---
slug: the-error-they-share
passed: 2026-10-18
by: the assay line (research/assay-line/claude-1018b/), Claude fixer (claude-opus-5-5) directed by claude-assaying-boyle-1014; source record by a Claude agent (claude-sonnet-5-5), checked by a Claude agent (claude-opus-5-5)
checked: 185
confirmed: 171        # 157 CONFIRMED + 14 MINOR
wrong: 4
unverifiable: 0
observed: 7
stale: 3
---

This was a directed repair pass over the 21 triage items in research/assay-line/claude-1018b/the-error-they-share/triage.json, read at e3d6ec7c89. The checker (N2) re-read the layer against the N1 source record at 9ddc1e9045 and raised four WRONG lines, fourteen MINOR and three STALE; the claim lines below are its record, with the fixer's outcome appended to each triage line. The fixer re-read the evidence for every WRONG: score.mjs's position test and a recount from results.json (59 wrong answers, chi2 1.051; the 470 chosen slots give 1.200), the verifier's own code, the Neitz, Geist and Jacobs abstract (PMID 2487095) beside the dog page's 2026-10-14 correction, and the page's own account of the dropped closest-planet question. All four were fixed with dated Corrected lines, the three STALE items share one Updated line, two MINOR items were fixed because their sentences were already being rewritten, and twelve MINOR items are listed. The page is generated by research/myth-consensus/build-page.mjs but carries blocks added by shared post-processors, so the generator and the published page were edited identically (a scratch rerun of the generator reproduces the edited markdown byte for byte and the edited quiz and corrections exactly). After the fixes verify-the-error-they-share.mjs passes (37 passed, 0 failed; it failed on the two moved quotes before).

## Claims
- CONFIRMED | The real cause is the 23.4° tilt of Earth's axis - and you can prove it below by switching each effect off and watching which one the seasons actually depend on. | public/strata/what-causes-the-seasons/index.html (git show 9ddc1e9045) | [quote-seasons] present on the cited page at the pin (verify-the-error-they-share.mjs check 1 finds every quote except dog-colour and gum); the page shows it as the quiz Why line, identical to items.json evidence.checked
- CONFIRMED | The real cause is the 23.4° tilt of Earth's axis - and you can prove it below by switching each effect off and watching which one the seasons actually depend on. | https://en.wikipedia.org/wiki/Season | [world-seasons] sources.mjs FOUND; "It is Earth's axial tilt that causes the Sun to be higher in the sky during the summer months"; tropics at 23.4 degrees. Same claim, same strength.
- CONFIRMED | The near-side spot logs about 14.8 days of sunlight and 14.8 of night. The far-side spot logs exactly the same … There is no hemisphere the Sun neglects. | public/strata/the-moon-has-no-dark-side/index.html (git show 9ddc1e9045) | [quote-far-side] present on the cited page at the pin (verify-the-error-they-share.mjs check 1 finds every quote except dog-colour and gum); the page shows it as the quiz Why line, identical to items.json evidence.checked
- MINOR | The near-side spot logs about 14.8 days of sunlight and 14.8 of night. The far-side spot logs exactly the same … There is no hemisphere the Sun neglects. | https://en.wikipedia.org/wiki/Far_side_of_the_Moon | [world-far-side] excerpt FOUND but weaker than the page: "both the near and far sides receive, on average, almost equal amounts of light", and the next sentences: "Lunar eclipses mean that the side facing Earth receives fractionally less sunlight than the far side when considered over a long period of time." The quote says the far-side spot logs "exactly the same" and the quiz checked option is "Exactly as much, sunlit half of every month". The author's own log concedes it (memory/log.d/2026-08-14T2330Z-the-error-they-share.md:51-52, "The Moon's far side really does get slightly more sun than a near side Earth periodically eclipses"). Fixer (K1): declined, tier B listed.
- CONFIRMED | Its angular size is the same all over the sky - and at the horizon it's actually about 1.6% smaller, because you're sitting roughly one Earth-radius farther from it. | public/strata/moon-illusion/index.html (git show 9ddc1e9045) | [quote-moon-illusion] present on the cited page at the pin (verify-the-error-they-share.mjs check 1 finds every quote except dog-colour and gum); the page shows it as the quiz Why line, identical to items.json evidence.checked
- CONFIRMED | Its angular size is the same all over the sky - and at the horizon it's actually about 1.6% smaller, because you're sitting roughly one Earth-radius farther from it. | https://en.wikipedia.org/wiki/Moon_illusion | [world-moon-illusion] FOUND: "only about 1.5% farther away and thus smaller in angular size when at the horizon". Page says about 1.6%; one Earth radius over the mean distance is 6371/384400 = 1.66%, so 1.6% is the closer rounding of the same fact.
- CONFIRMED | Water is blue on its own, in the dark, in a white-walled pool, because it absorbs red light. | public/strata/why-water-is-blue/index.html (git show 9ddc1e9045) | [quote-water-blue] present on the cited page at the pin (verify-the-error-they-share.mjs check 1 finds every quote except dog-colour and gum); the page shows it as the quiz Why line, identical to items.json evidence.checked
- CONFIRMED | Water is blue on its own, in the dark, in a white-walled pool, because it absorbs red light. | https://en.wikipedia.org/wiki/Color_of_water | [world-water-blue] FOUND; red component absorbed in a white pool, the intrinsic-colour claim.
- CONFIRMED | The air going over the top reaches the trailing edge first, with time to spare, and never meets its partner again. | public/strata/the-air-that-got-there-first/index.html (git show 9ddc1e9045) | [quote-equal-transit] present on the cited page at the pin (verify-the-error-they-share.mjs check 1 finds every quote except dog-colour and gum); the page shows it as the quiz Why line, identical to items.json evidence.checked
- CONFIRMED | The air going over the top reaches the trailing edge first, with time to spare, and never meets its partner again. | https://en.wikipedia.org/wiki/Lift_(force) | [world-equal-transit] FOUND: "the upper markers arriving at the trailing edge long before the lower ones". Supports top-first and no rejoining.
- CONFIRMED | Liquid water absorbs over a broad band by dielectric relaxation - polar molecules trying to follow the flipping field and lagging behind. | public/strata/why-microwaves-heat-food/index.html (git show 9ddc1e9045) | [quote-microwave] present on the cited page at the pin (verify-the-error-they-share.mjs check 1 finds every quote except dog-colour and gum); the page shows it as the quiz Why line, identical to items.json evidence.checked
- CONFIRMED | Liquid water absorbs over a broad band by dielectric relaxation - polar molecules trying to follow the flipping field and lagging behind. | https://en.wikipedia.org/wiki/Microwave_oven | [world-microwave] FOUND: not a resonance; "all polar molecules participate, and dielectric heating ... over a wide range of frequencies". Same claim.
- CONFIRMED | Nothing about it is cold - it's just better at stealing the heat from your hand. | public/strata/metal-colder-than-wood/index.html (git show 9ddc1e9045) | [quote-metal-cold] present on the cited page at the pin (verify-the-error-they-share.mjs check 1 finds every quote except dog-colour and gum); the page shows it as the quiz Why line, identical to items.json evidence.checked
- CONFIRMED | Nothing about it is cold - it's just better at stealing the heat from your hand. | https://en.wikipedia.org/wiki/Thermal_effusivity | [world-metal-cold] FOUND: similar temperatures, the high-effusivity metal "is detected as cool". Same claim.
- CONFIRMED | Breathing helium does not raise the pitch of your voice, it changes the timbre. Your vocal folds buzz at the same rate, but sound travels about 2.9 times faster in helium, so the resonances of your vocal tract slide upward. | src/content/strata/helium-voice.md (git show 9ddc1e9045) | [quote-helium] present on the cited page at the pin (verify-the-error-they-share.mjs check 1 finds every quote except dog-colour and gum); the page shows it as the quiz Why line, identical to items.json evidence.checked
- CONFIRMED | Breathing helium does not raise the pitch of your voice, it changes the timbre. Your vocal folds buzz at the same rate, but sound travels about 2.9 times faster in helium, so the resonances of your vocal tract slide upward. | https://en.wikipedia.org/wiki/Helium | [world-helium] FOUND: sound "nearly three times" faster (page 2.9), fundamental frequency "does not change", timbre changes.
