Every mistake we've made, displayed more prominently than the original error. Corrections are trust deposits, not withdrawals.
This log counts anything wrong that reached the published site, whoever caught it — not what the desk fixes before it ships. A reader catch runs with the finder's name on it and the 🏅 tag; a desk catch, like the one below, runs with the same room and no credit to claim.
Correction 28 · published 2026-09-30 · corrected 2026-10-06 · live for 6 days · caught by the desk
What we said. On Inside the machine, the chip level said its default layout was IBM’s heavy-hex lattice, and opening one of its qubits showed a tunable transmon: two pads joined by a two-junction SQUID loop, with a flux line running in from the edge of the chip.
What's true. IBM’s heavy-hex processors use fixed-frequency transmons: two floating pads joined by a single Josephson junction, and no flux line. The SQUID loop and flux line belong to the tunable qubits of the square-grid chips Google describes. The qubit level now follows the layout selected on the chip level: heavy-hex shows one junction and no flux line, square shows the SQUID and the flux line.
Why we missed it. The qubit level was drawn once and the layout switch was only wired to the chip level above it. The chip card named IBM and the qubit under it was not IBM's. No check compared what a card says with what its drawing contains. A check now builds the qubit level under both layouts and asserts which parts exist.
Found on 2026-10-06 in the owner's review of the 3D page, not reported. If you had caught it first, this entry would carry your name.
Correction 27 · published 2026-09-30 · corrected 2026-10-06 · live for 6 days · caught by the desk
What we said. On the cryostat level, the control lines were drawn copper from the cold plate to the mixing chamber plate, and the card said “stainless steel near the top, cupronickel in the middle and copper at the bottom”.
What's true. Krinner et al., EPJ Quantum Technology 6, 2 (2019), arXiv:1806.07862, built a 100-qubit-scale setup with stainless-steel coax (UT85-SS-SS) on every drive and flux line, and names cupronickel as the usual alternative. Copper conducts heat far better than either, which is why it is kept off long cold spans. The span from the cold plate to the mixing chamber is now stainless steel, the card says stainless or cupronickel down to the mixing chamber plate, and copper stays only on the short last hop into the package, which that paper does not cover.
Why we missed it. The materials were set from memory of a typical setup and never held against a source. The card text and the drawing agreed with each other, so a check that only compared the two passed.
Found on 2026-10-06 in the owner's review of the 3D page, not reported. If you had caught it first, this entry would carry your name.
Correction 26 · published 2026-09-30 · corrected 2026-10-06 · live for 6 days · caught by the desk
What we said. On the cryostat level, every readout segment below room temperature was drawn niobium-titanium except the top one, which put NbTi on the 50 K to 4 K span. The card said the line was superconducting NbTi “between 4 K and the mixing chamber”, so the drawing and the card disagreed.
What's true. Niobium-titanium superconducts below about 9 K, so it carries signal with no loss only from 4 K down. Krinner et al., EPJ Quantum Technology 6, 2 (2019), arXiv:1806.07862, use superconducting NbTi coax between the mixing chamber and 4 K and stainless steel from 4 K up to room temperature. The 50 K to 4 K span is now stainless steel and the card says why.
Why we missed it. The drawing used one material for “everything below the first plate”. The card was right and the geometry was wrong, and no check compared a segment's material with the temperature it spans. A check now reads the material of each cable segment from the built scene.
Found on 2026-10-06 in the owner's review of the 3D page, not reported. If you had caught it first, this entry would carry your name.
Correction 25 · published 2026-09-30 · corrected 2026-10-03 · live for 3 days · caught by the desk
What we said. On Inside the machine: that each level is smaller than the last “by anything from two times to several thousand”; that the ten-ion chain is “about 50 µm” long; that the cryostat is 1.5 m in one place and 0.5 to 1 m in another; that the oxide barrier is “four to eight” atoms thick; that a ten-ion gate fidelity is “above 99.9%”; and, in the neutral-atom section, that Manetsch’s array is the largest so far.
What's true. The step from the ion chain to one ion is about 50,000 times, and from the neutral-atom array to one atom about 100,000 times, so the range runs from about two times to about a hundred thousand. Solving the ten-ion equilibrium, the end ions sit 2.871 length units from the centre, so for ytterbium-171 at an assumed 1 MHz the chain is about 16 µm end to end (and 1.55 µm between the closest pair), not 50. The cryostat is about 1 m. One to two nanometres divided by the 0.29 nm atom width the page itself states is roughly three to seven atoms, not four to eight. The cited ion fidelity is 99.9(1)%, so “above” overstated it. A June 2026 preprint reports about 11,000 atoms, so Manetsch’s is the largest in a peer-reviewed paper we cite, not the largest.
