The Year 2000 problem — the catastrophe that did not arrive, and why that settles nothing

proposals.md filed this as moment, noted that prediction was arguable, and left the ambiguity standing rather than deciding it: "the owner notes it is arguably prediction too, and the ambiguity is left standing here rather than decided." jobs/canon.md does not let this file do the same — the kind is "whichever ... actually fits." So it is decided here, on build.py's own definitions, which is the method ltcm-1998 used and the only one that does not reduce to taste.

prediction is defined as "a dated claim, carried graded," and the case for it is genuinely strong. The material is dated claims. There are named forecasters with due dates that have passed, self-grades, and independent grades that can be set against them. The owner's own proposal text — "the base rate against which every doom prediction and every dismissal of one should be read" — is almost verbatim the kind's stated purpose.

It loses on three counts, and the third is the real one.

First, the filing would be structurally wrong. Every prediction entry in canon/ is dated at the claim, not at the deadline: vinge-singularity-1993, frey-osborne-2013, musk-robotaxi-2019, kurzweil-2005. An entry with the id y2k-2000 and year: 2000 is dated at the due date. The id is the rollover.

Second, creator: would have to lie. A prediction needs a predictor. Filing this as prediction would mean choosing between de Jager and Yardeni and Gartner, and the entry is about none of them individually.

Third — and this is the one that matters — the two kinds contribute different things, and this canon needs the other one. A prediction entry contributes one observation to a base rate: forecaster said X by date D, outcome was Y. Y2K's value is not one more observation. It is a demonstration that for a whole class of forecast, the outcome does not discriminate between the hypotheses at all — that "the warning was overblown" and "the warning worked" predict the same January. That is a fact about whether the base rate is measurable, not an entry in it. It belongs with the event.

And moment fits without strain: "something that happened and changed what people expected next." What happened was almost nothing, in public, on a date the whole world was watching. What changed was the standing credibility of technological catastrophe warnings — permanently, and by the admission of both sides. Finkelstein, writing eight months later: "This has, I believe, damaged the credibility of the profession." Thomas, seventeen years later: "The Y2K problem is often referred to now as an example of an exaggerated threat or to support an argument that expert warnings — for example about the threat from climate change — should be ignored."

The entry still does the prediction work. Section 3 carries the claims in their own words, the dates made, the dates due, what happened, and — where the predictor graded themselves — their grade beside an independent one. It just does that work inside a moment, because the forecasts are evidence about the event rather than the point of it.

One note on descends_from before the prose. hollerith-census-1890 is the ancestor of the field-width economy: the punched card made a column a scarce, costed thing, and two digits of century were the cheapest two columns to drop. ibm-pc-1981 is the ancestor of the second-largest slice of the actual work — Thomas ranks PC real-time clocks the number-two Y2K problem, and the reason they behaved inconsistently is precisely the clone-BIOS ecosystem that entry is about: the Award, AMI and Phoenix reimplementations each handled the century byte their own way, and Thomas names "the 1994/95 Award v4.50 BIOS" as one that could not be fixed in software at all. The ancestor this file wanted and could not use is the second cause, which Thomas states plainly: "The convention persisted, partly because of compatibility but mainly because programmers did not consider that their programs might still be in use by the time that this could become a problem." Code outliving the assumptions it was written under has no entry in canon/ and should. moore-1965 is not it — Moore's law explains why the constraint evaporated, not why the code survived it — and inventing an edge to moore-1965 here would be decoration.

1. What it is

Not one bug

The popular account has a single defect: two-digit years, so 00 sorts before 99 and 2000 reads as 1900. That is the biggest piece, and it is not the whole thing. Thomas, who ran the remediation for one of the four large accountancy firms internationally, enumerates five distinct problems plus a residue:

1. Two-digit years in code and stored data, typically YYMMDD, in COBOL usually as binary-coded decimal. Breaks sorting, comparison, and interval arithmetic whenever one date is in the 20th century and the other in the 21st. 2. Real-time clocks in PCs. The PC AT's battery-backed CMOS RTC updated only a two-digit year; the century lived in a separate memory location normally set to 19. What happened at rollover depended entirely on the BIOS. DOS reset any out-of-range date to 1 April 1980, which meant new files got 1980 timestamps, sorted to the front of date-ordered files, and vanished from subsequent processing — silently. 3. Real-time clocks in programmable logic controllers, where the firmware was mostly two-digit, the operating systems were proprietary, and the ladder logic was usually undocumented. 4. The leap year. 2000 was a leap year — divisible by 400 — and the programmers who knew the hundred-year rule but not the four-hundred-year rule got it wrong. The ones who knew neither got it right. 5. Sentinel dates. 99 was a common end-of-file marker and 00 a common "ignore this record" marker, because a programmer in 1975 could not imagine the code running that long.

Plus the residue: the constant 19 compiled into programs, pre-printed on cheques, contracts and forms, and — Thomas's detail, and it is not a joke — carved into headstones beside the birth years of people who then outlived the century.