- CONFIRMED | Then why is the old window thicker at the bottom? Because of how it was made. | public/strata/glass-is-liquid/index.html (git show 9ddc1e9045) | [quote-glass-flow] present on the cited page at the pin (verify-the-error-they-share.mjs check 1 finds every quote except dog-colour and gum); the page shows it as the quiz Why line, identical to items.json evidence.checked
- CONFIRMED | Then why is the old window thicker at the bottom? Because of how it was made. | https://en.wikipedia.org/wiki/Glass | [world-glass-flow] FOUND: sags "were already there the day it was made"; non-uniform thickness from manufacture.
- CONFIRMED | A slab does not dry to get hard. It reacts: cement and water grow into interlocking crystals that consume the water. | public/strata/does-concrete-dry/index.html (git show 9ddc1e9045) | [quote-concrete] present on the cited page at the pin (verify-the-error-they-share.mjs check 1 finds every quote except dog-colour and gum); the page shows it as the quiz Why line, identical to items.json evidence.checked
- CONFIRMED | A slab does not dry to get hard. It reacts: cement and water grow into interlocking crystals that consume the water. | https://en.wikipedia.org/wiki/Concrete | [world-concrete] FOUND: cement reacts with water by hydration and hardens. Supports react-not-dry.
- CONFIRMED | Oxygen-rich blood is bright scarlet; oxygen-poor venous blood is dark maroon. | public/strata/why-veins-look-blue/index.html (git show 9ddc1e9045) | [quote-veins] present on the cited page at the pin (verify-the-error-they-share.mjs check 1 finds every quote except dog-colour and gum); the page shows it as the quiz Why line, identical to items.json evidence.checked
- CONFIRMED | Oxygen-rich blood is bright scarlet; oxygen-poor venous blood is dark maroon. | https://en.wikipedia.org/wiki/Vein | [world-veins] FOUND: venous blood "is dark red"; page says dark maroon, a colour name for the same thing.
- CONFIRMED | taste buds across the whole tongue and palate contain cells that respond to all five basic qualities - sweet, salty, sour, bitter, umami | public/strata/the-zones-that-were-never-there/index.html (git show 9ddc1e9045) | [quote-tongue-map] present on the cited page at the pin (verify-the-error-they-share.mjs check 1 finds every quote except dog-colour and gum); the page shows it as the quiz Why line, identical to items.json evidence.checked
- CONFIRMED | taste buds across the whole tongue and palate contain cells that respond to all five basic qualities - sweet, salty, sour, bitter, umami | https://en.wikipedia.org/wiki/Tongue_map | [world-tongue-map] FOUND: "all parts are equally capable of conveying the qualia of all tastes".
- STALE | A human retina has three cone pigments; a dog has two, with measured peaks at 429 nm (a blue cone) and 555 nm (a yellow-green cone). Missing the long-wave "red" cone you have, a dog can't separate red from green | git show 368a1b9cd53^:public/strata/the-colours-the-dog-keeps/index.html | [quote-dog-colour] the sentence was on the dog page when this layer was built (2026-08-14) and was removed there on 2026-10-14 (commit 368a1b9cd53); at 9ddc1e9045 the verifier prints "missing: dog-colour/checked". The quiz still presents it as the words of "the layer that checked it". The sentence is also false, see world-dog-colour. Fixer (K2): dated 2026-10-18 (Updated line).
- WRONG | A human retina has three cone pigments; a dog has two, with measured peaks at 429 nm (a blue cone) and 555 nm (a yellow-green cone). Missing the long-wave "red" cone you have, a dog can't separate red from green | https://eutils.ncbi.nlm.nih.gov/entrez/eutils/efetch.fcgi?db=pubmed&id=2487095&retmode=text&rettype=abstract ; public/strata/the-colours-the-dog-keeps/index.html | [world-dog-colour] excerpt FOUND and it contradicts the second sentence: the dog has "two classes of cone photopigment ... spectral peaks of about 429 nm and 555 nm"; the 555 nm cone is a long/medium-wave cone, so the dog is not "Missing the long-wave red cone you have". The dog page itself now says "Corrected 2026-10-14: The page said dogs lack a long-wave cone. They retain a long/medium-wave cone near 555 nm and a short-wave cone near 429 nm". This page still prints the old sentence (index.html line 433, QUIZ dog-colour why; items.json evidence.checked). The checked option "Blue and yellow; red and green look alike" is fine. Fixer (K3): fixed 2026-10-18.
- CONFIRMED | and then nothing happens to the rate until about 63, where it finally starts to ebb at roughly 0.7% a year. | public/strata/the-metabolism-that-didnt-slow/index.html (git show 9ddc1e9045) | [quote-metabolism] present on the cited page at the pin (verify-the-error-they-share.mjs check 1 finds every quote except dog-colour and gum); the page shows it as the quiz Why line, identical to items.json evidence.checked
- CONFIRMED | and then nothing happens to the rate until about 63, where it finally starts to ebb at roughly 0.7% a year. | https://researchonline.ljmu.ac.uk/id/eprint/15386/1/Pontzer%20et%20al%202021%20Accepted.pdf | [world-metabolism] UNREADABLE to the tool; I downloaded the PDF and ran pdftotext: "Segmented regression analysis identified a break point at 63.0 y (95% CI: 60.1, 65.9), after which adjusted TEE begins to decline" and "Adjusted total expenditure declined by -0.7 +/- 0.1% per year"; "stable from age 20 to 60". Both parts of the excerpt are there and support the claim.
- CONFIRMED | The head is about 9% of your skin, and it sheds about 9-10% of your heat. | public/strata/heat-lost-through-your-head/index.html (git show 9ddc1e9045) | [quote-head-heat] present on the cited page at the pin (verify-the-error-they-share.mjs check 1 finds every quote except dog-colour and gum); the page shows it as the quiz Why line, identical to items.json evidence.checked
- CONFIRMED | The head is about 9% of your skin, and it sheds about 9-10% of your heat. | https://en.wikipedia.org/wiki/List_of_common_misconceptions_about_science,_technology,_and_mathematics | [world-head-heat] FOUND: "Heat is lost from the body in proportion to the amount of exposed skin. The head accounts for around 7-9% of the body's surface"; the page's about 9% of skin and 9-10% of heat sit at the top of that range, and the quiz answer (About 10 percent, against 4, 45, 60) is the right one.
- CONFIRMED | A peptic ulcer is, overwhelmingly, an infection (a spiral bacterium, Helicobacter pylori) or the slow erosion of long-term painkillers. | public/strata/does-spicy-food-cause-ulcers/index.html (git show 9ddc1e9045) | [quote-ulcers] present on the cited page at the pin (verify-the-error-they-share.mjs check 1 finds every quote except dog-colour and gum); the page shows it as the quiz Why line, identical to items.json evidence.checked
- CONFIRMED | A peptic ulcer is, overwhelmingly, an infection (a spiral bacterium, Helicobacter pylori) or the slow erosion of long-term painkillers. | https://www.niddk.nih.gov/health-information/digestive-diseases/peptic-ulcers-stomach-ulcers/symptoms-causes | [world-ulcers] FOUND: "The most common causes of peptic ulcers are Helicobacter pylori (H. pylori) infection and nonsteroidal anti-inflammatory drugs".
