--- slug: nothing-here-is-a-bug 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: 102 confirmed: 69 wrong: 23 # 11 WRONG + 12 MINOR unverifiable: 2 observed: 8 stale: 0 --- This was a directed repair pass over the 23 triage items in research/assay-line/claude-1018b/nothing-here-is-a-bug/triage.json (11 WRONG, 12 MINOR, no STALE), read at db2da072f8, after a source record (N1, 92 entries) and one checker (N2, 102 claim lines, 10 of them added). Before editing, the fixer re-read the evidence for every WRONG item itself: the HTML standard's hardwareConcurrency clause (workers.html, fetched), ECMA-262 2025 6.1.6.1.20 and 21.3.2 and the 1997 first edition's 15.8.2 (pdftotext), UAX #15 section 3, ECMA-402 6.5 with a local Node (tzdata 2026a, CLDR 48) for Europe/Kyiv, CLDR en.xml at release-41, -42 and -43 with the CLDR 42 and 43 release notes, the battery's Cancún instant, and the verifier's code; it recomputed in python3 that six 17-digit strings round-trip to 0.1 + 0.2. All eleven WRONG items held and were fixed, with eight dated Corrected lines on the page and in the markdown. The largest fix moves String(0.1 + 0.2) out of the controls, which changes the page's counts from five of eighteen and four controls to five of nineteen and three controls; the captured answers are unchanged, and the verifier now asserts that the probe code is the code the runtimes ran. One MINOR item (capture.mjs "every runtime it can find") was fixed because it shares a sentence with a WRONG one; the other eleven MINOR items are declined under tier B. The claim lines below are N2's, with the lines whose page wording changed restated as corrected. After the fixes the layer's verifier printed 61/61. ## Claims - CONFIRMED | [exp-1997-first-edition] ECMA-262 has never specified what Math.exp returns and has not specified it since 1997 | https://ecma-international.org/wp-content/uploads/ECMA-262_1st_edition_june_1997.pdf | pdftotext of the 1997 PDF, 15.8.2.8 exp(x): "Returns an implementation-dependent approximation to the exponential function of the argument". The sentence is redundant ("never ... and not since 1997") but true. - CONFIRMED | [exp-current-edition] ECMA-262 has never specified what Math.exp returns and has not specified it since 1997 | https://tc39.es/ecma262/2025/ | 21.3.2.14 Math.exp: "Return an implementation-approximated Number value representing the exponential function of R(n)". Same in ES2024, ES2026. - WRONG | [html-any-number-prose] the HTML standard says a browser may report any number of processors it likes | https://html.spec.whatwg.org/multipage/workers.html | The getter "must return a number between 1 and the number of logical processors potentially available to the user agent"; lower values are allowed only for UA limits or "to limit fingerprinting possibilities". Not "any number it likes": it may not report more than are available. Same text in archived 2016 and 2018 copies. Fixed (triage K1): the page now says the standard lets a browser report fewer processors than it has, down to one, to limit fingerprinting. - WRONG | [html-any-number-relates] the HTML specification says the processor count may be any value the user agent likes | https://html.spec.whatwg.org/multipage/workers.html | Same as above: the value is bounded above by the logical processors potentially available; "any value the user agent likes" is not what the clause says. Fixed (triage K2): the relates note now says "any processor count from one up to the number it has"; the nearby block copy updated to match. - WRONG | [html-hwconc-quote] HTML: the attribute "may return a lower value than the actual number of logical processors" and the user agent may return any value it likes to resist fingerprinting. | https://html.spec.whatwg.org/multipage/workers.html | The quoted words "may return a lower value than the actual number of logical processors" are not in the current HTML standard nor in web.archive.org copies of workers.html from 2016-06-02 and 2018-05-31 (positive control: the same copies contain "hardwareConcurrency attribute's getter must return a number between 1 and"). The clause permits lower values, not "any value it likes". Fixed (triage K3): the battery clause now quotes the standard's own words. - CONFIRMED | [date-parse-quote] ECMA-262 21.4.3.2: if the String does not conform to the Date Time String Format, "the function may fall back to any implementation-specific heuristics or implementation-specific date formats". | https://262.ecma-international.org/15.0/ | ES2024 21.4.3.2 Date.parse: "If the String does not conform to that format the function may fall back to any implementation-specific heuristics or implementation-specific date formats." Quote exact, clause number right. - CONFIRMED | [tanh-implementation-approximated] ECMA-262 21.3.2: the listed functions are "implementation-approximated"; an implementation is only "encouraged" to use the approximation that is generally available. | https://tc39.es/ecma262/2025/ | Each listed function "Return[s] an implementation-approximated Number value". Record error: Math.tanh is 21.3.2.35 in ES2025, not 21.3.2.37. - MINOR | [tanh-encouraged] ECMA-262 21.3.2: the listed functions are "implementation-approximated"; an implementation is only "encouraged" to use the approximation that is generally available. | https://tc39.es/ecma262/2025/ | "encouraged" is in 4.2 (implementations "are encouraged to strive to approximate the ideal"), i.e. to approach the exact value; the 21.3.2 Note uses "recommended (but not specified by this standard) that implementations use the approximation algorithms ... contained in fdlibm". The page puts "encouraged" in 21.3.2 and attaches it to using "the approximation