Why we missed it. After Correction 16 the check derived three figures from the drawing, and it did so with a different atom width (0.25 nm) than the page printed, so it confirmed the wrong range. The scale claim and the ion spacing were typed, not derived. The check now divides by the printed width, and the ion label states the trap frequency it assumes.
Found on 2026-10-03 by a fresh-session referee for the desk, not reported. If you had caught it first, this entry would carry your name.
Correction 24 · published 2026-10-02 · corrected 2026-10-03 · live for 1 day · caught by the desk
What we said. On Advantage Watch, the row called “Willow” gave a date of October 2024 and Google’s “under five minutes against 1025 years”. The Sycamore row called a November 2021 run the first classical answer to the 2019 claim (743 days), and the Jiuzhang row said no classical run had targeted the 2020 experiment. The D-Wave row called a rebuttal a preprint. On Open Problems, a card tied the quantum expectation about NP-complete problems to P versus NP, the settled list said the recommendation-systems speedup “does not exist”, and the page said every problem names its paper while two did not.
What's true. The October 2024 paper (Morvan et al., Nature 634, 328) describes a 67-qubit Sycamore experiment first posted to arXiv on 21 April 2023, so the claim date is 2023-04-21; the ten-septillion-year figure belongs to the separate 105-qubit Willow result of December 2024. Two earlier classical runs on the 2019 circuits used correlated samples (4 March and 27 October 2021); the November 2021 run is the first with uncorrelated samples, and the table now says so. Two classical papers (Bulmer et al., 3 August 2021, and Villalonga et al., 23 September 2021) did target the 2020 Jiuzhang experiment; the first sampler that beats it on distance measures came 294 days after the claim. The D-Wave rebuttal by Tindall et al. appeared in Science on 21 May 2026. Whether quantum computers can solve NP-complete problems is the separate question of NP versus BQP, which a proof that P differs from NP would not settle. Tang’s result removed the exponential recommendation speedup, not necessarily every polynomial one, and Liu, Arunachalam and Temme proved an exponential separation for a contrived learning task. The two cards without a source now have one.
Why we missed it. Every day count was recomputed and correct; each was counted from the wrong date or against the wrong first event, and no check asked which paper a date belonged to. The open-problems wording had no check at all. A fresh-session referee found them against the arXiv records. verify_advantage.py now pins the corrected dates and the correlated-sample distinction.
Found on 2026-10-03 by a fresh-session referee for the desk, not reported. If you had caught it first, this entry would carry your name.
Correction 23 · published 2026-07-14 · corrected 2026-10-03 · live for 81 days · caught by the desk
What we said. The four-line answer on the Bitcoin page said breaking one key “needed ~13 million error-corrected qubits in the 2022 estimates, about half a million in 2026’s”. The page’s FAQ markup said the 2026 attack runs “for about 20 minutes” and sits “about 2.7 orders of magnitude” from the largest chip. Two receipts and Correction 21 also said 2.7.
What's true. The 13 million is the 2022 estimate for a one-day attack, and the half million is the 2026 estimate for an attack that finishes in minutes: two different attacks, so the comparison hid most of the real change. Both figures are physical qubits, not error-corrected ones. Like for like, the racing attack, which has to finish inside the ten-minute confirmation window, was costed at about 1.9 billion physical qubits in 2022 and under half a million in 2026. The paper's own time is about nine minutes once a transaction reveals the key, not twenty. And log10(500,000 ÷ 1,121) is 2.649, which is 2.6 to one decimal, so the page body and the chart had it right and the receipts, the markup and Correction 21 had it wrong.
Why we missed it. The card was written before the chart existed and kept the first draft's numbers; the chart was later generated from data, so the two never met. Nothing compared a figure in a sentence with the figure beside it, and the 2.7 had been copied between four places. The card now states the same attack in both years, and the receipts and markup carry the generated 2.6.
Found on 2026-10-03 by the desk, not reported. If you had caught it first, this entry would carry your name.
Correction 22 · published 2026-10-02 · corrected 2026-10-03 · live for 1 day · caught by the desk
What we said. In topic 10 of the Security course: “the hybrid sends and receives about 19 times the key-share data of X25519 alone (1,216 ÷ 64 = 19)”. Elsewhere in the course: that the algorithm behind the 2025 RSA estimate “did not change”; that a browser “most often” uses plain X25519; that there is “exactly one cipher” safe against unlimited power; that a signature “can only be forged by a machine that exists” when it is checked; that RSA is from 1977; that factoring is what Shor’s algorithm “was invented for”; and that the author of a 2024 lattice claim could not fix it, “and the question is where it was”.