- STALE | Ordinary peristalsis pushes the base out in a day or two, on the same schedule as everything else you eat. | git show 3ceca30e6fa^:public/strata/does-swallowed-gum-stay-in-your-stomach/index.html | [quote-gum] the sentence was on the gum page at 2026-08-14 and was changed there on 2026-09-27 (3ceca30e6fa) to "on the ordinary gut-transit clock (perhaps a little slower than most food)", with a dated Corrected line; the verifier prints "missing: gum/checked". The quiz still shows the old words as the checking page's. Fixer (K4): dated 2026-10-18 (Updated line).
- MINOR | Ordinary peristalsis pushes the base out in a day or two, on the same schedule as everything else you eat. | https://en.wikipedia.org/wiki/Chewing_gum ; https://www.scientificamerican.com/article/fact-or-fiction-chewing-gum-takes-seven-years-to-digest/ | [world-gum] the Wikipedia excerpt is there and says "as quickly as any other food", but the page this line is quoted from withdrew "on the same schedule as everything else you eat" as overstated on the evidence of Milov in Scientific American, which I fetched: "It probably passes through slower than most foodstuffs, but eventually the normal housekeeping waves in the digestive tract will sort of push it through". The quiz answer (one to two days) stands; the clause does not. Fixer (K5): fixed 2026-10-18.
- CONFIRMED | in Miranda & Schaffner (2016) transfer was measurable at the shortest tested contact time, <1 s, so there is no safe window and no cliff at 5 s | public/strata/is-the-5-second-rule-real/index.html (git show 9ddc1e9045) | [quote-five-second] present on the cited page at the pin (verify-the-error-they-share.mjs check 1 finds every quote except dog-colour and gum); the page shows it as the quiz Why line, identical to items.json evidence.checked
- CONFIRMED | in Miranda & Schaffner (2016) transfer was measurable at the shortest tested contact time, <1 s, so there is no safe window and no cliff at 5 s | https://eutils.ncbi.nlm.nih.gov/entrez/eutils/efetch.fcgi?db=pubmed&id=27590818&retmode=text&rettype=abstract | [world-five-second] FOUND: "Some transfer takes place instantaneously, at times of <1 s, disproving the five-second rule."
- CONFIRMED | The line is real. It is in Jean-Jacques Rousseau's Confessions, Book VI, where he reaches for a half-remembered anecdote about an unnamed aristocrat | public/strata/let-them-eat-cake/index.html (git show 9ddc1e9045) | [quote-cake] present on the cited page at the pin (verify-the-error-they-share.mjs check 1 finds every quote except dog-colour and gum); the page shows it as the quiz Why line, identical to items.json evidence.checked
- CONFIRMED | The line is real. It is in Jean-Jacques Rousseau's Confessions, Book VI, where he reaches for a half-remembered anecdote about an unnamed aristocrat | https://en.wikipedia.org/wiki/Let_them_eat_cake | [world-cake] FOUND: book six of the Confessions, words of an unnamed "great princess". Supports the quote and the checked option.
- CONFIRMED | When Caesar reformed the calendar in 46 BCE he added exactly ten days to reach 365, sprinkled across the other months, and never touched February. | public/strata/why-february-has-28-days/index.html (git show 9ddc1e9045) | [quote-february] present on the cited page at the pin (verify-the-error-they-share.mjs check 1 finds every quote except dog-colour and gum); the page shows it as the quiz Why line, identical to items.json evidence.checked
- CONFIRMED | When Caesar reformed the calendar in 46 BCE he added exactly ten days to reach 365, sprinkled across the other months, and never touched February. | https://en.wikipedia.org/wiki/Julian_calendar | [world-february] FOUND: ten days added to a 355-day year, February "not changed in ordinary years". The lead says the calendar "was proposed in 46 BC" by Caesar, matching 46 BCE.
- CONFIRMED | But the velocities that made Hubble's famous plot were measured by Vesto Slipher more than a decade earlier - and Hubble's paper names no source for them. | public/strata/who-discovered-the-expanding-universe/index.html (git show 9ddc1e9045) | [quote-redshifts] present on the cited page at the pin (verify-the-error-they-share.mjs check 1 finds every quote except dog-colour and gum); the page shows it as the quiz Why line, identical to items.json evidence.checked
- CONFIRMED | But the velocities that made Hubble's famous plot were measured by Vesto Slipher more than a decade earlier - and Hubble's paper names no source for them. | https://en.wikipedia.org/wiki/Hubble%27s_law | [world-redshifts] FOUND: redshifts "earlier measured and related to velocity by Vesto Slipher in 1917"; 1912-1917 to 1929 is more than a decade.
- CONFIRMED | and Hubble's paper names no source for them. | https://archive.org/download/hubble736874/A%20relation%20between%20distance%20and%20radial%20velocity%20among%20extra-galactic%20nebulae%20by%20Edwin%20Hubble.pdf | [world-redshifts-hubble-paper] UNREADABLE to the tool; I read it. The excerpt is at pp. 169-170. No "Slipher" (nor "lip") in the text layer or in a fresh tesseract OCR of pp. 169-170; positive controls in the same text: Humason (2), Lundmark, Shapley. The paper does credit Humason for four velocities (N.G.C. 6822, 221, 224, 5457), so "names no source" is true of the Slipher velocities, which is what "them" refers to.
- CONFIRMED | But the velocities that made Hubble's famous plot were measured by Vesto Slipher more than a decade earlier | https://en.wikipedia.org/wiki/Vesto_Slipher | [world-redshifts-slipher-first] FOUND: "In 1912, he was the first to observe the shift of spectral lines of galaxies".
- CONFIRMED | 21 questions this project had already checked | public/data/the-error-they-share/items.json ; src/content/strata/ (the 21 stratum files) | [dek-21-questions] 21 items; the 21 stratum pages are dated 2026-06-22 to 2026-07-16, all before this layer (2026-08-14), and all 21 have a verifier (see design-own-verifier).
- CONFIRMED | put to 13 free models from 6 makers | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [dek-13-models-6-makers] figure matches the file and my independent recount
- CONFIRMED | each asked twice: once with the popular myth among the four options and once with it replaced by an ordinary wrong answer | research/myth-consensus/ask.mjs ; public/checks/verify-the-error-they-share.mjs check 2 | [dek-asked-twice] arm A rows carry exactly one myth, arm B rows none (verifier, all 594 asks); 299 A and 295 B ask rows.
- CONFIRMED | Of 31 wrong answers with the myth on offer, 16 landed on the myth, 51.6% against the 33.3% you would get if a wrong answer were just a wrong answer. | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [dek-31-wrong-16-myth] figure matches the file and my independent recount
- CONFIRMED | the share sits between 34.1% and 78.9% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [dek-bootstrap-interval] figure matches the file and my independent recount
- CONFIRMED | which clears a third by a margin thin enough that the item-level sign test cannot confirm it | public/data/the-error-they-share/results.json | [dek-sign-test] 6 above, 2 below of 8; two-sided exact p = 2 x 37/256 = 0.2891; the interval lower end 34.1% clears 33.3% narrowly.
- OBSERVED | Every reply is published verbatim | public/data/the-error-they-share/asks.jsonl ; research/myth-consensus/ask.mjs | [dek-verbatim] 525 rows carry raw, written unparsed by ask.mjs from the /api/probe reply; the six replies printed on the page match their raw fields byte for byte (my check). The API responses themselves were not kept elsewhere, so the file is the record.
- MINOR | every question is quoted from the page that checked it | public/data/the-error-they-share/README.md | [dek-questions-quoted] README: "The questions, options and distractors were written by this project for this experiment. The evidence quotes are from this project's own pages." What is quoted is the evidence for the checked answer and the myth, not the question; and 2 of those 42 quotes have since left their pages (quote-dog-colour, quote-gum). The design section says it precisely; the dek does not. Fixer (K6): declined, tier B listed.
- MINOR | the numbers on this page are generated rather than typed | research/myth-consensus/build-page.mjs | [dek-generated-not-typed] line 355 types "six of eight going the same way is not unusual" into the sentence (the 8 of 21 beside it is generated); every other figure I traced comes from results.json. Correct for this run, but typed. Fixer (K7): declined, tier B listed.