that is generally available". Substance (non-binding) right, quotation misplaced. - MINOR | [sqrt-ctl-es2024] CONTROL. IEEE 754 requires square root to be correctly rounded, and ECMA-262 21.3.2.32 defers to it rather than listing it as implementation-approximated. Exactly one answer is conforming. | https://262.ecma-international.org/15.0/ | In ES2024 21.3.2.32 is Math.sqrt and says "Return an implementation-approximated Number value representing the result of the square root", and the 21.3.2 Note lists sqrt (24 names). So the clause cited does list sqrt as implementation-approximated in the only edition where 21.3.2.32 is sqrt. - CONFIRMED | [sqrt-ctl-es2025] CONTROL. IEEE 754 requires square root to be correctly rounded, and ECMA-262 21.3.2.32 defers to it rather than listing it as implementation-approximated. Exactly one answer is conforming. | https://tc39.es/ecma262/2025/ | ES2025 (and ES2026, draft) Math.sqrt: "Return F(the square root of R(n))", sqrt absent from the Note. True in substance, but the clause is 21.3.2.33; 21.3.2.32 is Math.sinh in ES2025/2026 (see finding). - CONFIRMED | [sqrt-ctl-ieee] CONTROL. IEEE 754 requires square root to be correctly rounded, and ECMA-262 21.3.2.32 defers to it rather than listing it as implementation-approximated. Exactly one answer is conforming. | https://en.wikipedia.org/wiki/IEEE_754 | Correct rounding is the default requirement and square root is a required operation. - CONFIRMED | [sum-ctl-es] CONTROL. IEEE 754 addition is correctly rounded and ECMA-262 requires it. The famous answer is not an error; it is the only conforming one. | https://tc39.es/ecma262/2025/ | 6.1.6.1.7 Number::add: "It performs addition according to the rules of IEEE 754-2019 binary double-precision arithmetic". - CONFIRMED | [ieee-correct-rounding-prose] IEEE 754 requires correct rounding, so exactly one bit pattern conforms | https://en.wikipedia.org/wiki/IEEE_754 | Correct rounding under the default mode gives one result; for Math.sqrt this holds in ECMA-262 only from ES2025. - WRONG | [tostring-ctl] CONTROL. ECMA-262 6.1.6.1.20 specifies Number::toString exactly, digit by digit: the shortest decimal that round-trips. One conforming answer. | https://tc39.es/ecma262/2025/ | 6.1.6.1.20 step 5 asks only that k be as small as possible "and that the least significant digit of s is not necessarily uniquely determined by these criteria"; Note 1 repeats it; choosing the closest digit is Note 2, "recommended". For 0.1 + 0.2 the shortest k is 17 and six strings satisfy step 5: 0.30000000000000002 through 0.30000000000000007 all round to the same double (checked with exact rationals). String(0.1 + 0.2) is latitude nobody took, not a determined control. Fixed (triage K6): the row is now counted as latitude (kind declined), so 3 controls, 19 comparable, 14 agreed, 5 divided; answers unchanged. - CONFIRMED | [normalize-ctl-policy] CONTROL. Unicode normalisation is stable by policy: a composition, once defined, may never change | https://www.unicode.org/policies/stability_policy.html | Strong Normalization Stability: NFC of a string of assigned characters is identical in every later version; "its canonical combining class and decomposition mapping will not be changed in any way". The second half of the old clause is the next line (fixed). - WRONG | [normalize-ctl-tr15] CONTROL. Unicode normalisation is stable by policy: a composition, once defined, may never change, and no new composition may be added. | https://www.unicode.org/reports/tr15/ | UAX #15 section 3: "It would be possible to add more compositions in a future version of Unicode ... Z must be a new character, and either X or Y must be a new character. However, the Unicode Consortium strongly discourages new compositions". New compositions are discouraged and restricted, not forbidden. The measured control (e + U+0301) is unaffected. Fixed (triage K7): the clause now says a new composition is allowed only for a newly encoded character and strongly discouraged even then. - CONFIRMED | [normalize-frozen-prose] Unicode normalisation (frozen by policy: a composition, once defined, may never change) | https://www.unicode.org/reports/tr15/ | Composition version fixed at Unicode 3.1.0 plus the stability policy: an existing composition does not change. - CONFIRMED | [es402-no-version] ECMA-402 requires the IANA time zone database but names no version of it. | https://tc39.es/ecma402/ | 6.5 requires the IANA database and names no release. - CONFIRMED | [cancun-2024b-news] The answer for this instant changed in tzdata 2024b. | https://raw.githubusercontent.com/eggert/tz/main/NEWS | NEWS, Release 2024b (2024-09-04): Mexico historical transitions updated for 1981-1997, America/Cancun among the zones. - CONFIRMED | [cancun-2024a-zone] The answer for this instant changed in tzdata 2024b. | https://raw.githubusercontent.com/eggert/tz/2024a/northamerica | At 2024a Cancun is "-5:00 Mexico E%sT" from 1981 Dec 23 to 1998; the Mexico rules have no rule in force in 1990, so 12:00Z is 07:00. - CONFIRMED | [cancun-2024b-zone] The answer for this instant changed in tzdata 2024b. | https://raw.githubusercontent.com/eggert/tz/2024b/northamerica | At 2024b Cancun is "-6:00 Mexico C%sT" from 1983 Jan 4 to 1997 Oct 26, so 12:00Z on 1990-07-15 is 06:00. Local node v22.23.1 (tz 2026a) also says 06:00. - CONFIRMED | [kyiv-rename] IANA renamed Europe/Kiev to Europe/Kyiv in tzdata 2022b. | https://raw.githubusercontent.com/eggert/tz/main/NEWS | NEWS Release 2022b: "Rename Europe/Kiev to Europe/Kyiv." - WRONG | [kyiv-canonical-cldr] IANA renamed Europe/Kiev to Europe/Kyiv in tzdata 2022b; whether a runtime reports the new name or canonicalises back to the old one is its data