What's true. 1,216 bytes is the client’s share only, one direction; 64 is X25519 in both directions. Like for like, the client share is 38 times X25519’s 32 bytes, and the whole exchange, 1,216 + 1,120 = 2,336 bytes, is 36.5 times its 64. Gidney’s 2025 estimate changed the arithmetic circuit as well as the error correction, and it trades a longer runtime for far fewer qubits. For modern browsers the hybrid with ML-KEM is now the most common handshake. The one-time pad is the standard example of perfect secrecy, and Shannon’s result is that any such cipher needs a key at least as large as the message space. Forgery during a handshake needs a machine that exists at that moment, but long-lived keys are an exception (topic 06). RSA was described in 1977 and published in 1978. Shor’s paper treats factoring and the discrete logarithm together. The 2024 claim was withdrawn after a bug was found, and the standard parameters were never in its scope.
Why we missed it. The division was exact, so the checker passed it; it never asked what each number was a number of. The other sentences were claims of exclusivity or origin that no script can test. A fresh-session referee found all of them, against primary sources, the day after the course went up. verify_crypto_course.py now pins the like-for-like ratios.
Found on 2026-10-03 by a fresh-session referee for the desk, not reported. If you had caught it first, this entry would carry your name.
Correction 21 · published 2026-10-01 · corrected 2026-10-01 · live for under a day · caught by the desk
What we said. Nothing in words. For about an hour after the 2026-10-01 publish, the Question on the home page was drawn with each of its four answer options listed twice, and the page's saved snapshot of it held nested copies of its own markers.
What's true. The question has four options. The doubling came from the build step that saves a ready-drawn copy of the page so it shows something before its script runs: the widget added its options without clearing the ones already there, so running it on a saved copy drew them twice. The widget now clears the list before it draws, and the saver strips its own markers before it writes.
Why we missed it. The saved copy was checked for the right words being present, not for each thing appearing once. A check now counts the options in the saved snapshot and fails if there are not exactly four.
Found on 2026-10-01 while checking the published page, not reported. If you had caught it first, this entry would carry your name.
Correction 20 · published 2026-07-29 · corrected 2026-09-30 · live for 63 days · caught by the desk
What we said. After Correction 13, the menu still described this page as “where we were wrong, and paid for it”. The diagram on this page still showed the hostile referee as “a fresh reader paid to attack it”. The Ledger's submission box still read “Checking sign-in status…” on a site with no accounts, and the project README still said an error report “pays the biggest reward we offer”, counted “five games” and used the old tier names.
What's true. Nobody is paid to find an error here and nobody is paid to referee. The referee is a fresh AI session, and the checks are scripts, AI referee passes and then one person, all described on How we check. A reader who reports an error is credited by name if they want that. There are no accounts, so there is no sign-in status to check, and the submission form appears as soon as the page loads.
Why we missed it. Correction 13 searched for the sentences it listed, not for the idea behind them, so the same promise in a menu label, a diagram and a README file, each worded differently, survived it. This is the third time a fix for one wording left another standing, after Correction 2. The menu subtitle is now written in the one script that generates the menu, so it cannot differ from page to page.
Found on 2026-09-30 by the desk, not reported. If you had caught it first, this entry would carry your name.
Correction 19 · published 2026-09-01 · corrected 2026-09-30 · live for 29 days · caught by the desk
What we said. On The Race, the caption of the hardware chart and the notes in our data file said “only the 2024 Google entry” was a below-threshold demonstration, and that “the others” were qubits encoded and operated but not proven to improve with scale.
What's true. Bluvstein et al. (Nature 649, 39; arXiv:2506.20661) report a surface-code memory on up to 448 neutral atoms in which distance 5 had 2.14(13) times lower error per round than distance 3. He et al. (Physical Review Letters 135, 260601, December 2025) report a distance-7 surface code on USTC's Zuchongzhi 3.2 with a suppression factor of 1.40(6). Both are below-threshold results. The Race now lists three, and says which.
Why we missed it. The sentence was written when Google's was the only result we had checked, and was kept as a standing claim about the whole field instead of a dated statement about our own table. The count of below-threshold entries is now read from a flag in data/hardware.json and written into the page by a script, so the sentence cannot fall behind the data, only the data can fall behind the field. Two of the six labs on the old scoreboard still have no entry, and the page says so.
Found on 2026-09-30 by the desk, not reported. If you had caught it first, this entry would carry your name.
Correction 18 · published 2026-08-18 · corrected 2026-09-30 · live for 43 days · caught by the desk
What we said. On The Machinery, topic 2 wrote a mixed state as ρ = Σᵢ pᵢ |ψᵢ⟩⟨ψᵢ⟩, closing the bra with a ket bracket. Topics 2, 6, 7 and 10 described the quantum Fourier transform, Holevo's bound and entanglement entropy as “a later topic”, “still to be built” or “not yet built”, although the page's own header says all 20 topics are live. Topic 8 said “constant vs balanced” covers every 1-bit function because the only other options were “the identity or its flip in disguise”, but the identity and NOT are exactly the two balanced functions. The header counter read “0 claims without a proof or a source”.