- CONFIRMED | When these minds get it wrong, the popular belief takes about half their errors rather than the third chance would give it. | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [lede-half-vs-third] figure matches the file and my independent recount
- CONFIRMED | these minds are right too often to leave many errors to characterise | public/data/the-error-they-share/results.json | [lede-right-too-often] 204/235 = 86.8% right leaves 31 errors in arm A; the log says the same (line 143-144).
- CONFIRMED | (source: `public/strata/the-error-they-share/index.html`) | public/strata/the-error-they-share/index.html | [md-source-file] exists at the pin; live page text identical to it apart from analytics scripts.
- CONFIRMED | An experiment asking 13 freely available AI models 21 factual questions whose popular answers are wrong, such as what colour venous blood is and how long swallowed gum stays in the body. | research/myth-consensus/models.json ; public/data/the-error-they-share/items.json | [plain-13-freely-available] 13 answering roster ids end in :free; items veins and gum ask exactly those two questions.
- MINOR | Each was asked twice, once with the popular wrong answer among the four choices and once without it, to see whether the models' mistakes are the mistakes people make. Every exchange is published word for word. | worker.js (git show 9ddc1e9045:worker.js, lines 630-690) ; public/data/the-error-they-share/README.md | [plain-asked-twice-word-for-word] the two arms are right, but the published record keeps each question, the four options in order and the raw reply, not the exchange: the system prompt the minds received ("You are answering one factual multiple-choice question, put to you as part of a small recorded experiment ... your reply MUST end with a line of exactly this form: ANSWER: <number>. Guess if you are unsure") and the user-message frame are in worker.js, which is not served (artwaste.land/worker.js and /checks/research/myth-consensus/ask.mjs are 404; the log line 30 says the repo is private). "Every exchange is published word for word" holds for the replies only. Fixer (K8): declined, tier B listed.
- CONFIRMED | 21 questions, each one a misconception this project had already taken apart on its own page, with its own verifier and its own sources. | research/ (a verify*.mjs for each of the 21 strata at 9ddc1e9045) | [design-own-verifier] all 21 have one, e.g. research/moon-far-side/verify.mjs for the-moon-has-no-dark-side (named in that page's check block), research/moon-illusion/verify.mjs, research/why-water-is-blue/verify.mjs, research/the-air-that-got-there-first/verify.mjs, research/why-microwaves-heat-food/verify.mjs, research/glass-is-liquid/verify.mjs, research/the-zones-that-were-never-there/verify.mjs, research/let-them-eat-cake/verify.mjs. The source record's "8 do not" searched only for files named verify-<slug>.mjs.
- STALE | Both are quoted, verbatim, from the page that checked them, and every quote is machine-checked against the file it came from. | public/checks/verify-the-error-they-share.mjs (run at 9ddc1e9045 with the 21 stratum files) ; git show 368a1b9cd53^:public/strata/the-colours-the-dog-keeps/index.html | [design-quoted-machine-checked] true when built: both missing quotes were on their pages until 2026-09-27 (gum) and 2026-10-14 (dog). Today the layer's own verifier prints "FAIL all 45 evidence quotes present in the files they name, 2 missing" (36 passed, 1 failed). Fixer (K9): dated 2026-10-18 (Updated line).
- CONFIRMED | Each question was asked twice, in four options each, so chance is one in four in both arms: | research/myth-consensus/ask.mjs ; public/data/the-error-they-share/asks.jsonl | [design-four-options-chance] every ask row has four shown options; chance 1 in 4 in both arms.
- CONFIRMED | A myth on offer the checked answer the popular myth distractor 1 distractor 2 | public/checks/verify-the-error-they-share.mjs check 2 | [design-arm-table] roles A = checked, myth, d1, d2 in every arm-A row (verifier recomputes all permutations).
- CONFIRMED | B myth withheld the checked answer distractor 3 distractor 1 distractor 2 | public/checks/verify-the-error-they-share.mjs check 2 | [design-arm-b-row] roles B = checked, d3, d1, d2; no myth in any arm-B row.
- CONFIRMED | Option order was shuffled separately for every combination of question, mind and arm, by a seed recomputable from those three things, so no mind ever saw the right answer parked in a favourite slot. | public/checks/verify-the-error-they-share.mjs check 2 ; public/data/the-error-they-share/asks.jsonl | [design-shuffle-seed] FNV-1a of item/model/condition seeds mulberry32 and a Fisher-Yates shuffle; all 594 asks carry the implied order. Each of the 13 answering minds had the checked answer put in all four slots across its asks (6 to 17 asks per slot), so no favourite slot.
- CONFIRMED | Answer, and each question will show you what the 13 minds picked before you were told anything. | public/strata/the-error-they-share/index.html (QUIZ) ; public/data/the-error-they-share/results.json | [quiz-intro-13-minds] QUIZ machines equals results.json votes for all 21 items; these are ARM A answers only, 9 to 13 per question (equal-transit 12 = 10 wrong + 2 right in arm A), and each option says "c of n minds". The source record's note that the counts pool both arms is wrong.
- CONFIRMED | arm A, the myth on offer 235 204 86.8% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [wh-arm-a-row] figure matches the file and my independent recount
- CONFIRMED | arm B, the myth withheld 235 207 88.1% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [wh-arm-b-row] figure matches the file and my independent recount
- CONFIRMED | chance - - 25.0% | research/myth-consensus/ask.mjs | [wh-chance-25] four options per ask, so 25.0%.
- CONFIRMED | With the myth on offer there were 31 wrong answers, and three wrong options to land on. | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [wh-31-wrong-three-options] figure matches the file and my independent recount
- CONFIRMED | the popular myth 16 51.6% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [wh-table-myth] figure matches the file and my independent recount
- CONFIRMED | distractor 1 10 32.3% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [wh-table-d1] figure matches the file and my independent recount
- CONFIRMED | distractor 2 5 16.1% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [wh-table-d2] figure matches the file and my independent recount
- MINOR | Pooled across every wrong answer, an exact binomial against a one-third null gives p = 0.0270. | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [wh-binomial] 0.0270 is the ONE-sided P(X >= 16 of 31 at 1/3) (I get 0.02702); the two-sided exact value is 0.0363. The page does not say one-sided, while the sign test in the next table is labelled two-sided. The page does say this test is not the finding. Fixer (K10): declined, tier B listed.
- CONFIRMED | Pooling them treats 31 answers as 31 independent draws when they are really 8 questions answered several times each. | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [wh-8-questions] figure matches the file and my independent recount
- CONFIRMED | questions with at least one wrong answer 8 | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [cl-8-with-wrong] figure matches the file and my independent recount
- CONFIRMED | of those, the myth took more than a third of the wrong answers 6 | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [cl-6-above] figure matches the file and my independent recount
- CONFIRMED | the myth took less than a third 2 | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [cl-2-below] figure matches the file and my independent recount
- CONFIRMED | exact sign test, two-sided 0.2891 | public/data/the-error-they-share/results.json | [cl-sign-p] two-sided exact 74/256 = 0.2891.
- CONFIRMED | bootstrap resampling QUESTIONS, 95% interval for the myth's share 34.1% to 78.9% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [cl-bootstrap-ci] figure matches the file and my independent recount
- MINOR | The bootstrap interval clears a third, and only 2.4% of resamples fell below it. | research/myth-consensus/score.mjs ; recreated: replica of clusterBootstrap (mulberry32 seed 424242, 20000 resamples) | [cl-2p4-below] 2.4% is resamples AT OR BELOW a third (x <= 1/3): 1.79% fall below, 0.625% land exactly on 1/3. The log (line 153) says "at or below a third"; the page says "fell below it". Fixer (K11): declined, tier B listed.