version talking. | https://tc39.es/ecma402/ | Second half not so. ECMA-402 6.5: implementations "should initially add each replacement time zone identifier as a non-primary time zone identifier that resolves to the existing renamed time zone identifier" and promote it after a "rename waiting period" (recommended two years): a policy, not a data release. V8 keeps CLDR's canonical Europe/Kiev whatever its data: node v22.23.1 here with process.versions.tz 2026a and CLDR 48 still resolves Europe/Kyiv to Europe/Kiev (recreated: node -e on this machine). Fixed (triage K8): the clause now says it is not the tzdata version talking, citing the rename waiting period and the 2026a Node. - WRONG | [cldr42-changed-tr-notes42] ECMA-402 leaves the display names to the implementation's locale data. Every browser takes them from CLDR, and CLDR 42 (October 2022) changed this one. | https://cldr.unicode.org/downloads/cldr-42 | CLDR 42 kept "Turkey" as the standard English name and only added "Türkiye" as alt="variant"; its note says the switch would come "possibly as early as CLDR 43". en.xml at release-42 has Turkey, at release-43 Türkiye (raw.githubusercontent.com/unicode-org/cldr/release-4x/common/main/en.xml). Fixed (triage K9): the clause now says CLDR 43 (April 2023). - WRONG | [cldr42-changed-tr-notes43] ECMA-402 leaves the display names to the implementation's locale data. Every browser takes them from CLDR, and CLDR 42 (October 2022) changed this one. | https://cldr.unicode.org/downloads/cldr-43 | "In v42, the customary English name for the country code TR was “Turkey” ... In v43, the customary English name was changed to “Türkiye”". The change Intl.DisplayNames reports came in CLDR 43 (published 2023-04-11), not CLDR 42. Fixed (triage K10), same edit as K9. - CONFIRMED | [cldr43-date] (as corrected) Every browser takes them from CLDR, and CLDR 43 (April 2023) changed this one. | https://api.github.com/repos/unicode-org/cldr/releases/tags/release-43 | release-43 "published_at": "2023-04-11"; cldr.unicode.org/downloads/cldr-43: "In English, Türkiye is now the primary country name for the country code TR, and Turkey is available as an alternate." The N2 line here confirmed only the date of CLDR 42 (2022-10-18), which the page no longer gives. - CONFIRMED | [es402-displaynames-locale-data] ECMA-402 leaves the display names to the implementation's locale data. | https://tc39.es/ecma402/ | Intl.DisplayNames [[LocaleData]] "implementation-defined within the constraints described in 9.1". - CONFIRMED | [every-browser-cldr] Every browser takes them from CLDR | https://cldr.unicode.org/ | Record had no source. CLDR home page: "CLDR is incorporated into all modern operating systems and browsers". (CLDR's own statement about its users.) - CONFIRMED | [welsh-revised] The plural categories come from CLDR, which has revised Welsh. | https://raw.githubusercontent.com/unicode-org/cldr/release-1-9/common/supplemental/plurals.xml | Record had no source (searched release-26 onward only). CLDR revised Welsh between 1.8 and 1.9: release-1-8 plurals.xml has cy one/two/"many" n is 8 or n is 11; release-1-9 has zero/one/two/few (n is 3)/many (n is 6). True, though 2010, so no captured runtime could differ on it. - CONFIRMED | [node-order] Node 21.7.3 is a higher version than Node 20.20.2 and an older build, because the 20 line kept getting patch releases after the 21 line stopped. | https://nodejs.org/dist/index.json | dist index: v21.7.3 2024-04-10 (the last v21); v20.20.2 2026-03-24 (the last v20). - CONFIRMED | [tz-count-es402] ECMA-402 specifies the function, not the list. The list is whatever database this build was compiled against. | https://tc39.es/ecma402/ | supportedValuesOf("timeZone") returns AvailablePrimaryTimeZoneIdentifiers(); 6.5 derives it from whatever IANA release the implementation carries. - CONFIRMED | [host-timezone-clause] The host decides what to say about where it is, and is permitted to say UTC. | https://tc39.es/ecma262/2025/ | SystemTimeZoneIdentifier is implementation-defined, "If the implementation only supports the UTC time zone, return \"UTC\"". - CONFIRMED | [host-locale-clause] The default locale is the host environment's current locale, and what that means is up to the host. | https://tc39.es/ecma402/ | DefaultLocale is implementation-defined, "the host environment's current locale". - CONFIRMED | [collate-clause] ECMA-402 requires a locale-sensitive comparison and specifies no order; the order is the collation data's. | https://tc39.es/ecma402/ | [[SortLocaleData]] implementation-defined. - CONFIRMED | [numberformat-clause] ECMA-402 defers the grouping pattern to locale data. | https://tc39.es/ecma402/ | Stronger than the record's excerpt: "The positions of grouping separators, and whether to display grouping separators for a formatted number, is implementation-defined." - CONFIRMED | [segment-es402] ECMA-402 requires segmentation and defers the boundaries to the implementation's data, which for Thai is a dictionary, not a rule. | https://tc39.es/ecma402/ | Intl.Segmenter [[LocaleData]] implementation-defined. - CONFIRMED | [segment-uax29-thai] ECMA-402 requires segmentation and defers the boundaries to the implementation's data, which for Thai is a dictionary, not a rule. | https://www.unicode.org/reports/tr29/ | UAX #29: Thai word boundaries require "dictionary lookup or other mechanisms". - MINOR | [hrtime-clause] High Resolution Time lets a user agent coarsen the clock to any resolution it judges safe, and they choose differently. | https://www.w3.org/TR/hr-time-3/ | "Let time resolution be 100 microseconds, or a higher implementation-defined value" (5 microseconds when cross-origin isolated): a floor, not "any