — the Fort Lauderdale meeting convened by the Wellcome Trust in January 2003, which introduced the "community resource project" framing and a publication-priority protection for data producers, and the Toronto workshop of May 2009, which extended pre-publication sharing to other data types with a negotiated moratorium on downstream publication. Immediate unconditional release survived as an ideal and was traded, in stages, for something producers would actually accept. That is what happens to open-release norms over time, and it is part of the precedent.

---

5. Commonly misused as

Not required for moment. Included because four of these are live in 2026 arguments about AI.

5.1 "The human genome was completed in 2003"

It was declared complete in 2003, at 99% of the euchromatic fraction. About 8% of the molecule remained unsequenced until 1 April 2022 and the Y until 23 August 2023 (§2.2). Sulston said in 1996 that completion would be asymptotic. The general form of the error — a scoped technical claim losing its scope on the way to the public — is the single most transferable thing in this file.

5.2 "$141 back for every $1 — big science pays for itself"

The number comes from Economic Impact of the Human Genome Project, Battelle Technology Partnership Practice, May 2011, funded by the Life Technologies Foundation: for 1988–2010, $796 billion of economic output, over $244 billion in personal income, 3.8 million job-years, and a return of 141 to 1; a 2013 update put U.S. impact near $1 trillion.

What is wrong with it is not arithmetic, it is attribution. The study uses an input–output model that traces essentially all downstream genomics activity back to the original federal investment. Economists objected on the record; Robert Topel of Chicago Booth's objection is the cleanest — "These are all costs" — because jobs created and salaries paid are inputs consumed, not benefits produced, and counting them as return double-counts the spending. Nature ran the methodological criticism in May 2011. And the report was industry-funded.

This is the exact shape of the AI-investment-return figures circulating in 2026, and this entry is what a reading cites when one appears: an input–output multiplier, sponsored by a party with a position, presented with three significant figures. bin/ethos-check.py refuses to distribute a reading that fakes precision. A 141:1 ROI is fake precision wearing an economist's coat.

5.3 "ENCODE showed 80% of the genome is functional"

ENCODE's September 2012 papers reported "biochemical functions for 80% of the genome," which was widely read as the end of junk DNA. Graur et al., "On the immortality of television sets: 'function' in the human genome according to the evolution-free gospel of ENCODE," Genome Biology and Evolution 5(3):578–590 (2013), showed the claim rests on a causal-role definition of "function" (this DNA does something detectable in an assay) rather than a selected-effect one (this DNA is maintained by selection because it does something), and that under the latter the number collapses — by their mutational-load argument, for 80% to be functional "each couple in the world would have to beget on average 15 children and all but two would have to die or fail to reproduce."

A definitional choice presented as an empirical result. The AI analogue needs no explanation: it is every benchmark where the metric's definition does the work the result is credited with.

5.4 "The genome flopped, so this kind of thing always flops"

The inverse misuse, and the one this file is most likely to cause if quoted carelessly. §4.4 is the answer: rare-disease diagnostic yields of 25–55%, a 2.6× enrichment in drug success for genetically supported mechanisms, a 30% reduction in adverse drug reactions in the largest pharmacogenomics implementation trial, routine tumour profiling, an approved CRISPR medicine, and a bespoke base editor built for one infant in six months. Doubt applied evenly is calibration; doubt applied only to the results one finds inconvenient is the corruption README rule 7 names. The correct use of this entry is to discount a timeline and a mechanism claim, never to discount a field.

5.5 "It came in early and under budget, so central planning works"

It did, and the reasons were a private competitor arriving in 1998 and a machine arriving in 1998 (§4.1). Cite the outcome; do not cite the plan as the cause of the outcome.

5.6 "Bermuda proves open release always wins"

§4.5, all four breaks. The short form: the Bermuda Principles were an anti-monopoly rule about a non-rival map, enforced by a funder's renewal criteria, in a domain where no misuse argument existed — and when one did exist, the same field redacted.

---

6. Where this file could be wrong

---

7. Sources

Primary — the project and the announcement

Primary — Bermuda

Primary — the predictions

Independent grading

The delivered record

The commons and the patents

The gene count

The function dispute

In this repository