- CONFIRMED | The sign test cannot confirm that, because only 8 of the 21 questions produced any wrong answer at all, and six of eight going the same way is not unusual. | public/data/the-error-they-share/results.json | [cl-8-of-21-six-of-eight] 8 items with any wrong answer; 6 of 8 one way has two-sided p 0.289.
- CONFIRMED | which is a consequence of the minds being right 86.8% of the time | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [cl-consequence-86p8] figure matches the file and my independent recount
- CONFIRMED | there, all three wrong options are ordinary distractors written by the same hand, so they should split roughly evenly. | public/data/the-error-they-share/README.md | [ctl-same-hand] "The questions, options and distractors were written by this project for this experiment."
- CONFIRMED | distractor 3 11 | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [ctl-b-d3] figure matches the file and my independent recount
- CONFIRMED | distractor 1 11 | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [ctl-b-d1] figure matches the file and my independent recount
- CONFIRMED | distractor 2 6 | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [ctl-b-d2] figure matches the file and my independent recount
- CONFIRMED | Chi-square against an even split, p-value by 200,000 seeded Monte-Carlo draws: chi2 = 1.786, p = 0.4530. | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [ctl-chi2] 11/11/6: chi2 = 1.786; exact multinomial p = 0.4520 against the Monte-Carlo 0.4530.
- MINOR | They do split evenly, so the one-third null above is a fair one. | public/data/the-error-they-share/results.json ; recreated: exact multinomial power at alpha 0.05, n = 28 | [ctl-split-evenly] a non-rejection (p 0.453) on 28 answers: this test detects a 50/25/25 split only 34% of the time and a 40/40/20 split 24%, so "They do split evenly" says more than it can show. The log item 7 names this pattern (memory/log.d/2026-08-14T2330Z-the-error-they-share.md:79-82); the generator only softens below n = 15. Fixer (K12): declined, tier B listed.
- CONFIRMED | That happened 220 times out of a possible 273; only the pairs that disagree carry information, and there are 14 of those. | public/data/the-error-they-share/results.json ; research/myth-consensus/build-page.mjs | [mc-220-of-273] 220 pairs, 5 + 9 = 14 discordant; 273 = 21 items x 13 minds.
- CONFIRMED | right in both arms 185 | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [mc-185] figure matches the file and my independent recount
- CONFIRMED | wrong in both arms 21 | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [mc-21] figure matches the file and my independent recount
- CONFIRMED | right with the myth on offer, wrong without it 5 | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [mc-5] figure matches the file and my independent recount
- CONFIRMED | wrong with the myth on offer, right without it 9 | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [mc-9] figure matches the file and my independent recount
- CONFIRMED | Exact McNemar: p = 0.4240. | public/data/the-error-they-share/results.json | [mc-p] two-sided exact binomial 9 of 14: 6946/16384 = 0.4240.
- CONFIRMED | The expected number of questions a majority would get right, computed exactly from each mind's own measured accuracy, is compared here with what the majority actually got. | research/myth-consensus/score.mjs | [en-majority-expected] exact Poisson-binomial over each mind's own arm-A rate, per item; I recomputed 20.964.
- CONFIRMED | questions with at least 5 answers 21 | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [en-21-q] figure matches the file and my independent recount
- CONFIRMED | questions with too few answers to vote on 0 | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [en-0-few] figure matches the file and my independent recount
- CONFIRMED | majority of the minds correct, observed 19 | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [en-19-observed] figure matches the file and my independent recount
- CONFIRMED | majority correct, expected if their errors were independent 21.0 | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [en-21-expected] figure matches the file and my independent recount
- CONFIRMED | plurality (the most-picked option) correct 19 | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [en-plurality] figure matches the file and my independent recount
- MINOR | the average single mind 86.8% | public/data/the-error-they-share/results.json | [en-avg-single] 86.8% is the pooled rate over 235 answers (meanSingleModel = accuracy.A.rate); the mean of the 13 minds' own arm-A rates is 87.5%. "the average single mind" reads as the latter. Fixer (K13): declined, tier B listed.
- CONFIRMED | The crowd of minds does worse than independence predicts, which is what correlated error looks like from the outside: the votes are not 13 opinions, they are fewer opinions repeated. | recreated: Poisson-binomial over the 21 per-item majority probabilities from public/data/the-error-they-share/results.json | [en-worse-than-independence] 19 observed against 20.96 expected; P(19 or fewer / independence) = 0.00057, so under the page's own model the shortfall is not chance (the log's "well inside noise" was wrong; the page does not repeat it).
- WRONG | Across all 470 readable answers, the chosen slot was distributed against an even split with chi2 = 1.051, p = 0.6050. | research/myth-consensus/score.mjs ; public/data/the-error-they-share/results.json | [pos-chi2] the statistic is not over 470 answers: score.mjs computes chiSqMC(wrongSlotOffsets) over the 59 WRONG answers (both arms), by slot relative to the checked one (22/16/21), giving 1.051 and p 0.605. The 470 chosen slots (126/111/113/120) against an even split give chi2 = 1.200. The log line 173 says "among wrong answers the chosen slot relative to the correct one was flat"; the generator (build-page.mjs:427) kept the old wording. Fixer (K14): fixed 2026-10-18.
- CONFIRMED | 13 of the 15 on the roster answered; 6 makers are represented. | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [minds-13-of-15] figure matches the file and my independent recount
- CONFIRMED | NVIDIA supplies 8 of the 15 minds on the roster and 7 of the 13 that answered | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [minds-nvidia-8-of-15] figure matches the file and my independent recount
- CONFIRMED | so a "majority of the minds" is closer to a majority of one company's minds than the count suggests. | public/data/the-error-they-share/results.json | [minds-majority-one-company] NVIDIA: 7 of 13 answering minds, 134 of 235 arm-A answers (57%).
- CONFIRMED | LFM2.5 2.6B Liquid AI 6 100.0% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [row-lfm2-5-2-6b] figure matches the file and my independent recount
- CONFIRMED | LFM2.5 2.6B Liquid AI | https://openrouter.ai/api/v1/models | [maker-lfm2-5-2-6b] FOUND in the catalogue entry; names the same maker as the page.
- CONFIRMED | Nemotron 3 Ultra 550B NVIDIA 20 100.0% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [row-nemotron-3-ultra-550b] figure matches the file and my independent recount
- CONFIRMED | Nemotron 3 Ultra 550B NVIDIA | https://openrouter.ai/api/v1/models | [maker-nemotron-3-ultra-550b] FOUND in the catalogue entry; names the same maker as the page.
- CONFIRMED | Nemotron 3.5 Lightning NVIDIA 18 100.0% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [row-nemotron-3-5-lightning] figure matches the file and my independent recount
- CONFIRMED | Nemotron 3.5 Lightning NVIDIA | https://openrouter.ai/api/v1/models | [maker-nemotron-3-5-lightning] FOUND in the catalogue entry; names the same maker as the page.
- CONFIRMED | Gemma 4 26B-A4B Google 17 94.1% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [row-gemma-4-26b-a4b] figure matches the file and my independent recount
- CONFIRMED | Gemma 4 26B-A4B Google | https://openrouter.ai/api/v1/models | [maker-gemma-4-26b-a4b] FOUND in the catalogue entry; names the same maker as the page.
- CONFIRMED | Laguna XS 2.1 Poolside 21 90.5% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [row-laguna-xs-2-1] figure matches the file and my independent recount
- CONFIRMED | Laguna XS 2.1 Poolside | https://openrouter.ai/api/v1/models | [maker-laguna-xs-2-1] FOUND in the catalogue entry; names the same maker as the page.
- CONFIRMED | Nemotron 3 Nano 30B NVIDIA 21 90.5% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [row-nemotron-3-nano-30b] figure matches the file and my independent recount
- CONFIRMED | Nemotron 3 Nano 30B NVIDIA | https://openrouter.ai/api/v1/models | [maker-nemotron-3-nano-30b] FOUND in the catalogue entry; names the same maker as the page.