What's true. The formula is ρ = Σᵢ pᵢ |ψᵢ⟩⟨ψᵢ|. The Fourier transform is topic 11, and Holevo's bound and von Neumann entropy are topic 15; all were built. There are four 1-bit functions: two constant, two balanced. The counter was never a count. No script can count claims in prose, and only 3 of the 20 topics link a primary paper, so the header now shows that figure, computed from the page.
Why we missed it. The forward references were written while the page was growing and were never revisited when it finished, the bra was a typing slip no check reads, and the zero was typed, not counted. Every count on the page, and the matching subtitles in the menu, now comes from a script, and a check fails if a number drifts. Sourcing the other 17 topics is open work, logged in the build log.
Found on 2026-09-30 by the desk, not reported. If you had caught it first, this entry would carry your name.
Correction 17 · published 2026-07-24 · corrected 2026-09-30 · live for 68 days · caught by the desk
What we said. On the home page and on the Bitcoin page, the March 2026 re-costing of the attack (arXiv:2603.28846) was labelled “Gidney et al.” and credited to “Gidney, Babbush, Drake, Boneh et al.”. The chart’s “today” bar read “a few thousand” with no machine named, and the text put the gap at “three orders of magnitude, not six”. The Bitcoin page also said the attack finishes in about nine minutes “inside the confirmation window”.
What's true. The paper's first author is Ryan Babbush, followed by Adam Zalcman and Craig Gidney. The largest superconducting chip is IBM's Condor, at 1,121 qubits, and Google's Willow, the machine that runs error correction below threshold, has 105. Against those the under-500,000-qubit estimate is at most 2.6 and 3.7 orders of magnitude away, and the 2022 estimate was 6.2. On timing, the paper gives about 9 or 12 minutes from a primed start, against blocks that arrive about every ten minutes, and puts the attacker's chance of winning that race at just under 41% under idealised assumptions. It is not a guarantee of finishing in time.
Why we missed it. The chart's figures were typed by hand from memory of the paper, and its bar was never named, so nothing tied it to the hardware table that already held the real counts. The receipt behind the gap sentence even cited a 1,180-qubit array that the bar did not show. The chart is now generated by a script from the hardware table and the papers' figures, so every bar length and every “orders of magnitude” is computed, and --check fails if a page drifts from it.
Found on 2026-09-30 by the desk, not reported. If you had caught it first, this entry would carry your name.
Correction 16 · published 2026-09-29 · corrected 2026-09-30 · live for 1 day · caught by the desk
What we said. On Inside the machine, at the bottom level, the Josephson junction: the aluminium-oxide barrier is “only a few dozen atoms thick” and “a few thousand times thinner than it is wide”, and its caption said the model draws it “about 5× too thick”. Higher up the same page: that the transmon's protection from charge noise is “the trick that gave the transmon its name”; that a pulse tube gets a machine to “about 4 K” and “cannot go colder”; and two sentences about how large machines are built, one about a control rack growing “long before the fridge does” and one about a cryostat hung from a frame “so the floor never touches the cold parts”.
What's true. The barrier in a transmon junction is aluminium oxide about 1 to 2 nm thick, which is about four to eight atoms, not dozens. It sits in an overlap roughly 100 to 300 nm across, so it is roughly a hundred times thinner than it is wide, not thousands of times; and the drawing exaggerates it about 10 times, not 5, so that it can be seen at all. “Transmon” abbreviates “transmission-line shunted plasma oscillation qubit”. The large shunt capacitor is what makes it insensitive to charge noise, but that is not where the name comes from. A pulse tube reaches a few kelvin, and the better ones a little under 4 K. The two sentences about how large machines are built were things we believed and had not sourced, so they are gone.
Why we missed it. The page was written in one sitting and its figures came from memory. The check for it tested that every part frames, opens and can be reached, and never read a sentence about a size, so three statements about one barrier disagreed with each other and with the drawing and nothing compared them. It now takes the drawn thickness and width from the model, the nanometres from the caption, and fails if the caption's “too thick” factor, the “times thinner” ratio or the atom count disagrees with them.
Found on 2026-09-30 by the desk, not reported. If you had caught it first, this entry would carry your name.
Correction 15 · published 2026-09-01 · corrected 2026-09-24 · live for 23 days · caught by the desk
What we said. In the hardware table and its data file, the November 2024 Atom Computing and Microsoft entry called its 24 logical qubits “the largest logical-qubit count claimed to date”. The data file also still said Google's Willow “is smaller than its own 2019 Sycamore”, the claim retracted as Correction 1.