resolution it judges safe". Only one user agent was captured (Chromium, 1e-1 ms = the floor); the split is with Node and Bun (5e-5 and 3e-5 ms), which are not user agents coarsening under this clause, so "they choose differently" is not what the row shows. - CONFIRMED | [canvas-spec-raster] The 2D canvas specification fixes the geometry of a drawing and says nothing about how it is rasterised. | https://html.spec.whatwg.org/multipage/canvas.html | The canvas spec prescribes no rasterisation algorithm; it mentions anti-aliasing only permissively ("can similarly be implemented using oversampling"). "Says nothing" is slightly strong, not false in substance. - CONFIRMED | [canvas-spec-kerning] Which font a generic family resolves to, how the edges are antialiased, whether the text is hinted: all of that is left to the implementation, correctly, because it depends on the machine. | https://www.w3.org/TR/css-fonts-4/ | Kerning "at the discretion of the user agent" (canvas spec); generic families "are expected to map to different fonts on different platforms. ... User agents should provide reasonable default choices" (CSS Fonts 4). - CONFIRMED | [canvas-clause-host] The 2D canvas specification fixes the geometry and says nothing about rasterisation: antialiasing, hinting, and which font a generic family resolves to are all the implementation's. This is the divergence commercial fingerprinting is built on. | https://inria.hal.science/hal-01285470v2 | Abstract: the Canvas API gives "highly discriminating attributes" because it "relies on multiple layers of the user's system". Paper text read from web.archive.org/web/2019id_/https://hal.inria.fr/hal-01285470v2/document (HAL itself returns a bot challenge). - CONFIRMED | [canvas-textwidth-clause] CSS names a generic family and the user agent chooses the face. Nothing in any standard says which font that is or how wide it will be. | https://www.w3.org/TR/css-fonts-4/ | Record had no source. CSS Fonts 4, generic families: "They are expected to map to different fonts on different platforms ... and are not particular about which specific font is chosen". No standard fixes the face, so none fixes the width. - CONFIRMED | [mowery-2012] Mowery and Shacham named the technique in 2012 (Pixel Perfect: Fingerprinting Canvas in HTML5, W2SP) | https://hovav.net/ucsd/papers/ms12.html | "Pixel Perfect: Fingerprinting Canvas in HTML5 By Keaton Mowery and Hovav Shacham. In Proceedings of W2SP 2012." - CONFIRMED | [laperdrix-2016] Laperdrix, Rudametkin and Baudry measured it in the wild in 2016 (Beauty and the Beast, IEEE S&P) | https://inria.hal.science/hal-01285470v2 | Title, authors, IEEE S&P May 2016; DOI 10.1109/SP.2016.57 matches at Crossref. - CONFIRMED | [laperdrix-discriminating] where canvas rendering was among the most discriminating attributes of the browser fingerprint. | https://web.archive.org/web/2019id_/https://hal.inria.fr/hal-01285470v2/document | PDF read: "We show that canvas fingerprinting is one of the most discriminating attributes"; "normalized entropy is at 0.491, putting it in the top 5 of the most discriminating attributes". - CONFIRMED | [canvas-stable-different] the difference is stable for one machine and different across machines. | https://web.archive.org/web/2019id_/https://hal.inria.fr/hal-01285470v2/document | PDF read: "canvas fingerprinting can be used to differentiate devices with pixel precision"; "one of the advantages of canvas fingerprinting is that it is stable. You can run it many times on the same computer and you will have the same result every time". Mowery and Shacham: "consistent, high-entropy". - CONFIRMED | [somebody-mining] Somebody is already mining it, and the reason they can is that everyone else reads a specification as a promise that the answer is the same everywhere. | https://web.archive.org/web/2019id_/https://hal.inria.fr/hal-01285470v2/document | PDF read: "This technique is gaining popularity in tracking scripts". Supports the first half; the second half is the page's argument. - UNVERIFIABLE | [worth-money] it is worth money precisely in proportion to how undetermined it is. | none: Laperdrix 2016, Mowery 2012, Wikipedia Canvas fingerprinting read; no pricing of fingerprinting | The page's own assertion; "precisely in proportion" is not measured anywhere on the page or in its sources. - CONFIRMED | [zip-twice] A ZIP records each entry's metadata twice, once in the local header and once in the central directory | https://pkware.cachefly.net/webdocs/casestudies/APPNOTE.TXT | APPNOTE 4.3.2: each local file header "MUST be accompanied by a corresponding \"central directory header\" record". - CONFIRMED | [font-three-extents] a TrueType font carries three different vertical extents | https://learn.microsoft.com/en-us/typography/opentype/spec/os2 | OS/2 carries sTypo* and usWin*; hhea is a third set. - CONFIRMED | [font-flag-not-everyone] OpenType's rule for choosing between them is a flag that not everyone sets | https://learn.microsoft.com/en-us/typography/opentype/spec/os2 | Record had no source. "Versions 0 to 3: ... Bits 7 to 15 are reserved and must be set to 0"; "Bits 7 to 9 were defined in version 4 (OpenType 1.5)". So older fonts cannot set it. On this machine 48 of 401 fonts under /usr/share/fonts set bit 7 and 353 do not (recreated: a 20-line OS/2 reader). - CONFIRMED | [font-browsers-line-heights] the reason two browsers can lay out the same font at different line heights while both conform | https://www.w3.org/TR/css-inline-3/ | The record's source (the font layer's concession) does not bear on browsers. CSS Inline 3: "It is recommended that implementations that use OpenType or TrueType fonts use the metrics sTypoAscender