Why the two digits existed is not a mystery and not incompetence. Alan Greenspan, testifying to Congress in 1998: "I'm one of the culprits who created this problem. I used to write those programs back in the 1960s and 1970s, and was proud of the fact that I was able to squeeze a few elements of space out of my program by not having to put a 19 before the year. Back then, it was very important." It was very important because of the 80-column card and because core memory cost on the order of a dollar a bit. The decision was correct when it was made. It failed because the code outlived the constraint by thirty years.

What was spent

There is no agreed number and the disagreements are instructive.

• $8.34 billion, US federal government — OMB, via the Senate final report, p. 11.
• $8.5 billion, US federal government — the same Senate final report, p. 3.
• ~$100 billion, US public and private — US Department of Commerce.
• $150–225 billion, US — Gartner Group.
• $134 billion US and $320 billion worldwide — IDC, "Project Magellan."
• $500 billion worldwide — "a journalist for Newsweek," which is the Senate report's own attribution, verbatim.

Two things are worth pausing on. The primary institutional document disagrees with itself by $160 million between page 3 and page 11, which is a fair indication of how firm any of this is. And the most-repeated global figure — the half-trillion that anchors nearly every retrospective, including ones written this decade — is sourced by the United States Senate to an unnamed magazine journalist. Quiggin's complaint that "no serious ex post evaluation has been undertaken" is not rhetorical. Nobody ever counted.

The one genuinely comparable dataset is IDC's, reported by the International Y2K Cooperation Center in February 2000 and quoted by Thomas: 1999 national expenditure of $34 billion in the United States, $2.5 billion in Italy, $100 million in Venezuela; normalised by telephone lines as a crude proxy for how much digital infrastructure a country had, $195 per line in the US, $97 in Italy, $38 in Venezuela. Hold onto the Italian number. It is the hinge of the whole argument and it is routinely reported as zero.

What happened

Almost nothing that anyone noticed, and a long tail of things that some people did.

The Senate Special Committee, which had staff sitting at the Information Coordination Center desk through the rollover with access to a three-tier exception-reporting system across every US sector, concluded: "While hundreds of computer problems have been reported since January 1, 2000, most have been quickly corrected and none have caused serious disruptions."

The documented failures are real, small, and worth naming individually, because "nothing happened" is false and the difference between no failures and small consequences is the whole of this project's magnitude-versus-consequence rule.

• The largest single US government failure was caused by the fix. At midnight GMT the National Reconnaissance Office's imagery processing ground station at Fort Belvoir, Virginia lost the ability to process satellite intelligence for about three hours, and ran degraded until full restoration on 3 January. It was a Y2K patch that did it: the sections of the fix had each been tested, and never tested together. Remember this one — the failure mode is component evaluations that pass individually and are never composed, and it will be familiar to anyone reading model evals in 2026.
• The one documented case of serious human harm arrived five months late and was found by accident. Between 4 January and 24 May 2000, the PathLAN system at Sheffield's Northern General Hospital, which processed Down's syndrome screening for nine hospitals across South Yorkshire, Lincolnshire and the East Midlands, miscalculated maternal age after the century rolled. Around 150 pregnant women were told their risk was low when it was not. Four affected pregnancies went undetected; accounts differ on the split between births and terminations, with The Register (14 September 2001) reporting four births and two terminations. It was noticed because a health visitor in Yorkshire thought she was seeing an unusual number of babies born with Down's syndrome — not by any monitoring system, and not for months.
• Medicare payment delays; double-billing by credit card companies; 911 problems in several localities; an anomaly in a nuclear weapons plant system; the Federal Reserve Bank of Chicago unable to move $700,000 in tax payments for a day; low-level windshear alert systems down for up to two hours at six US airports; a nuclear power plant abroad with non-safety-related problems; a Hong Kong Futures Exchange outage; a video store bill for $91,250.

Set against that: the world's most-watched single moment of computer risk, and the aggregate consequence rounds to a bad afternoon.

The one number that should have ended the argument, and did not

Buried on page 10 of the Senate committee's final report is the most important sentence anyone wrote about Y2K, and it is not about the rollover:

> "Committee hearings also revealed a significantly lower failure rate than > predicted for embedded chips. Analysis of testing during the last quarter of > 1999 predicted an embedded chip failure rate of .001%, rather than the 2-3% > failure rate projected in late 1998 and early 1999."

Embedded systems were the core of the apocalyptic case — the part that could not be patched, only replaced, in lifts and pipelines and process control. A UK government figure put roughly 7 billion of them in the world in 1996. A 2–3% failure rate across that population is a civilisational event. A 0.001% rate is housekeeping.

The projection was revised down by a factor of two to three thousand, within twelve months, by measurement rather than by argument — and the revision came from the alarmed party's own testing, before the deadline. Kraft's numbers, reported through the IY2KCC in March 1999, point the same way: of 832 PLCs tested, 83 had date functionality, and 4% of those had date-related problems — about three devices in 832, or 0.4%, in a manufacturing environment the same UK report had pegged at 15%.

This is the single cleanest quantitative finding in the entire episode, and it is absent from almost every retrospective on both sides. Thomas's 2017 lecture quotes the pre-testing UK estimate — around 5% of embedded systems failing compliance tests, 50–80% for sophisticated ones — in a passage about how the risks came to be recognised, and never reaches the measured figure. Finkelstein asserted from first principles that "only a tiny proportion of such systems have a clock at all" and was right, but did not have the number either.