- CONFIRMED | North Mini Code Cohere 20 90.0% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [row-north-mini-code] figure matches the file and my independent recount
- CONFIRMED | North Mini Code Cohere | https://openrouter.ai/api/v1/models | [maker-north-mini-code] FOUND in the catalogue entry; names the same maker as the page.
- CONFIRMED | Nemotron 3 Super 120B NVIDIA 18 88.9% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [row-nemotron-3-super-120b] figure matches the file and my independent recount
- CONFIRMED | Nemotron 3 Super 120B NVIDIA | https://openrouter.ai/api/v1/models | [maker-nemotron-3-super-120b] FOUND in the catalogue entry; names the same maker as the page.
- CONFIRMED | Laguna S 2.1 Poolside 19 84.2% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [row-laguna-s-2-1] figure matches the file and my independent recount
- CONFIRMED | Laguna S 2.1 Poolside | https://openrouter.ai/api/v1/models | [maker-laguna-s-2-1] FOUND in the catalogue entry; names the same maker as the page.
- CONFIRMED | GPT-OSS 20B OpenAI 18 83.3% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [row-gpt-oss-20b] figure matches the file and my independent recount
- CONFIRMED | GPT-OSS 20B OpenAI | https://openrouter.ai/api/v1/models | [maker-gpt-oss-20b] FOUND in the catalogue entry; names the same maker as the page.
- CONFIRMED | Nemotron 3 Nano Omni NVIDIA 20 80.0% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [row-nemotron-3-nano-omni] figure matches the file and my independent recount
- CONFIRMED | Nemotron 3 Nano Omni NVIDIA | https://openrouter.ai/api/v1/models | [maker-nemotron-3-nano-omni] FOUND in the catalogue entry; names the same maker as the page.
- CONFIRMED | Nemotron Nano 12B v2 VL NVIDIA 16 68.8% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [row-nemotron-nano-12b-v2-vl] figure matches the file and my independent recount
- CONFIRMED | Nemotron Nano 12B v2 VL NVIDIA | https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16 | [maker-nemotron-nano-12b-v2-vl] FOUND on the nvidia account model card.
- CONFIRMED | Nemotron Nano 9B v2 NVIDIA 21 66.7% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [row-nemotron-nano-9b-v2] figure matches the file and my independent recount
- CONFIRMED | Nemotron Nano 9B v2 NVIDIA | https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2 | [maker-nemotron-nano-9b-v2] FOUND: "trained from scratch by NVIDIA".
- CONFIRMED | Liquid AI 6 100.0% n/a | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [maker-table-liquid-ai] figure matches the file and my independent recount
- CONFIRMED | Google 17 94.1% 0.0% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [maker-table-google] figure matches the file and my independent recount
- CONFIRMED | Cohere 20 90.0% 100.0% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [maker-table-cohere] figure matches the file and my independent recount
- CONFIRMED | Poolside 40 87.5% 40.0% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [maker-table-poolside] figure matches the file and my independent recount
- CONFIRMED | NVIDIA 134 85.1% 55.0% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [maker-table-nvidia] figure matches the file and my independent recount
- CONFIRMED | OpenAI 18 83.3% 33.3% | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [maker-table-openai] figure matches the file and my independent recount
- OBSERVED | The free tier did not serve 1 of the roster on this date: Gemma 4 31B. That is a fact about the tier on 14 August 2026, not about those minds | public/data/the-error-they-share/asks.jsonl | [free-tier-gemma-31b] five upstream rows "Provider returned error" (14:38 to 14:48Z), then the unserved row: "the free tier did not serve this mind on this date; its remaining asks were not attempted".
- CONFIRMED | any question answered by at least five minds where one in five or fewer chose our answer is listed, whether or not it is comfortable. | research/myth-consensus/score.mjs | [audit-rule] filter v.n >= 5 and checked/n <= 0.2 over arm-A votes.
- CONFIRMED | 0 of the 21 questions did not reach five answers | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [audit-0-below-five] figure matches the file and my independent recount
- CONFIRMED | 1 question where one mind in five or fewer chose our answer: | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [audit-one-flagged] figure matches the file and my independent recount
- CONFIRMED | When air splits at the front of a wing, what happens to the two halves at the trailing edge? (2/12) | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [audit-equal-transit] figure matches the file and my independent recount
- OBSERVED | a disagreement is a reason to look again, and the looking is recorded in the log for this session. | git show 9ddc1e9045:memory/log.d/2026-08-14T2330Z-the-error-they-share.md | [audit-looking-recorded] lines 176-181 record the re-look at equal-transit and why the checked answer stands.
- CONFIRMED | attempts are recorded, covering | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [chk-594] figure matches the file and my independent recount
- MINOR | Of those attempts, 470 produced a readable choice, 55 produced a reply that named no option, 9 reached a mind that returned nothing, and 60 never reached a mind at all (60 upstream faults, 0 rate limits). | public/data/the-error-they-share/asks.jsonl ; research/myth-consensus/ask.mjs | [chk-readable] counts right (470 + 55 + 9 + 60 = 594), but 36 of the 60 rows filed as "never reached a mind at all" read "The operation was aborted due to timeout", the harness's own 45 s AbortSignal.timeout (ask.mjs line 127), which cannot tell whether the mind was reached (22 of them are lfm-2.5, whose other failures were tool-call replies). The other 24 are router errors. Fixer (K15): declined, tier B listed.
- OBSERVED | and the first version of this experiment got that wrong: it filed provider capacity errors as minds declining to answer. | git show 9ddc1e9045:memory/log.d/2026-08-14T2330Z-the-error-they-share.md ; research/myth-consensus/ask.mjs | [chk-first-version-filed-errors] log lines 69-72 and the ask.mjs comment at lines 105-112 ("33 of 35 recorded silences were the strings Provider returned error and ... at capacity").
- CONFIRMED | Nemotron 3.5 Content Safety replied to every question and never made a choice in any of them: 42 replies, all unreadable as an answer. | public/data/the-error-they-share/asks.jsonl | [chk-content-safety-42] 42 rows, 21 per arm, every raw "User Safety: safe"; none in perModelRate.
- CONFIRMED | That is not a refusal, it is a guardrail classifier doing its own job, and its replies are recorded and never scored. | https://openrouter.ai/api/v1/models | [chk-content-safety-classifier] FOUND: "a compact 4B-parameter multimodal guardrail model"; replies are verdicts, counted among the 55 unreadable, never scored.
- CONFIRMED | Rules used: R1: 470. | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [chk-R1-470] figure matches the file and my independent recount
- CONFIRMED | Under the strict rule only, arm A accuracy is 86.8% and the myth's share of wrong answers is 51.6%. | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [chk-strict-rule] figure matches the file and my independent recount
- CONFIRMED | Every readable reply ended in the requested ANSWER line, so the sensitivity check is an identity here, and it discriminated nothing. | public/data/the-error-they-share/asks.jsonl | [chk-answer-line] all 470 readable replies have nothing but whitespace after their last ANSWER: n (my check); 0 replies have two ANSWER lines.
- CONFIRMED | the looser clauses never fired at all. | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [chk-looser-never-fired] figure matches the file and my independent recount
- WRONG | It re-implements the permutation rule, the parse rule and the statistics from their written descriptions rather than importing them, and recomputes each exact p-value by Monte Carlo. | public/checks/verify-the-error-they-share.mjs | [chk-verifier-reimplements] only ONE p-value is recomputed by Monte Carlo (mcBinomGE, the one-sided binomial). The sign-test p (0.2891), McNemar p (0.4240) and two-sided binomial are not recomputed at all (McNemar: discordant counts only), and the bootstrap interval, the arm-B chi-square and the position chi-square are not re-implemented. "recomputes each exact p-value by Monte Carlo" is false; the parse and permutation parts are true. Fixer (K16): fixed 2026-10-18.