and sTypoDescender ... In the absence of these metrics, the \"Ascent\" and \"Descent\" metrics from the HHEA table should be used." A recommendation in a note, so two conforming engines may use different metrics. - CONFIRMED | [encoding-several] a text file's bytes are frequently valid under several encodings at once | https://en.wikipedia.org/wiki/Charset_detection | "all arrangements of bytes are valid. There is no technical way to tell these encodings apart". - MINOR | [cmu-standard] Ask the standard pronouncing dictionary of American English how a word is said | http://www.speech.cs.cmu.edu/cgi-bin/cmudict | CMU calls itself "an open-source machine-readable pronunciation dictionary for North American English"; nothing makes it "the standard pronouncing dictionary of American English", and Kind V then calls it "The authority itself". The parent layer says only "the CMU Pronouncing Dictionary". - CONFIRMED | [node21-stopped] Node 21.7.3 is a higher version than Node 20.20.2 and an older build, because the 20 line kept getting patch releases after the 21 line stopped. | https://raw.githubusercontent.com/nodejs/Release/main/schedule.json | v21 end 2024-06-01; v20 end 2026-04-30. - MINOR | [cannot-be-patched] That is the whole of canvas fingerprinting. It is not an exploit and it patches nothing, because there is no defect | https://brave.com/privacy-updates/4-fingerprinting-defenses-2.0/ | Dek: "it cannot be patched, because nothing is broken". Browsers ship changes that do close it: Brave "applies randomization-based protections to canvas" (farbling of toDataURL, getImageData, measureText), and Tor Browser offers to "return blank image data" (Wikipedia, Mitigation). Not bug fixes, but the latitude can be and is taken back by browsers. - CONFIRMED | [layer-rounding-23] twenty-three functions named as implementation-approximated, measured against a correctly-rounded answer built in BigInt | src/content/strata/the-rounding-you-were-allowed.md | Parent dek, and the ES2025 21.3.2 Note lists 23 names (acos ... tanh, without sqrt). - WRONG | [layer-rounding-sqrt-struck] ECMA-262 21.3.2 names twenty-three functions as implementation-approximated and has done since 1997. | src/content/strata/the-rounding-you-were-allowed.md | "has done since 1997" is false for the number: the 1997 first edition (15.8.2) names ten ("acos, asin, atan, atan2, cos, exp, log, pow, sin, and sqrt"), ES2024 names 24 (with sqrt), ES2025 23; the parent itself says sqrt was struck in 2025. Also 21.3.2 and the term "implementation-approximated" postdate 1997. Fixed (triage K14): the note now says the Math object has carried that latitude since 1997, when the first edition named ten. - MINOR | [layer-rounding-bigint] That layer builds the correctly-rounded answer in BigInt and measures five engines against it. | src/content/strata/the-rounding-you-were-allowed.md | The parent captured "Five arithmetics": Chromium, Firefox, WebKit, an older Node and glibc through python3. Two are the same engine (V8) and one is a C library, so "five engines" overstates it. - CONFIRMED | [layer-cores-webkit] That layer read the three engines' source and found WebKit returning one of exactly two constants. | src/content/strata/how-many-cores-do-you-have.md | Parent dek: "WebKit returns one of exactly two constants, 4 or 8". - CONFIRMED | [layer-cores-webkit-prose] (three engines' source read, and WebKit returning one of exactly two constants) | src/content/strata/how-many-cores-do-you-have.md | Parent dek: "Read the three engines' source". - CONFIRMED | [layer-archive] A ZIP stores every file's metadata twice and the two copies can drift, so a reader has to choose which is authoritative and the format does not say. | src/content/strata/the-archive-that-disagrees-with-itself.md | Parent: "generic ZIP differences receive no universal winner"; APPNOTE has no rule naming either copy authoritative (grep "authoritative": 0 hits; control: "MUST be preceded by" found). - CONFIRMED | [layer-font] a TrueType file carries three separate vertical extents, hhea, sTypo and usWin | src/content/strata/your-font-does-not-agree-with-itself.md | Parent: "can carry three different vertical extents"; names hhea, sTypo, usWin. - MINOR | [layer-encoding] That layer's refusal to name a winner is the behaviour every other member of Kind II ought to have and mostly does not. | src/content/strata/the-encoding-your-text-file-admits-to.md | The parent does show "A statistical favorite ... separately". "every other member of Kind II ... mostly does not": the other two Kind II members decline too (archive: "no universal winner"; font: "not predictions of any named browser"). - CONFIRMED | [layer-past-249] That layer compiled 249 archived releases of the time zone database and found 145 of 248 consecutive releases changing a clock reading that had already happened. | src/content/strata/the-past-has-a-version-number.md | Parent dek: 249 releases; "145 of 248 consecutive releases". - CONFIRMED | [layer-dpd] a complete census finds 1,000 of 1,024 ten-bit patterns written by the encoder and exactly 24 more that still decode | src/content/strata/the-twenty-four-codes-no-encoder-writes.md | Parent dek: 1,000 of 1,024; "exactly 24 accepted decoder inputs with no encoder preimage". - CONFIRMED | [layer-twowords-3928] 39.28% of running words carry more than one reading, thirteen named variations account for 36.63 points | public/strata/two-words-in-every-five/data.json | data.json ladder: multiPct 39.2833, explainedPct 36.6289, 13 rules. Record excerpt is a reformatted copy of the compact JSON (values match). - CONFIRMED | [layer-twowords-0218] what the argument about spelling reform is actually about turns out to