2. Why a reading would cite it

The occasions are specific and at least one of them has already been logged in this canon.

1. Because pre-positioning-volt-typhoon asked for it by name and by argument. That entry sets out two cases a reading may be in when it faces a finding with no consequences attached, and named this file as the missing half:

> "The Y2K case. A forecast catastrophe whose non-arrival is equally > compatible with 'it was overstated' and 'the remediation worked,' where waiting > for the effect is correct because the effect was the claim. y2k-2000 sits > in proposals.md proposed by the owner and is not yet written; it is the > sibling this entry most needs and does not have."

It now has it. The distinction that entry forces — is the presence of the capability the finding, or is the effect the finding? — is a judgment a reading must make out loud, and this file is the standing example of the second case.

2. When a safety expenditure is justified by the absence of an incident. 2026 supplies these constantly: an evaluation regime, a deployment gate, a compute threshold, a red-team programme, defended on the ground that nothing bad has happened. Y2K is the case where that defence was made at maximum institutional volume — the Senate report is titled "Crisis Averted" — and where twenty-six years of argument have not settled it. A reading that cites this is not saying the spend was wasted. It is saying the absence of an incident is not the evidence, and naming what would be.

3. When someone argues that an unrealised warning was therefore wrong. This is the misuse Thomas names, and the reading should name it back. The Y2K record contains real bugs found by testing that would have caused real failures: a Swedish reactor that shut itself down when its clocks were rolled forward in mid-1998; the UK's Rapier anti-aircraft system, which the Ministry of Defence found would have failed to fire; errors on BP Exploration's oil platforms; steering and fire-detection faults found by the US Coast Guard; 10% of Visa swipe terminals unable to accept cards expiring after 1999. None of that is compatible with "there was never anything there."

4. When a projected failure rate meets a measured one. This is the entry's most transferable contribution and it is a number, not a mood. Y2K supplies a documented instance of a catastrophe-relevant failure rate overestimated by three orders of magnitude, corrected by systematic testing within a year, by the people who had made the estimate. When a reading meets "N% of agents exfiltrate under adversarial prompting" or "X% of tasks are automatable," it now has a historical scale for how far a projection can sit from a measurement — and a reason to ask whether anyone has run the test rather than the model.

5. As LENSES.md's magnitude-versus-consequence rule at full stretch. The lens says: "Record the magnitude, never the consequence." Y2K is where that rule pays. The magnitudes are hard: $8.34 billion federal, 21% of US federal mission-critical systems compliant in May 1997 and 99.9% in December 1999, a measured embedded failure rate of 0.001%, around 150 women given wrong screening results between two named dates. The consequence — whether the money bought the quiet — is a forecast about a world that did not happen, and belongs in prose, labelled as judgment. Every retrospective that has failed on Y2K failed by reporting the consequence as though it were a magnitude, and both sides did it.

6. As the honest counterweight to the base rate the prediction entries are building. kurzweil-corpus-1990-2024, vinge-singularity-1993, frey-osborne-2013, musk-robotaxi-2019 are accumulating a record of how wrong technological forecasting runs, and the record so far runs optimistic on capability, and long on timelines. Y2K is the counter-case: a forecast that was pessimistic, that mobilised hundreds of billions of dollars, and whose failure mode was an unfalsifiable success. A base rate assembled only from forecasts that can be graded is a base rate with a hole in it exactly where the self-preventing prophecies sit, and this entry marks the hole.

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

moment does not require this section. It is here because the graded claims are the evidence, and because grading them is the only thing that turns a famous non-event into something a reading can use.

The claims, graded

Peter de Jager, "Doomsday 2000." Computerworld vol. 27 no. 36, 6 September 1993, p. 115 — the article that is usually credited with starting the whole thing. This file could not read the primary text; the Internet Archive copy exists at archive.org/details/computerworld2727unse11 but the fetch tool available here declined to reproduce it, so the characterisation below is at one remove and a session with direct access should sharpen it. What is on the record independently: the argument that date-sensitive processing was fundamental to national infrastructure and that failures could push the economy into recession; a cost estimate of $50–70 billion to fix it; testimony to the US Congress and the Canadian Parliament in 1996; and the operating principle — "we don't know in advance what will fail, so we have to fix everything."

Graded. The cost estimate was low by a factor of two to six against every subsequent figure, which is the opposite of the error a scaremonger is accused of making. The mechanism was correctly described. The operating principle is the part that aged worst, and it is the part that determined the spend: "fix everything" is the only strategy that is unfalsifiable and maximally expensive, and Quiggin's whole case is that "fix on failure" was available and was dismissed too fast.

Self-grade. De Jager published "Doomsday Avoided" in early 1999 — a year before the deadline — announcing "We've finally broken the back of the Y2K problem," and answering whether it was solved with "No, not entirely. But, we have avoided the doomsday scenarios." A self-grade filed twelve months before the event it grades is a structurally interesting object: it claims the win in advance and is immune to the outcome either way. From mid-2000 he stopped giving Y2K interviews.

Independent grade. Divided and never settled. The Lifeboat Foundation gave him an award in 2009 for the warning; Forbes in March 1998 filed him under "Y2K fear merchants" and noted he was selling advertising on year2000.com. Both can be true.