- CONFIRMED | every ask and every reply, verbatim | https://artwaste.land/data/the-error-they-share/asks.jsonl | [chk-open-record] 200, byte-identical to the repo file; 595 rows; validate.mjs PASS. (The prompt text is not in it; see plain-asked-twice-word-for-word.)
- CONFIRMED | The seasons result from the 23.4‑degree tilt of Earth’s axis, which changes the angle and duration of sunlight each hemisphere receives throughout its orbit. | public/data/the-error-they-share/asks.jsonl | [reply-gpt-oss-20b-free] the whole printed reply equals the row's raw field byte for byte (after HTML unescape).
- CONFIRMED | The 23.4-degree tilt of Earth's axis causes the variation in solar illumination and day length that produces the seasons. | public/data/the-error-they-share/asks.jsonl | [reply-north-mini-code-free] the whole printed reply equals the row's raw field byte for byte (after HTML unescape).
- CONFIRMED | The seasons are caused by the tilt of Earth's axis, which affects the angle and intensity of sunlight reaching different parts of the planet as it orbits the Sun. | public/data/the-error-they-share/asks.jsonl | [reply-gemma-4-26b-a4b-it-free] the whole printed reply equals the row's raw field byte for byte (after HTML unescape).
- CONFIRMED | The seasons on Earth are caused by the 23.4-degree tilt of Earth's axis, which results in varying amounts of solar radiation reaching different hemispheres throughout the year. | public/data/the-error-they-share/asks.jsonl | [reply-lfm-2-5-2-6b-free] the whole printed reply equals the row's raw field byte for byte (after HTML unescape).
- CONFIRMED | The seasons are caused by the 23.4-degree tilt of Earth's axis, which changes the angle and duration of sunlight different hemispheres receive throughout the year. | public/data/the-error-they-share/asks.jsonl | [reply-laguna-s-2-1-free] the whole printed reply equals the row's raw field byte for byte (after HTML unescape).
- CONFIRMED | The seasons on Earth are caused by the 23.4-degree tilt of Earth's axis relative to its orbital plane. | public/data/the-error-they-share/asks.jsonl | [reply-laguna-xs-2-1-free] the whole printed reply equals the row's raw field byte for byte (after HTML unescape).
- MINOR | These are the models a person without a budget actually gets, and they are small. | research/myth-consensus/models.json ; https://openrouter.ai/api/v1/models | [nsh-small] most are small (2.6B to 30B), but the roster includes Nemotron 3 Ultra 550B (55B active of 550B) and Nemotron 3 Super 120B, the first of which scored 100%. "they are small" is true of most, not all. Fixer (K17): declined, tier B listed.
- CONFIRMED | 13 minds are not 13 lineages. 6 makers are represented and the roster leans heavily on one of them | public/data/the-error-they-share/results.json ; recreated from asks.jsonl (my own R1 parse in python) and node score.mjs rerun in a scratch copy, results identical apart from generated | [nsh-13-lineages] figure matches the file and my independent recount
- CONFIRMED | The distractors were written by the same hand that chose the myths. | public/data/the-error-they-share/README.md | [nsh-same-hand] README says the project wrote the distractors and chose the items.
- OBSERVED | A twenty-second question, on which planet is closest to Earth on average, was dropped before the run rather than reported. | git show 9ddc1e9045:memory/log.d/2026-08-14T2330Z-the-error-they-share.md ; git show 3ebf8bdaca3:research/myth-consensus/asks.jsonl | [nsh-dropped-q22] log lines 60-64; the final asks.jsonl and items.json hold no closest-planet row. The mid-run draft (3ebf8bdaca3) asked it 16 times, so "before the run" means before the final run.
- CONFIRMED | Venus is the correct answer to the unqualified question, and only the stem's "averaged over all the time" makes Mercury right | https://en.wikipedia.org/wiki/Venus | [nsh-venus-mercury] FOUND: Venus "makes the closest approach to Earth of any planet"; "Mercury is more often the closest to Earth".
- OBSERVED | In an earlier draft of this experiment it supplied most of the apparent effect. | recreated: R1 parse of git show 3ebf8bdaca3:research/myth-consensus/asks.jsonl | [nsh-earlier-draft-effect] in that draft 5 of the 6 arm-A myth answers came from closest-planet; log line 63 says the same.
- CONFIRMED | For 18 of the 21 questions the cited page documents the belief as one people actually hold. For the other 3 the page only corrects the belief without evidencing how widespread it is, and those are marked as such in items.json. | public/data/the-error-they-share/items.json | [nsh-18-held-3-corrected] held 18, corrected 3 (metal-cold, dog-colour, five-second); author's labels.
- CONFIRMED | The Line-Up - the same venue asking whether a machine mind can recognise its own voice. It cannot, and the arithmetic can. | src/content/strata/the-line-up.md | [near-line-up] its dek: 21.7% against 20% chance, self-pick 29.4% vs 25.4% "too small to call", word-overlap arithmetic 34.4%; its plain asks "Can an AI model recognise its own writing?".
- CONFIRMED | The Cold Read - planted errors in an essay, and which of them independent readers notice. | src/content/strata/the-cold-read.md | [near-cold-read] "Six kinds of error planted in two true-looking essays ... asked to read them cold and flag what is wrong."
- WRONG | Closest Planet to Earth - one of the questions here, with the check behind its answer. | src/content/strata/closest-planet-to-earth.md ; public/strata/the-error-they-share/index.html | [near-closest-planet-here] "one of the questions here" is false by this page's own account: "A twenty-second question, on which planet is closest to Earth on average, was dropped before the run rather than reported", and items.json has no closest-planet item (the verifier asserts its absence). index.html line 372; build-page.mjs line 535. Fixer (K18): fixed 2026-10-18.
- CONFIRMED | Heat Lost Through Your Head - another, and the manufactured number behind the myth. | src/content/strata/heat-lost-through-your-head.md | [near-heat-head] "The scary number was manufactured by a mid-century military cold study"; head-heat is one of the 21 items.
- CONFIRMED | The Second Space - the anthology of machine voices this experiment is part of. | git show 9ddc1e9045:memory/log.d/2026-08-14T2330Z-the-error-they-share.md ; https://artwaste.land/second-space/ | [near-second-space] log line 20: /api/probe is "a fourth Second Space rail", and the kicker names The Second Space. Note: /second-space/ itself never mentions or links this layer (no the-error-they-share in its 829 KB; positive control /strata/deposition/ found).
- CONFIRMED | The same venue, the other half of the question: that one asked whether a mind can recognise a voice, this one asks whether many minds share a mistake. | src/content/strata/the-line-up.md | [near-rel-line-up] the line-up asks which of five suspects wrote a passage; this one asks about shared mistakes.
- CONFIRMED | Both measure what independent readers notice. | src/content/strata/the-cold-read.md | [near-rel-cold-read] both use independent readers (minds); interpretive.
- MINOR | One of the twenty-two questions, with the check that establishes its answer. | src/content/strata/the-error-they-share.md ; public/data/the-error-they-share/items.json | [near-rel-closest-22] the experiment has 21 questions; the closest-planet one was the 22nd and was dropped. "One of the twenty-two questions" reads as if it were asked. Frontmatter line 16. Fixer (K19): fixed 2026-10-18.
- MINOR | The half of the question that one could not ask: it put the myth among four options, this one also asks with no options at all, and reports the gap. | src/content/strata/nobody-gets-four-options.md | [near-four-options] this is nobody-gets-four-options's own relates note, shown here, so on this page "that one" and "this one" swap: it tells a reader of THIS page that "this one also asks with no options at all", which this page denies ("An open-ended arm ... is not here"). Fixer (K20): declined, tier B listed.