be 0.218% | public/strata/two-words-in-every-five/data.json | homograph 0.20602% + case 0.01195% = 0.21797%, and the parent's "0.22%, about one word in 459" (1/459 = 0.2179%). - CONFIRMED | [layer-twowords-prose] The standard pronouncing dictionary of American English gives more than one reading for 39.28 per cent of running words | src/content/strata/two-words-in-every-five.md | Parent: "for 39.28% of running words the dictionary gives more than one pronunciation". - CONFIRMED | [layer-price] That one walks functions cheap forward and expensive backward and is about the UNIT the price is quoted in; the standards it audits disagree because they pin one curve at different points. | src/content/strata/the-price-of-going-backward.md | Parent: "a function that is cheap forward and expensive backward"; "one curve staked 7.16 bits apart". - CONFIRMED | [own-counts] (as corrected) five of the nineteen with latitude came back with more than one answer, all three controls came back identical | research/correct-is-not-determined/data/capture.json | Recounted from the rows by the verifier after the fix: divided 5, agreed 14, comparable 19, controls 3, controlsDivided 0 (61/61). Before the fix the page said five of eighteen and four controls; see [tostring-ctl]. - CONFIRMED | [own-thirteen] (as corrected) which is why fourteen of the nineteen rows here came back agreed | research/correct-is-not-determined/data/capture.json | agreed 14 of comparable 19, recounted by the verifier. It said thirteen of eighteen until the String(0.1 + 0.2) row was moved out of the controls. - OBSERVED | [own-cancun-dek] (as corrected) including what o'clock it was in Cancun on a July morning in 1990, where the runtimes split 06:00 against 07:00 and the version numbers do not run in order | research/correct-is-not-determined/data/capture.json | Row tz.cancun.1990: 06:00 (node22, node20, chromium), 07:00 (node21, bun); the instant is 12:00Z (battery.mjs). "afternoon" was the separate WRONG line below, fixed. - OBSERVED | [own-cancun-relates] gets 06:00 from three and 07:00 from two, with the older Node saying 07:00 and the still older one saying 06:00 | research/correct-is-not-determined/data/capture.json | Three 06:00, two 07:00; node21 07:00, node20 06:00. - OBSERVED | [own-same-machine] Every runtime above sits on the same machine | research/correct-is-not-determined/capture.mjs | capture.mjs runs each runtime it finds at the paths it names, all on the machine it runs on; capture.json "machine": "linux x64". - OBSERVED | [own-engines] Two distinct JavaScript engines: V8 (three Node builds and Chromium) and JavaScriptCore (Bun). | research/correct-is-not-determined/data/capture.json | runtimes: V8 x3, JavaScriptCore (bun 1.3.11), V8 in Chromium 141.0.7390.37. - OBSERVED | [own-tanh-expm1] on the five runtimes captured here, Math.tanh and Math.expm1 agree exactly | research/correct-is-not-determined/data/capture.json | tanh 3fe85efab514f394 and expm1 3e45798ee3fdb764 on all five. - OBSERVED | [own-controls-identical] (as corrected) All three came back identical across all five implementations. | research/correct-is-not-determined/data/capture.json | One value per control row (ctl.sqrt2, ctl.sum, ctl.normalize); the verifier re-ran them here (node v22.23.1) and got the same values. It said "All four" until the String(0.1 + 0.2) row was moved out. - OBSERVED | [own-no-firefox] there is no Firefox or Safari here, because the container this was captured in could not download either. | memory/log.d/2026-08-23T1800Z-two-words-and-the-gap-that-follows-it.md | Build session log: "npx playwright install firefox webkit fails at egress in this container"; README repeats it. - CONFIRMED | [own-canvas-hash] the canvas probe reduces a drawing to a 32-bit hash and discards the image. | research/correct-is-not-determined/battery.mjs | 32-bit FNV-1a over the PNG data URL; the image is not kept. - MINOR | [own-nothing-transmitted] Nothing is transmitted, no identifier is derived | public/strata/nothing-here-is-a-bug/page.js | Nothing derived leaves the tab (page.js fetches only ./capture.json). But "no identifier is derived" is not literally so: the page computes a 32-bit hash of a canvas rendering, which is the canvas-fingerprint value the page itself describes, alongside time zone, locale, cores and text width. Accurate wording: derived, shown, never sent or stored. - CONFIRMED | [own-compared-locally] the result is compared with the captured column in your browser and nowhere else. | public/strata/nothing-here-is-a-bug/page.js | The only fetch is ./capture.json; comparison runs in the tab. - CONFIRMED | [own-byte-for-byte] (as corrected) Your own browser then answers the identical battery, the same probe code byte for byte, in your tab and nowhere else. | public/strata/nothing-here-is-a-bug/page.js | page.js imports ./battery.js, a verbatim copy of battery.mjs (verifier: byte-identical). The clause notes in battery.mjs were corrected 2026-10-18 after the capture, so the file is no longer byte-identical to the one the runtimes ran (commit 625bead4a8c); the ids, questions and run functions are (sha256 7f49f28c796e..., now asserted by the verifier). The dek said "byte-for-byte the same file" and was narrowed with the fix. - MINOR | [own-capture-mjs] node research/correct-is-not-determined/capture.mjs re-runs the battery on every runtime it can find on the machine and rewrites data/capture.json | research/correct-is-not-determined/capture.mjs | It looks only at five hard-coded paths (/opt/node22, /opt/node21, /opt/node20, /root/.bun, /opt/pw-browsers/chromium-1194). Run in a scratch copy on this machine, which has node v22.23.1 on PATH, it printed "skip" for all five, wrote