What's true. The December 2023 entry directly above it lists 48 logical qubits. What Microsoft and Atom Computing claimed was a record for logical qubits entangled together, and the entry now says that. Willow has 105 qubits against Sycamore's 53; Correction 1 fixed three copies of that sentence and missed a fourth, in a caveat inside data/hardware.json that the page does not display but the file publishes.
Why we missed it. The record claim was copied from the announcement without its qualifier. The check that guards this table compared its prose with the data but never tested one row's superlative against the rows around it, and the Correction 1 sweep searched the pages, not the data files behind them. Both are checked now.
Found on 2026-09-24 by the desk, not reported. If you had caught it first, this entry would carry your name.
Correction 14 · published 2026-09-08 · corrected 2026-09-24 · live for 16 days · caught by the desk
What we said. The Signal on IBM and the University of Chicago said they “ran 70 logical qubits” with 468 T gates and 2,415 logical two-qubit operations, and that “because the circuit was error-corrected” its error rate came out ten times lower.
What's true. The paper uses error detection: runs where the code flagged an error were thrown away, and the tenfold improvement is measured on the runs that were kept. Nothing was corrected. Its July version did describe a 70-qubit circuit with 468 T gates, but the authors revised it in September to a 64-qubit circuit with 314 T gates on 76 physical qubits. The Signal no longer quotes a two-qubit gate count, which we could not confirm against the paper's abstract. It now reports the current version, says what changed, and explains detection versus correction.
Why we missed it. The story was written from the first version of the paper and a news report of it, and the report's framing was carried over. Nobody re-read the abstract for the one word the whole site is about, and nothing watches a cited preprint for new versions.
Found on 2026-09-24 by the desk, not reported. If you had caught it first, this entry would carry your name.
Correction 13 · published 2026-07-29 · corrected 2026-09-24 · live for 57 days · caught by the desk
What we said. The homepage offered a weekly letter (“the first issue will find you”) and a Discord where your streak would “count”, and a box after a won game offered the same letter. Every page could show a sign-in button. Most pages' footers said a reported error is “the biggest bounty we pay”. The Arcade announced a board game as “designed”. The Frontier described answers reviewed by a second person and a panel of AI models. The About page said our calculations live in “open code that runs again before every release”, that three checks can each “stop a release”, that the correction report “always gets a reply”, and that “everything” is “checked twice”.
What's true. There is no newsletter yet; the form only delivered an address to our inbox. The Discord has two members, and streaks have never counted anywhere but your own browser. No sign-in has ever completed, because the database behind it has never been set up. There is no bounty. The board game has no design yet. The Frontier's model panel has never run. Most of our checking scripts live in a private repository and are run by the desk, not on every release; one guard runs automatically on each change. All of these are now hidden or reworded, and The Frontier is unlisted until its panel runs. What is still offered works today: the correction report, the Ledger's claim form, the games, and the progress your browser keeps for you.
Why we missed it. Each feature was written as if the service behind it was about to exist, and the copy was never taken down when it didn't. The checks on this site compare numbers with their sources and sentences with their tables; nothing compared a promise with whether it could be kept.
Found on 2026-09-24 by the desk, not reported. If you had caught it first, this entry would carry your name.
Correction 12 · published 2026-08-27 · corrected 2026-09-24 · live for 28 days · caught by the desk
What we said. On The Ledger, rule 2: “Whoever captures a claim does not resolve it. A second person writes the verdict.” On the About page: “Resolving a claim takes two people”, claims are “judged by rules written before the answer is known”, and criteria are frozen “long before its deadline”.
What's true. SymbiQ is one person working with AI tools. The two verdicts published on 22 September were drafted by an AI model, in a pass separate from the one that logged the claims, and published by the desk; no second person reviewed them. Both claims were also logged on 27 August 2026, eight months after their 31 December 2025 deadlines, so their criteria were written when the outcome could already be known. Rule 2 now says what happens: a verdict is written after the deadline, in its own pass, against the criteria frozen when the claim was logged, and it cites its evidence. The About page and both claims now say that these two were logged after their deadlines.
Why we missed it. The rules were written for a desk of more than one person and for claims logged as they are made. The first verdicts passed our checker's version of rule 2, which only tests that the verdict's author field differs from the capture field. Nothing asked who those authors were, or put each claim's capture date beside its deadline.
Found on 2026-09-24 by the desk, not reported. If you had caught it first, this entry would carry your name.
Correction 11 · published 2026-09-22 · corrected 2026-09-24 · live for 2 days · caught by the desk
What we said. On The Ledger, rule 5: “Every claimant gets a chance to respond before a Not verified verdict ships.”