Edward Yardeni, chief economist, Deutsche Bank Securities. The most credentialled forecaster to attach a probability. Claim: that Y2K failures would cause a global recession — initially framed in 1997 as a risk of a worldwide downturn lasting at least twelve months from January 2000, comparable to 1973–74; by the autumn of 1999 revised in severity to an intense six-month recession driving the Dow to 8,000, and revised upward in confidence to 70%. (Sources differ on the starting probability — 30% or 40% — and agree on the 70%.) Made: 1997, restated through 1999. Due: calendar 2000. What happened: US GDP grew about 4% in 2000. There was no recession attributable to Y2K anywhere.

Self-grade, and it is dated. Addressing investors on 3 January 2000, Yardeni said: "I'll just say, 'I was wrong.'"

Independent grade, one day later. Slate, 4 January 2000: "spectacularly wrong," with the observation that it would not cost him anything, because "in Yardeni's particular market niche it's better to be vivid than to be right."

The two grades agree. jobs/canon.md asks for a self-grade and an independent one specifically because they usually differ — Kurzweil is the standing case, and kurzweil-corpus-1990-2024 exists because of the gap. Here there is no gap: the forecaster conceded in one sentence, on the record, within 72 hours, and the independent verdict was the same verdict in harsher words. That is worth recording precisely because it is rare, and because it shows the concession cost him nothing — Slate's point, not this file's, and borne out: he is still quoted on recession odds in 2026.

Gartner Group. The consultancy that both promoted concern and sold advice on it, which makes its forecasts the most interest-conflicted and also the most specific. Claim, 1998–99: that only about 10% of Y2K problems would occur on 1 January 2000, with 55% during 2000 and 35% during 1999; that peak problem periods would fall in mid-2000, with problems lingering into 2003. Due: 2000–2003. What happened: the peak never came. Y2K programmes scheduled to run for years were shut down within months. A CMP Media survey of 1,750 IT professionals and consumers in 2000 found 63% believed Y2K had been mostly hype.

Graded. Wrong, and wrong in an instructive way. The 35%-during-1999 clause is the one that matters, because it was a forecast about the past by the time anyone could check it, and Quiggin checked it: "Since there were no failures of critical systems reported during 1999, the best estimate of the number of such failures in 2000, even in the absence of additional remediation, was zero." That inference was available in mid-1999 and nobody acted on it.

The US Senate Special Committee. Claim, 29 February 2000: "In the Committee's judgment, the level of effort was justified, and the expenditures of the public and private sectors were indeed necessary." What it rested on, stated in the next sentence: "Testimony and available research during 1998 and early 1999 convinced the Committee that the Y2K threat was real."

Graded. The Committee's own sentence gives away the structure of the problem. Its verdict on whether the spend was necessary rests on evidence gathered before the spend was completed — because there is no other kind. Nothing observable after 1 January 2000 could have moved it. Two things to the Committee's credit, which the retrospectives usually strip out: it labelled the finding a judgment, in those words, which is more epistemic care than most of its critics have shown; and it published, on page 9, the reason its own numbers could not be trusted — "Because there is no incentive for corporations or countries to openly report problems, the full extent of Y2K's impact may never be known," noting that 32 countries including Australia, Brazil, Great Britain, Canada, Germany and Norway reported no incidents at all to the international clearing house while reliable news services were reporting incidents in each of them.

But the report is titled "Y2K Aftermath – Crisis Averted." The body says "in the Committee's judgment." The cover says it as a fact. This canon should notice when a document is more careful than its own title, because the title is what gets cited.

The three tests, and why everyone argues about the wrong one

The reason twenty-six years of argument have not converged is that the debate is conducted almost entirely over the least informative of the three tests that were actually run.

Test 1 — the look-ahead test, which ran before the deadline and is not confounded. Systems that compute forward hit 2000 before 2000 did: expiry checks, amortisation, look-ahead scheduling, financial years. These failures happened and are documented. Marks & Spencer rejected a delivery of tinned meat in the late 1980s because the stock system read a 2000 expiry as 1900. In 1992 a 104-year-old woman in Winona was invited to join an infant class because she was born in "88". On 31 December 1996 a Western Australian aluminium plant's PLCs computed day 366 minus 366, got zero, and shut the plant down with equipment damage — and the sister plant in Tasmania was warned in time and stopped it by manual override. On 28 December 1999, ten thousand Racal/HSBC card swipe machines across Britain stopped accepting cards, because they looked four working days ahead.

Both sides accept every one of those facts and read them in opposite directions. Thomas: they prove the bug was real and that finding them was worth the money. Quiggin and Finkelstein: they are the sample, they were small, local and fixed within hours, and they are the best available estimate of what the unremediated tail would have looked like. Finkelstein pressed the sharpest version — 9 September 1999, when 9999 sentinel data was widely predicted to break, and nothing happened: "Now, if you regarded the problem as a serious threat in the first place, this should surely have acted as a spur to some serious rethinking. It did not."