a capture with 0 runtimes over data/capture.json, after which the verifier passed 23/34. The README says it "finds every runtime present on the machine". - WRONG | [own-verify] node research/correct-is-not-determined/verify-nothing-here-is-a-bug.mjs re-derives the counts, re-runs the controls, and checks that every figure in this page's prose matches the capture. | research/correct-is-not-determined/verify-nothing-here-is-a-bug.mjs | It does not check "every figure in this page's prose". The divided count is found only in the meta descriptions. In a scratch copy I changed the body's "Five of eighteen" to "Six of eighteen" and "Node 21.7.3" to "Node 23.7.3": verifier 56/56, exit 0. Control: changing "Five of the eighteen" in the meta tags fails it (55/56). Also unchecked: "thirteen", "06:00", "07:00", "36". Fixed (triage K19): the verifier now also checks the body's quoted count, the comparable and control counts in the body and the Node versions named (red on purpose: "Six of nineteen" and "Node 23.7.3" in a scratch copy fail 59/61), and the sentence says headline counts. - CONFIRMED | [own-no-network] Neither reaches the network. | research/correct-is-not-determined/capture.mjs | capture.mjs imports no network module and evaluates the battery in Chromium on a blank page; the verifier spawns only python3. - CONFIRMED | [own-verify-headline] checks the headline row through a non-JavaScript implementation | research/correct-is-not-determined/verify-nothing-here-is-a-bug.mjs | True as stated; the check passes if CPython agrees with either 06:00 or 07:00, so it cannot say which build is current (CPython over system zoneinfo says 06:00 here). - CONFIRMED | [own-footer-scout] The spine it walks was identified by a survey pass on 2026-08-20 recorded in oversight/ideas/scouted-2026-08-20.md, which called it the sharpest genuinely unwalked epistemics spine in the corpus and did not build it. | oversight/ideas/scouted-2026-08-20.md | Item 5 "Conformance permits disagreement"; file opens "Scouted 2026-08-20, and not taken". - CONFIRMED | [own-timer-onesig] Reported to one significant figure, because the aim is the clamp and not the jitter around it. | research/correct-is-not-determined/battery.mjs | toPrecision(1). - UNVERIFIABLE | [only-ii-iv-closable] Only Kind II and Kind IV are closable. | none: the page's taxonomy; no outside source makes the claim | The page's own argument from its taxonomy, not a claim about a source. - CONFIRMED | [register-os-update] because the register can be updated by the operating system beneath it. | https://docs.python.org/3/library/zoneinfo.html | Record had no source. "By default, zoneinfo uses the system's time zone data if available" (the verifier's own CPython path is an example of a register the OS updates). - MINOR | [none-said-the-shape] none of them said the shape was the finding. | src/content/strata/two-words-in-every-five.md | One of the eight said it, before the portal (commit a9054f010f1, 16:00Z; portal 625bead4a8c, 16:39Z): its relates note to the rounding layer says "a standard authorises its own indeterminacy ... In both cases the gap is not between right and wrong, it is between correct and determined." It was written by the same session, so "independently" is also loose for that member. - WRONG | (added) what o'clock it was in Cancun on a July afternoon in 1990 / The clock reading in Cancún on a July afternoon in 1990 | research/correct-is-not-determined/battery.mjs | The instant is 12:00Z (Date.UTC(1990, 6, 15, 12)); the page's own answers are 06:00 and 07:00 local, a morning in Cancun under either reading (the parent layer says "six in the morning"). Fixed (triage K21): now "a July morning" in the dek and the body. - MINOR | (added) The portal borrows two of its functions as probes (Math.tanh and Math.expm1) | public/strata/the-rounding-you-were-allowed/data.js | The parent measured eight functions: FUNCTIONS = ["exp","log","sin","cos","tan","atan","cbrt","sqrt"]. tanh and expm1 are in clause 21.3.2, not among that layer's functions. - CONFIRMED | (added) math.expm1 clause: ECMA-262 21.3.2, as above. | https://tc39.es/ecma262/2025/ | expm1 is in the 21.3.2 Note list (ES2024, 2025, 2026) and returns an implementation-approximated value. - CONFIRMED | (added) Neither reading is wrong; NaN is also a correct answer. | https://262.ecma-international.org/15.0/ | 21.4.3.2: implementation-specific heuristics are permitted, and "Strings that are unrecognizable ... shall cause this function to return NaN". - CONFIRMED | (added) ECMA-402 defers the skeleton and its literals to locale data. | https://tc39.es/ecma402/ | "[[LocaleData]].[[]] must have a [[styles]] field"; DateTimeStyleFormat reads styles from [[LocaleData]], which is implementation-defined. - MINOR | (added) Then your own browser answers the same twenty-four ... without sending anything anywhere. | https://artwaste.land/strata/nothing-here-is-a-bug/ | The answers are not sent (page.js). The served page does send a page view: an inline script POSTs the path to /api/pulse and it loads https://static.cloudflareinsights.com/beacon.min.js (not in the repo HTML; added at serving). Neither reads the battery. - CONFIRMED | (added) The portal puts one of those instants, Cancun at noon UTC on 15 July 1990 | src/content/strata/the-past-has-a-version-number.md | Parent dek: Cancún "on a July afternoon in 1990 ... said six in the morning, then said seven for twenty-three years, then in September 2024 went back to six". - OBSERVED | (added) four command-line runtimes and one browser, two distinct JavaScript engines between them | research/correct-is-not-determined/data/capture.json | node22, node21, node20, bun (kind runtime) and