What's true. The first two verdicts, on IonQ and on Atom Computing with Microsoft, were published as Not verified on 22 September without either company being contacted first. The rule now says what the Ledger does: any claimant can answer a verdict at any time, the reply is published in full beside it, and a reply that shows the verdict is wrong changes it, with the change logged here. Both verdicts are open to a reply on those terms.
Why we missed it. Publishing a verdict ran through the checks a machine can do: criteria frozen and sources archived. Contacting the company is not one of those, and no step in the process owned it, so the verdicts shipped on the checks that did run.
Found on 2026-09-23 by the desk, not reported. If you had caught it first, this entry would carry your name.
Correction 10 · published 2026-08-27 · corrected 2026-09-24 · live for 28 days · caught by the desk
What we said. At the top of The Ledger: every claim is “logged the day it's made”. The page's search and link-preview descriptions said claims are “tracked from the moment they're made”, and the homepage said “predictions are logged the day they're made”.
What's true. None of the twenty claims was logged on the day it was made. The first ten were captured on 27 August 2026 from statements made between July 2024 and June 2026, and the next ten on 23 September 2026 from statements going back to January 2024. What each entry does carry is the date the claim was made, a quote checked against the source, and an archived copy of that source, so the gap between the claim and our logging of it is visible rather than hidden. The tagline, all three descriptions and the homepage now say that.
Why we missed it. The sentence described the Ledger we meant to run, a desk that logs claims as they are announced, and was written before any claim was captured. Seeding it from statements that were already months old made it false on the first day, and nothing compared the tagline with the dates sitting in the entries below it.
Found on 2026-09-23 by the desk, not reported. If you had caught it first, this entry would carry your name.
Correction 9 · published 2026-09-22 · corrected 2026-09-23 · live for 1 day · caught by the desk
What we said. Lower down this page, under “Claims we got wrong on The Ledger”: “Two claims have now passed their resolution date and carry a drafted verdict awaiting a second reviewer.”
What's true. Both verdicts were published on 22 September, as Not verified, and neither was waiting on anything. The sentence was typed by hand when the verdicts were drafted and was not touched when they shipped. It no longer states a count or a status. It says what void means and that every verdict so far was decided by its own pre-registered criteria.
Why we missed it. The same failure as Correction 5: a hand-typed number about data that lives somewhere else. Publishing the verdicts updated the claim files, the Ledger index and the scorecards. Nothing connected that step to a sentence on this page.
Found on 2026-09-23 by the desk, not reported. If you had caught it first, this entry would carry your name.
Correction 8 · published 2026-08-27 · corrected 2026-09-23 · live for 27 days · caught by the desk
What we said. The Ledger entry for Jensen Huang's remark that useful quantum computers are 15 to 30 years away quoted him in full. It gave Quantum Computing Report's article on the remark as the source, with a note saying that article “directly transcribes the Q&A exchange.”
What's true. That article discusses the remark and never prints it. None of its twelve archived copies, from January 2025 to this month, contains the quoted words. The quote is real: Constellation Research printed the same three sentences on 8 January 2025, and an archived copy from the next day still shows them. The entry now cites that report and links the copy, and its sourcing note records the change. The claim and its resolution criteria are unchanged. We also logged something the entry had missed: Huang walked the remark back in March 2025, without naming a new range.
Why we missed it. The quote was repeated so widely that it looked checked. The sourcing note described the source we meant to cite, and nobody opened the linked page to look for the words. It surfaced on 23 September, when every Ledger source was archived for the first time and each copy was searched for the text it is cited for.
Found on 2026-09-23 by the desk, not reported. If you had caught it first, this entry would carry your name.
Correction 7 · published 2026-08-27 · corrected 2026-09-23 · live for 27 days · caught by the desk
What we said. On The Ledger, rule 3: “Every source gets an independent snapshot the moment it's captured, so an edited or deleted press release can't make a claim unprovable later.”
What's true. All ten claims shipped with no archived copy, and so did the evidence behind the first two verdicts, published on 22 September. Both verdicts say a named company missed a dated claim, and one leaned partly on that company's own investor-relations pages, which is the situation the rule exists to prevent. Every source and every piece of verdict evidence now links to a dated Wayback Machine copy. Seventeen of the eighteen copies were taken on or before the day we relied on the page. The eighteenth, the 2 September article behind the Atom Computing verdict, had never been archived successfully. It was saved on 23 September, a day after the verdict, and its link shows that date.
Why we missed it. The Ledger already showed an “archived” link whenever the field was filled, so the feature looked finished. The checker that guards the claim files tested the other rules on that list and not this one, so an empty field passed. It now fails any tracking claim without an archived source, and any verdict whose evidence lacks one.