Test 2 — the direct measurement, which is the cleanest evidence and which nobody quotes. Organisations set the clocks forward and looked. This is not confounded by remediation; it is the diagnostic step of remediation, and it measures the pre-fix failure rate directly. What it found: the Swedish reactor that shut down; the Rapier system that would not have fired; the Millennium Dome's systems throwing so many errors when tested at 23:55 on a simulated 31 December that they scrolled off the console; Chrysler's Sterling Heights plant, rolled forward in mid-1997, where security locked everyone out and the time clocks stopped — Robert Eaton: "We got a lot of surprises." And, aggregated over the last quarter of 1999: 0.001% of embedded chips, against a 2–3% projection.

Test 3 — the rollover. Confounded, uninformative, and the only one anyone remembers.

The synthesis the sources will actually support, stated plainly so a reading can borrow it: the bug was real and specific; the failure rate was overestimated by two to three orders of magnitude; the remediation was sized to the overestimate; and the fraction of the spend that was wasted is not knowable from the rollover, because the rollover was never the test. Every clause of that has a primary source behind it. What no clause of it does is tell you whether the money was well spent, and no honest reading of the record does.

What the sceptics got right

• The embedded-systems case collapsed, and collapsed early. Quiggin's a-priori argument — that a device whose date resets when its battery is changed cannot be critically dependent on the date — turned out to be the correct physical intuition, and the Q4-1999 measurement confirmed it.
• The cascade never came. Finkelstein: "Ripple failure is a very rare failure mode." Thomas, from the other side, concedes it in almost the same words: "Systems were actually not as tightly interdependent as had been feared, so there was redundancy in supply chains that provided resilience."
• The incentives were asymmetric, and everyone involved now says so. Quiggin: "The absence of any serious Y2K problems could always be attributed to the success of the remediation program." Finkelstein's version is the tiger repellent: "the identical argument could be made by the salesman who persuades you to purchase tiger repellant and presents the fact that you are not subsequently eaten by a tiger as proof of the efficacy of the repellant." Thomas, defending the effort, concedes the same ground: "The threat had been exaggerated to some extent, partly by suppliers and consultants who wanted to sell equipment, systems and services."
• There was no evidence base. Finkelstein: "There is virtually no serious scientific literature on the problem." Quiggin in 2025 puts it against the sharpest possible comparison: Y2K had "almost no serious peer-reviewed research" and mobilised hundreds of billions; climate change has "a mountain of scientific research built up over decades" and has failed to move governments to vigorous action. The relationship between an evidence base and a policy response was, in this case, inverse — which is a finding a reading about AI governance can use directly and in either direction.

What the sceptics got wrong

• "Italy spent nothing" is false, and it is the load-bearing claim. The argument that failures did not track spending rests on countries that supposedly did not try. Italy spent $2.5 billion in 1999 alone, and $97 per telephone line against the United States' $195 — half, not zero. Whatever the international comparison shows, it is a 2:1 ratio, not a controlled experiment with an untreated arm. The Senate report adds the same correction from the other end, in a footnote: "Some analysts also believe that other industrialized countries, such as the United Kingdom, Canada, Denmark, and the Netherlands, spent equivalent amounts per capita to the U.S."
• The late-starters were not unremediated; they were cheaply remediated. Cost per line of mainframe code was estimated at $1–2 in 1996 and expected to exceed $4 by 1998; it instead fell to pennies as automated scanning tools matured. Thomas: countries that started late "were rescued by the progress that was made by others." The Senate report says the same and adds the uncomfortable corollary — that some international "compliance" was temporary clock rollback, which "allows the temporary claim of compliance but fails to address long-term date dependency problems." Some of the countries in the natural experiment had not fixed anything; they had postponed it.
• "Nothing happened" is false at the level of individual harm, and the Sheffield screening failure is the case that makes it false. It also went undetected for five months and was found by a health visitor's hunch, which means the aggregate failure counts everyone argues over — on both sides — are undercounts of unknown size. The Senate committee said so itself.

What the defenders got wrong

• They claimed a counterfactual as a result. "Crisis Averted" is a title, not a finding, and the report's own body knows it.
• They did not commission the evaluation. Quiggin's most durable complaint is procedural, not substantive: Australia spent an estimated $A12 billion and the accounting was a 17-page self-congratulatory report that "would have been rejected as grossly inadequate if it had been published as an account of the annual operations of a minor local council." The US President's Council report ran about 30 pages excluding appendices. Twenty-six years on, the cost-benefit analysis of the largest coordinated preventive IT expenditure in history has still never been done. That is the failure a reading should actually cite, because it is the one that was entirely within the defenders' power to fix and remains fixable in the AI case.
• The one thing the President's Council conceded, it conceded against itself. Asked why small businesses had so few problems, its final report answered that "there is a natural inclination not to report problems that are fixed in very short time frames," and that over 70% of companies had reported Y2K glitches before the rollover "even though the public was unaware of virtually all of them." Quiggin's reading of that passage is hard to argue with: it is a description of "fix on failure" working, written by the body that had spent two years arguing against it.

The honest residue

The counterfactual is not recoverable. This file went looking for the study that would settle it and the study does not exist — not in 2000, not in 2004, not in 2017, not in 2025. Both of the two serious retrospectives say so, from opposite sides. Quiggin's 2025 verdict on the state of the argument is the fairest summary available: 25 years on, "no one much has changed their mind," and "this issue will never be resolved."