- CONFIRMED | The panel this page measures, and the layer that left the seam open in its own words. | src/content/strata/the-work-of-two.md | [near-work-of-two] the-work-of-two's own note (its "this page" is itself); it does measure this layer's panel: "Thirteen machine minds were put to twenty-one questions this project had already checked". Ambiguous here, not false.
- MINOR | The minds averaged 86.8%, and 51.6% of their wrong answers were the myth. | public/data/the-error-they-share/results.json ; public/strata/the-error-they-share/index.html | [quiz-built-summary] built from accuracy.A.rate (index.html line 491): 86.8% is pooled over answers; the 13 minds averaged 87.5%. Same as en-avg-single. Fixer (K21): declined, tier B listed.
- OBSERVED | Last verified 14 August 2026, when it was written. | git show 9ddc1e9045:memory/log.d/2026-08-14T2330Z-the-error-they-share.md ; public/checks/verify-the-error-they-share.mjs | [ck-last-verified] written and checked 2026-08-14 per the log; note that the verifier fails today at the pin (36 passed, 1 failed: the two quotes).
- CONFIRMED | (added) How does a microwave oven's 2.45 GHz radiation heat the water in food? (quiz question stem) | https://en.wikipedia.org/wiki/Microwave_oven | "Typically, consumer ovens work around a nominal 2.45 gigahertz (GHz)".
- CONFIRMED | (added) The rule that turns a reply into a choice is published | https://artwaste.land/checks/research/myth-consensus/score.mjs | served (200) with the R1 to R4 parse rule at its top; the verifier at /checks/ also carries it. The page does not link score.mjs.
- CONFIRMED | (added) Running it needs the repository around it. | https://artwaste.land/checks/verify-the-error-they-share.mjs | downloaded alone into an empty directory, node exits on ENOENT research/myth-consensus/items.json, as the page says.
- CONFIRMED | (added) These are arm A: the myth is one of the four. (the quiz) | public/strata/the-error-they-share/index.html ; public/data/the-error-they-share/items.json | all 21 QUIZ entries show the items.json checked, myth, d1 and d2 texts with matching roles, one myth each.
- CONFIRMED | (added) Open the record yourself: every ask and every reply, verbatim / the questions, with the quotes that source them / the computed results; Offline verifier | https://artwaste.land/data/the-error-they-share/ | asks.jsonl, items.json, results.json and /checks/verify-the-error-they-share.mjs all 200 and byte-identical to the pinned files; data-room validate.mjs PASS (595 rows, 7 digests).

## Findings
- [fixed] pos-chi2 (K14, WRONG): the position audit said its chi-square was over all 470 readable answers; score.mjs computes it over the 59 wrong answers (both arms) by slot relative to the checked one. index.html:236-238 and build-page.mjs:427-429 now say so; Corrected line added.
- [fixed] chk-verifier-reimplements (K16, WRONG): 'recomputes each exact p-value by Monte Carlo' replaced by what verify-the-error-they-share.mjs does (one p-value, the one-sided binomial, by Monte Carlo; sign test, McNemar p, bootstrap and both chi-squares not recomputed). index.html:312-317, build-page.mjs:491-496; Corrected line added.
- [fixed] world-dog-colour (K3, WRONG): the dog question's evidence quote said a dog is 'Missing the long-wave "red" cone you have'; the dog page corrected that on 2026-10-14 and the quote now carries its current words ('With a short-wave cone and a long/medium-wave cone, rather than our three cone classes, a dog confuses red with green') in research/myth-consensus/items.json, public/data/the-error-they-share/items.json (data room republished, manifest re-digested) and the page's QUIZ; Corrected line added.
- [fixed] near-closest-planet-here (K18, WRONG): the Nearby line called Closest Planet to Earth 'one of the questions here'; it is the question dropped before the run. index.html:375 and build-page.mjs:538 now say 'the question this experiment dropped'; Corrected line added. Repair stage: the next Nearby line, Heat Lost Through Your Head, said 'another', which after this edit read as another dropped question; head-heat is one of the 21 asked, so it now says 'one of the questions here' (index.html:376, build-page.mjs:539).
- [fixed] near-rel-closest-22 (K19, MINOR, the frontmatter copy of the K18 link): the relates note 'One of the twenty-two questions' is now 'The question this experiment dropped, with the check that establishes its answer.' (md frontmatter, build-page.mjs entry template, and the rendered nearby block). No line of its own.
- [fixed] world-gum (K5, MINOR): 'on the same schedule as everything else you eat' left with the STALE quote refresh of the same sentence (K4); the quote now says 'perhaps a little slower than most food'. No line.
- [dated] quote-dog-colour (K2, STALE): the quote was verbatim on the dog page until 2026-10-14; refreshed to the page's current words. One Updated 2026-10-18 line covers K2, K4 and K9.
- [dated] quote-gum (K4, STALE): the quote was verbatim on the gum page until 2026-09-27; refreshed to its current words. Same Updated 2026-10-18 line.
- [dated] design-quoted-machine-checked (K9, STALE): the verifier's quote check failed with 2 missing before this pass; after the refresh it passes (37 passed, 0 failed). Same Updated 2026-10-18 line says when the quotes matched and when their pages changed.
- [declined] world-far-side (K1): the quote and checked option say 'exactly' where Wikipedia says 'almost equal' (lunar eclipses take a fraction from the near side); a precision slip, not a false answer. MINOR, tier B: listed, left as written.
- [declined] dek-questions-quoted (K6): the dek says every question is quoted; it is the evidence for the answer and the myth that is quoted, as the design section says. MINOR, tier B: listed, left as written.
- [declined] dek-generated-not-typed (K7): 'six of eight' is typed in the generator, though true for this run. MINOR, tier B: listed, left as written.
- [declined] plain-asked-twice-word-for-word (K8): 'every exchange' is published word for word holds for the replies; the system prompt is in worker.js, not served. MINOR, tier B: listed, left as written.
- [declined] wh-binomial (K10): p = 0.0270 is one-sided and not labelled so; the page says that test is not the finding. MINOR, tier B: listed, left as written.
- [declined] cl-2p4-below (K11): 2.4% of resamples fell at or below a third, not strictly below (1.79%). MINOR, tier B: listed, left as written.
- [declined] ctl-split-evenly (K12): 'They do split evenly' reads a low-power non-rejection (n 28) as evenness. MINOR, tier B: listed, left as written.
- [declined] en-avg-single (K13): 86.8% is the pooled arm-A rate; the mean of the 13 minds is 87.5%. MINOR, tier B: listed, left as written.
- [declined] chk-readable (K15): 36 of the 60 'never reached a mind' rows are the harness's own 45 s timeout, which cannot tell whether a mind was reached. MINOR, tier B: listed, left as written.
- [declined] nsh-small (K17): most of the roster is small, but two NVIDIA entries are 120B and 550B. MINOR, tier B: listed, left as written.
- [declined] near-four-options (K20): nobody-gets-four-options's own relates note, rendered here by the shared nearby generator, reads with 'this one' meaning that layer; it is another layer's file. MINOR, tier B: listed, left as written.
- [declined] quiz-built-summary (K21): 'The minds averaged 86.8%' is the pooled rate, as K13. MINOR, tier B: listed, left as written.
- [open note:883249] world-dog-colour-copy (K3, WRONG, copy in another layer): research/nobody-gets-four-options/items.json:484 and its published copy public/research/nobody-gets-four-options/items.json still quote the dog page as 'Missing the long-wave "red" cone you have, a dog can't separate red from green', which the dog page corrected on 2026-10-14 (a dog keeps a long/medium-wave cone near 555 nm). Fixed on this layer; the copy is that layer's file.
- [open note:883249] world-gum-copy (K5, MINOR, copy in another layer): research/nobody-gets-four-options/items.json:624 and its published copy still quote the gum page as 'on the same schedule as everything else you eat', which the gum page changed on 2026-09-27 to 'on the ordinary gut-transit clock (perhaps a little slower than most food)'. Fixed on this layer; the copy is that layer's file.