Found on 2026-09-23 by the desk, not reported. If you had caught it first, this entry would carry your name.
Correction 6 · published 2026-08-27 · corrected 2026-09-21 · live for 25 days · caught by the desk
What we said. What is a qubit actually made of? introduces The Calibration in its own words — “the game below is what recalibrating against that drift looks like” — and the page shipped with a game below that sentence.
What's true. There was no game below that sentence for most readers. Its heading sat immediately after the page's 🔴 marker and carried no data-tier-keep, so tiers.js filed the entire section under 🔴 and hid it from anyone who had not chosen that depth. At the page's default depth, and at 🟢, and at 🟡, the prose promised a game and then stopped. Only a reader who switched to 🔴, or who arrived on the #calibration link from the Arcade, ever saw it. Every other interactive widget on this site carries that attribute for exactly this reason. It does now too, and the detuned-Rabi formula that used to open the section has moved up into the 🔴 derivation where it belongs, so the game is introduced in language that works at any depth.
Why we missed it. Nothing was broken, so nothing failed. The HTML was valid, the game mounted correctly when reached, every checker passed, and the Arcade's own deep link worked — which is the link a builder clicks when testing. The defect only exists for a reader who scrolls the page normally at the depth it opens at, and no check, and no author, had ever done that. It was found while measuring something else entirely: the widget reported zero height in an automated layout pass, which was supposed to be a measurement and turned out to be a symptom. A check now loads the page at every depth and fails if the game is not on screen.
Found on 2026-09-21 by the desk, not reported. If you had caught it first, this entry would carry your name.
Correction 5 · published 2026-08-27 · corrected 2026-09-08 · live for 12 days · caught by the desk
What we said. On The Ledger, in the tagline at the top: “Nothing has resolved yet: the Ledger opened with 7 claims still tracking.”
What's true. Ten claims were on the page. Seven was right on the day the Ledger opened and was never touched again — and the element holding it, <span id="ldg-seed-count">, looked dynamic while being nothing of the kind: no JavaScript anywhere on this site ever wrote to that id. Two of the ten had also passed their resolution deadline by then, so “still tracking” was wrong about those two in a second, worse way. The sentence is now computed from data/claims/index.json when the page loads, and states the live position — how many claims are tracked, how many have resolved, and how many have a verdict drafted and waiting on a second reviewer.
Why we missed it. This page's whole promise is that its numbers are computed rather than asserted, and its own headline number was asserted, sitting directly above ten claims that each carried a machine-checked schema. A check validated every claim file and cross-checked them against the index and passed every time, because the wrong number was in prose — and prose is the one thing none of this site's checkers read. That is the identical failure to Correction 1, one page over, which is the part worth admitting. The fix is not a better number: the hand-typed one is gone, replaced by a computed one, so there is no longer anything there that can go stale.
Found on 2026-09-08 while reviewing an outside critique of the site, not reported. If you had caught it first, this entry would carry your name.
Correction 4 · published 2026-08-21 · corrected 2026-09-08 · live for 18 days · caught by the desk
What we said. On the home page, as a section heading and in the description that search engines and social cards quote: “Nine games where you cannot fake a score.”
What's true. Twelve. Six cabinets in The Arcade — Circuit Golf, Grover's Escape, Max-Cut, the Annealing Volcano, quantum tic-tac-toe and The Calibration — plus six more playable things out on the pages they belong to: The Bench, The Decoder Duel, the CHSH game, The Question, The Bottleneck and The Prune. The Arcade had been counting correctly the whole time (“six games”, “six more playable things”), so for eighteen days two of our own pages contradicted each other in public. Nine was true when it was written; three games shipped after it and nothing went back.
Why we missed it. The same shape as the correction above it and as Correction 1: a count written once, by hand, in prose, on a page nothing recomputes — and this one was also in the meta description, so the wrong number was what the site said about itself everywhere it was linked. A check now derives the real figure from the cabinet list in games.js and the entries in play.html, and fails if any of the three prose counts — the Arcade's six, the six out in the wild, or the home page's total — stops matching the games that actually exist.
Found on 2026-09-08 while reviewing an outside critique of the site, not reported. If you had caught it first, this entry would carry your name.
Correction 3 · published 2026-08-23 · corrected 2026-09-04 · live for 12 days · caught by the desk
What we said. On the Ising-machine page, under the “Let it settle” widget: “that enumeration also shows there is only one discrete local minimum on this graph, the global optimum itself.”
What's true. There are four minimising spin assignments — two arrangements and their mirror images. The widget's own readout says so: it prints the count, and on this graph it prints “4 ways.” The argument the sentence was making survives intact, and is the part that matters: every discrete local minimum on this graph is a global optimum, so a settle that falls short has genuinely not found some other valid arrangement and stopped there. Only the count was wrong.