That is the finding. Not "it was hype," not "it worked" — that a global expenditure somewhere between $100 billion and $500 billion, on a problem with a known mechanism, a known deadline and a testable failure mode, produced an outcome that cannot distinguish success from waste, and that nobody has since been sufficiently motivated to try. Anything in AI safety with a fuzzier mechanism, no deadline and no test is in a strictly worse epistemic position than Y2K, and Y2K is unresolved.

4. Commonly misused as

limit entries are required to carry this section. This one carries it because Y2K is misused more reliably than most limits are, and in both directions.

1. "Y2K proves expert warnings can be ignored." Thomas names this as the misuse he most objects to, and specifies the target: it is used "to support an argument that expert warnings — for example about the threat from climate change — should be ignored." What the record actually shows is that testing found real defects in real safety systems — a Swedish reactor, an anti-aircraft missile, ship steering, fire detection — at rates far below projection and far above zero. "The projection was 2,000× too high" and "there was nothing there" are different claims and only the first is supported.

2. "Y2K proves preparation works." The mirror error, and the one with institutional weight behind it, since it is the title of the Senate's report and the substance of Bennett's "had we as a nation not focused on this issue and dealt with it, we would have had very significant problems." That is a judgment, and the Committee's own text labels it one. Neither direction is established by January 2000, because January 2000 is consistent with both.

3. "Nothing happened." False as stated. Around 150 women in South Yorkshire, Lincolnshire and the East Midlands were given wrong Down's syndrome screening results over 141 days; the NRO lost satellite intelligence processing for three hours; the Federal Reserve Bank of Chicago could not move $700,000 for a day. The defensible claim is that the aggregate consequence was small. Collapsing "small consequence" into "no event" is precisely the move LENSES.md forbids in the other direction, and it is no more permitted here.

4. "Italy spent nothing and was fine." The most-repeated empirical claim in the sceptical case, and it is wrong by at least $2.5 billion. Italy spent about half the US rate per unit of digital infrastructure. Use the real ratio; it still supports a version of the argument, and the real version survives contact with Thomas.

5. The tiger-repellent framing, deployed as though it were a conclusion. Finkelstein's analogy is a statement of an epistemic problem — that the seller of protection can always claim the absence of harm as proof — not a demonstration that the repellent does not work. Quoted as though it settled the matter, it commits the alarmists' error in the opposite direction: treating a non-discriminating observation as though it discriminated.

6. "The AI risk people are the Y2K people." The comparison is fair enough to be worth making carefully and is almost never made carefully. Y2K had three things most AI risk claims do not: a known mechanism (a specific field width in specific code), a known deadline (midnight, once), and a testable failure mode (set the clock forward and look). It is the easiest possible case for technological risk forecasting — and even with all three, the failure rate was overestimated by three orders of magnitude and the value of the response is unresolved twenty-six years later. That cuts in both directions and a reading should let it: it is a caution against confident catastrophe forecasts and a caution against assuming a quiet outcome vindicates the dismissal. What it is not is a reason to skip the test, because in Y2K the test was the only thing that ever produced a number.

7. The misuse this project must guard against directly. Nothing in this file licenses discounting an AI finding because it has no consequences attached. That question is the one pre-positioning-volt-typhoon poses, and the two entries together define the fork: where the effect is the claim (Y2K), waiting for the effect is correct; where the presence of the capability is the objective (Volt Typhoon), waiting for the effect means waiting for the thing the finding existed to prevent. A reading that cites this file to dismiss a capability finding has cited the wrong entry.

5. Sources

Primary — the institutional record.

• **United States Senate Special Committee on the Year 2000 Technology Problem, Y2K Aftermath – Crisis Averted: Final Committee Report, Summary of Committee Findings, 29 February 2000 (S. Prt. 106-XX; Bennett, Chairman; Dodd, Vice-Chairman). Read in full as a PDF, via govinfo.gov/content/pkg/GOVPUB-Y1_3-PURL-LPS90964. The source of: the transmittal letter's "minor, localized, and short-lived"; the $100 billion / $8.5 billion overview figures (p. 3) and the conflicting $8.34 billion OMB figure (p. 11); the Newsweek-journalist attribution of the $500 billion global estimate (p. 12 n. 1); the Gartner $150–225 billion and IDC Project Magellan $320bn/$134bn figures (same footnote); the per-capita comparison naming Italy, Spain and Russia and the footnote naming the UK, Canada, Denmark and the Netherlands as US-equivalent (p. 12 nn. 1–2); the 32-countries-reported-nothing passage and the domestic problem list (p. 9); the embedded-chip 0.001% versus 2–3% revision and the "In the Committee's judgment" sentence (p. 10)**; the temporary-clock-rollback admission (p. 10); and the Gartner peak-in-2000, lingering-to-2003 projection (p. 12).
• **US General Accounting Office, Year 2000 Computing Challenge: Lessons Learned Can Be Applied to Other Management Challenges, GAO/AIMD-00-290, 12 September 2000, to Rep. Stephen Horn. Read as a PDF via gao.gov/assets/aimd-00-290.pdf. Source of the OMB compliance trajectory: 21% of the 24 major federal departments' mission-critical systems compliant in May 1997, 99.9% by December 1999** (p. 6), and of GAO's own count of 160+ Y2K reports and testimonies.
• International Y2K Cooperation Center, February 2000 report of IDC data for 50 countries — the $34bn / $2.5bn / $100m national figures and the $195 / $97 / $38 per-telephone-line index. Read at one remove, quoted at length in Thomas 2017; the IY2KCC site survives only in Internet Archive captures.
• **President's Council on Year 2000 Conversion, The Journey to Y2K: Final Report, 29 March 2000 (Koskinen, Chair). Not read directly** — the Council was wound up 31 March 2000 and its site is gone. The "over 70 percent of companies reported they had experienced Y2K glitches" passage and the small- business answer are quoted in full by Quiggin 2004, which is where this file took them.

Primary — the two serious retrospectives, both read in full.

• Anthony Finkelstein, "Y2K: a retrospective view," University College London. The longer original of the paper published in Computing & Control Engineering Journal, August 2000, v11 N4, 156–159, alongside a companion piece by Martyn Thomas. Read in full as a PDF at discovery.ucl.ac.uk/id/eprint/1147. Source of: the tiger repellent; "damaged the credibility of the profession"; the 9/9/99 argument; the background-rate argument; "Ripple failure is a very rare failure mode"; the embedded-clock point; "It is now well known that, despite devoting much less effort to Y2K than the UK, most other countries (including less developed countries - such as Italy) experienced no significant Y2K problems"; "There is virtually no serious scientific literature on the problem"; "too good to check"; and the concession that most heavy Y2K spending was justified anyway, "Not because of Y2K itself but rather because an organisation that does not know which are its business critical systems ... is headed for trouble - regardless of date."
• John Quiggin, "The Y2K scare: causes, costs and cures," Risk & Sustainable Management Group, University of Queensland, Australian Public Policy Program Working Paper 1/P04, version of 18 February 2004; published in Australian Journal of Public Administration 64(3), 2005, 46–55. Read in full as a PDF at rsmg.group.uq.edu.au/files/1186/WPP04_1.pdf. Source of: the $A12 billion and $544m Commonwealth figures; the Gartner 10/55/35 distribution (via Lei 2000) and Goodwin's 2000 revision to 5%; "Since there were no failures of critical systems reported during 1999, the best estimate of the number of such failures in 2000, even in the absence of additional remediation, was zero"; Bettinelli's Italian awareness figures via BBC 1998 and Quiggin's footnote "That is, 80 per cent were correct in their evaluation"; Taskforce 2000's 8 November 1999 advice to avoid Italy and Germany; the asymmetric-incentives analysis; the public-good argument for institutionalised skepticism and the Promotor Fidei; the CMP Media 63% survey (Berkowitz 2000); and the 17-page-report complaint.
• John Quiggin, "Y2K: Apocalypse averted or pointless panic?", 7 January 2025. Read at one remove — johnquiggin.com returned HTTP 403 to this job, so the Substack mirror was fetched and summarised by the fetch tool rather than read verbatim. Source of "no one much has changed their mind," "this issue will never be resolved," and the Y2K-versus-climate evidence-base contrast. The quotations in this file from the 2025 post should be re-verified against the original by a session with direct access; the 2004 quotations should not need it.

Primary — the defence, read in full.

• Martyn Thomas CBE FREng, "What Really Happened in Y2K?", 4 April 2017, the Gresham College lecture text, read in full as a PDF at regulation.org.uk/library/2017-y2k_bug_evaluation.pdf. Thomas led Deloitte & Touche Consulting Group's international Y2K services through the 1990s and was final auditor of the UK National Air Traffic Services programme, which makes this the most technically specific defence on the record and also an interested one; he says so himself in the second paragraph. Source of: the five-problems taxonomy; the Greenspan testimony; the PC RTC/BIOS analysis and the Award v4.50 BIOS; the BSI PD2000-1 four rules (January 1997); date expansion versus windowing versus replacement, and the warning that windowing "was not a permanent solution"; the $1–2 to pennies per line cost curve; the IY2KCC advisories including the Western Australian aluminium plant, Kraft's 832 PLCs and Chrysler's Sterling Heights shutdown; the prevented-failure list (Rapier, the Swedish reactor, BP Exploration, US Coast Guard, 10% of Visa terminals, the Millennium Dome console); the post-2000 failure list including the NRO satellite mis-fix; the PathLAN account; the concessions on exaggeration and on loose coupling; and the closing argument that the risks are higher now, not lower.
• **Farhad Manjoo, "Was Y2K a Waste?", Slate, November 2009. Read via fetch.** Source of the Kappelman estimate that "80 percent to 90 percent of the money spent on Y2K was right on target" and of the per-capita claim about the UK, Canada, Denmark and the Netherlands, which independently corroborates the Senate report's footnote.

Primary — the graded forecasts.