Why we missed it. The sentence was written during a referee pass that was checking something else and got the surrounding physics right. Nothing compared it against the widget three inches above it, which computes the number at page load and displays it. A check now enumerates all 32 assignments, checks every discrete local minimum is a global optimum, confirms the widget's settled states really are stationary out to 100,000 steps, and measures what its read-off actually achieves — about 70% of runs reach the optimum, and the shortfall is the unpinned global phase rather than a worse arrangement.
Found on 2026-09-04 while writing that verification script, not reported. If you had caught it first, this entry would carry your name.
Correction 2 · published 2026-08-18 · corrected 2026-09-04 · live for 17 days · caught by the desk
What we said. On Basics, in three places: “Google measured Λ = 2.14 on Willow, going from d = 5 to d = 7”; “Λ = 2.14 ± 0.02 on Willow is quoted for the d = 5 → d = 7 step… it is not an average over 3 → 5 → 7”; and, in the source list, “measured d = 5 → 7.”
What's true. Google obtains Λ = 2.14 ± 0.02 by linear regression of ln(εd) against d across the whole d = 3, 5, 7 series (Nature 638, 920 (2025); arXiv:2408.13687). It is a common-ratio parameter of the family, not a measurement of one step — and the step we named is the one furthest from it: on the paper's own decoder rates, ε3/ε5 = 2.15, ε5/ε7 = 2.12. The final sentence was the worst of it: “it is not an average over 3 → 5 → 7” asserted, specifically and confidently, the opposite of what the paper does.
Why we missed it. We didn't. A referee pass caught this exact claim on 30 August and fixed it on The Race and the QEC page. Nobody grepped the rest of the site, so the third copy — on the page a beginner is most likely to read first — stayed live five more days, and this is the most-repeated number we publish. A correction is not finished when the page you were looking at is fixed. There is now a site-wide guard: A check fails the build if any page attributes Λ = 2.14 to a single distance step, or reinstates the retracted wording.
Found on 2026-09-04 while adding receipts to this page, not reported. If you had caught it first, this entry would carry your name.
Correction 1 · published 2026-09-01 · corrected 2026-09-04 · live for 3 days · caught by the desk
What we said. On The Race: “Google's Willow (2024) is smaller than its own 2019 chip.” The milestone table said the same thing in its own words — “Smaller than Sycamore.”
What's true. Willow has 105 physical qubits (Google Quantum AI, Nature 638, 920 (2025); arXiv:2408.13687). Sycamore had 53 (Arute et al., Nature 574, 505 (2019); arXiv:1910.11333). Willow is about twice Sycamore's size, not smaller — and both numbers were sitting in the table directly under the sentence. The comparison the paragraph was actually reaching for is against 2023's biggest chips: Willow holds under a tenth the qubits of Atom Computing's 1,180-site array or IBM's 1,121-qubit Condor. The point survived; the evidence for it did not. The frontier really did stop competing on raw count. We just proved it with a number that says the opposite.
Why we missed it. The wrong figure was never computed — it was asserted in prose, and prose is the one thing none of this site's checkers read. A check validated the table against data/hardware.json, and the chart's inline copy against both, and passed every time, because all three agreed with each other and none of them was the sentence. Worse, the same wrong claim had been copied into three places (the JSON note, the inline chart data, the visible table row), so on any spot-check it read as corroborated. The checker now derives the page's comparative sentences from the milestone data — the twenty-fold growth, the under-a-tenth comparison, Willow against Sycamore — and fails if any of them stops being true, or if the retracted wording ever reappears.
Found during unrelated work on 2026-09-04, not reported. If you had caught it first, this entry would carry your name.
Every entry follows the same three lines: what we said → what’s true → why we missed it. Corrections found by readers are credited by name with the 🏅 “taught it better than us” tag, and run with more room than the claim they fix.
Different log, same rule. The Ledger tracks public claims from labs, executives
and regulators to a dated verdict, and when we write resolution criteria that turn out unscoreable,
that goes on our record as void, with the same prominence as any correction above, not
quietly dropped. This section mirrors the Ledger's own scorecard rather than
duplicating it by hand.
A claim is void only when our criteria turn out unscoreable. Every verdict published so far was decided by its own pre-registered criteria.
The most useful thing you can send us. What you get is your name on the correction, published with more room than the claim it fixes.
You don't need to be certain, and you don't need to be polite about it. "This number looks off" is a perfectly good report. If you turn out to be right, the correction runs with your name on it.
Read by a human, always.
Every 🔴 topic gets an independent review before it ships, and facts from secondary coverage are checked against primary sources. Those passes returned hundreds of findings, all fixed before publication. This page counts only what got past them.
← swipe the diagram to see all of it →