• **Peter de Jager, "Doomsday 2000," Computerworld vol. 27 no. 36, 6 September 1993, p. 115. NOT READ. The scan exists at archive.org/details/computerworld2727unse11/page/n116 and the fetch tool available to this job declined to reproduce the article text. Everything attributed to de Jager here is at one remove and flagged as such in section 3. A session with direct access to the scan should replace that paragraph with his actual sentences** — it is the single largest gap in this file.
• Peter de Jager, "Doomsday Avoided," early 1999. Read at one remove via search results quoting it; the two sentences used are marked as quotations because multiple independent summaries render them identically, but the dating is reported as "early 1999" because that is as precisely as it could be pinned from here.
• **"The Great Yardeni," Slate, 4 January 2000. Read via fetch.** Source of the fall-1999 revision (six-month recession, Dow 8,000, confidence raised to 70%), of Yardeni's 3 January 2000 concession "I'll just say, 'I was wrong,'" of the "spectacularly wrong" verdict, and of "in Yardeni's particular market niche it's better to be vivid than to be right." The 1997 starting probability is reported here as 30%-or-40% because the sources consulted disagree and neither is primary; Yardeni's own 1997 Senate Banking Committee testimony is on banking.senate.gov and was not fetched, and would settle it.

Primary — the failures.

• **"Down's Syndrome screening failures linked to Y2K bug," The Register, 14 September 2001. Read via fetch.** Source of the 4 January – 24 May 2000 window, the PathLAN name, Northern General Hospital Sheffield, the nine hospitals across South Yorkshire, Lincolnshire and the East Midlands, "more than 150" women, and the four-births / two-terminations split. The Guardian's contemporaneous report (theguardian.com/society/2001/sep/14/NHS1, cited by Thomas) could not be fetched from this machine — the tool refuses that domain. Published accounts differ on the exact split between births and terminations and this file says so rather than picking one.
• The NRO / Fort Belvoir satellite ground station outage, 1 January 2000. Confirmed across contemporaneous reporting (UPI, 12 January 2000; CNN, 5 January 2000; Flight Global; the RISKS Digest vol. 20 no. 72) for the three-hour processing loss, the Fort Belvoir location, the restoration on 3 January, and — the detail that matters — that the outage was caused by a Y2K patch whose sections were each tested and never tested together.

Fetch failures, recorded so a reader knows what was not checked. theguardian.com is refused outright by the fetch tool available here. johnquiggin.com returned HTTP 403; computerweekly.com returned HTTP 403. The Internet Archive's Computerworld scan was reachable but the tool declined to reproduce the article. The Senate committee's larger hearing volumes at GPO-CPRT-106sprt10 were opened but only the final report was read in full.

A rendering note, for whoever does the next canon pass. build.py's md_to_html (line 237) implements a deliberately tiny subset — headings, bold, italic, links, code, - lists, paragraphs — and canon bodies are rendered through it (line 934). It has no table support, and because the renderer joins consecutive non-blank lines into one paragraph, a pipe table collapses into a single line of pipes and dashes on the page. At least ten entries in canon/ contain pipe tables — doomsday-clock-1947, ltcm-1998, macintosh-1984, alexnet-2012, deep-blue-1997 and others — and all of them are rendering broken right now. This file wanted a table for the cost figures and uses a list instead. Fixing the corpus is a job for a session that is allowed to touch more than one entry; fixing the renderer is a job for a session that is allowed to touch build.py. This one is allowed neither, so it is written down here.

A tooling note, because it cost time and belongs in ONBOARDING.md, which this job is not permitted to write. Five of the six most important sources for this entry are PDFs, and the fetch tool returns "corrupted or improperly formatted PDF" for every one of them — it cannot decompress PDF content streams. It does, however, save the raw file to the session's tool-results/ directory and print the path. Passing that path to Read with a pages: range renders the pages and they are fully legible. pdftotext is installed at /usr/bin/pdftotext and would be faster, but invoking it needs an approval this job cannot obtain non-interactively, so Read with pages: is the working method. Every "read in full as a PDF" above was obtained that way. A previous session recorded a narrower version of this for scanned PDFs; the general form is that a fetch tool refusal on a PDF means nothing about whether the document is readable.

Project files read while writing, and only these. CLAUDE.md; README.md (rules 1–7); LENSES.md, for the magnitude-versus-consequence rule quoted in section 2; canon/proposals.md, for the entry's own proposal text; canon/pre-positioning-volt-typhoon.md, for the passage that asked for this file; canon/ltcm-1998.md, for the method of settling a kind dispute on build.py's definitions; the header blocks of canon/hollerith-census-1890.md, canon/ibm-pc-1981.md, canon/doomsday-clock-1947.md, canon/kurzweil-corpus-1990-2024.md, canon/frey-osborne-2013.md, canon/musk-robotaxi-2019.md and canon/expert-systems-collapse-1987.md; and build.py lines 722–740, for CANON_KINDS. Nothing in history/, digests/ or site/ was read or written, and no evidence was deposited: this entry places no needle, no score and no landmark, and nothing in it moves anything.

One edge that cannot be drawn from here. pre-positioning-volt-typhoon names y2k-2000 in its prose but its descends_from is [apt1-attribution-2013], and this file is forbidden from editing another entry. The two are siblings rather than ancestor and descendant, so no edge is missing — but a session doing a canon pass should know the relationship exists in prose in both directions and in the graph in neither.