chromium (kind browser); V8 and JavaScriptCore. - CONFIRMED | (added) they do not agree about what o'clock it was thirty-six years ago | research/correct-is-not-determined/battery.mjs | 1990 to 2026 is 36 years. - CONFIRMED | (added) Densely Packed Decimal defines both directions and the two are not inverse | https://en.wikipedia.org/wiki/Densely_packed_decimal | "The 3 x 8 = 24 non-standard encodings fill in the gap between 10^3 = 1000 and 2^10 = 1024." ## Findings - [fixed] html-cores-any-number (WRONG, K1): "the HTML standard says a browser may report any number of processors it likes" (index.html, Kind I) now says the standard lets a browser report fewer processors than it has, down to one, to limit fingerprinting. Corrected line on the page and in the markdown. - [fixed] html-cores-any-number (WRONG, K2): the same statement in the relates note to how-many-cores-do-you-have ("the processor count may be any value the user agent likes") now says "any processor count from one up to the number it has"; the nearby-block teaser on the page was updated to the generator's output (nearby-layers --check on an isolated copy: 0 pages need rebuilding). Same Corrected line. - [fixed] html-cores-misquote (WRONG, K3): the battery clause quoted the HTML standard as saying "may return a lower value than the actual number of logical processors", words it does not contain; restored to the standard's own words ("must return a number between 1 and the number of logical processors potentially available to the user agent"; "when the user agent desires to limit fingerprinting possibilities"). battery.js rebuilt with build-page.mjs; data/capture.json clause field updated to match. Same Corrected line as K1. - [fixed] tostring-not-determined (WRONG, K6): String(0.1 + 0.2) was a control on the claim that ECMA-262 specifies Number-to-string digit by digit; 6.1.6.1.20 leaves the last digit open. The probe's kind is now declined with a clause quoting step 5; data/capture.json row kind and counters recomputed from the rows (comparable 19, agreed 14, controls 3, divided 5, controlsDivided 0); every copy of the counts changed (seoDescription, dek, relates note, markdown body, meta, og, twitter, JSON-LD, body, check box, README, verifier needles). Corrected line. - [fixed] unicode-no-new-compositions (WRONG, K7): the normalisation control clause's "and no new composition may be added" now says a new one is allowed only for a newly encoded character and strongly discouraged even then. Corrected line. - [fixed] kyiv-data-version (WRONG, K8): the Kyiv clause's "is its data version talking" now says it is not the tzdata version talking, citing ECMA-402's rename waiting period and the Node with tzdata 2026a that still answers Europe/Kiev. Corrected line. - [fixed] cldr42-not-43 (WRONG, K9): the Türkiye clause's "CLDR 42 (October 2022) changed this one" now says CLDR 43 (April 2023): en.xml at release-42 still has Turkey as the standard name. Corrected line. - [fixed] cldr42-not-43 (WRONG, K10): the same sentence checked against the CLDR 43 release note ("Türkiye is now the primary country name"); fixed by the same edit and the same Corrected line. - [fixed] twenty-three-since-1997 (WRONG, K14): the relates note to the-rounding-you-were-allowed (and its nearby-block teaser) now says ECMA-262 21.3.2 names twenty-three functions and the Math object has carried that latitude since 1997, when the first edition named ten. Corrected line. - [fixed] verifier-every-figure (WRONG, K19): "checks that every figure in this page's prose matches the capture" (index.html) and "asserts that the page's prose matches the data" (markdown) now say headline counts; the verifier now also checks the body's quoted count, the comparable and control counts in the body, and the Node versions named, and went red on purpose in a scratch copy (59/61). Corrected line. - [fixed] cancun-afternoon (WRONG, K21): "on a July afternoon in 1990" (dek, body) is now "a July morning": the instant is 12:00Z, 06:00 or 07:00 in Cancún. Corrected line. - [fixed] capture-finds-every-runtime (MINOR, K18): "re-runs the battery on every runtime it can find on the machine" shares a sentence with K19, so it was fixed: "on each runtime it finds at the paths it names"; README and capture.mjs header comment likewise. No line on the page (MINOR). - [declined] tanh-encouraged-misplaced (MINOR, K4): "encouraged" is ECMA-262 4.2's word, 21.3.2 says "recommended"; substance right, tier B (triage K4). - [declined] sqrt-clause-number (MINOR, K5): the sqrt control cites 21.3.2.32, which is Math.sinh in ES2025 (sqrt is 21.3.2.33 and exact there); a clause-number slip, tier B (triage K5). - [declined] hrtime-any-resolution (MINOR, K11): High Resolution Time sets a floor, not "any resolution it judges safe", and the split row mixes one user agent with Node and Bun; loose, tier B (triage K11). - [declined] cmu-the-standard (MINOR, K12): CMUdict is "a machine-readable pronunciation dictionary for North American English", not "the standard pronouncing dictionary"; loose, tier B (triage K12). - [declined] cannot-be-patched (MINOR, K13): browsers can take the canvas latitude back (Brave farbling, Tor blank image data), though not as a bug fix; loose, tier B (triage K13). - [declined] five-engines (MINOR, K15): the parent's five arithmetics include glibc and two V8 builds; loose, tier B (triage K15). - [declined] kind-ii-mostly-not (MINOR, K16): the archive and font layers also decline to name a winner; overstatement, tier B (triage K16). - [declined] no-identifier-derived (MINOR, K17): nothing derived leaves the tab, but the canvas hash is computed and shown; loose, tier B (triage K17). - [declined] none-said-the-shape (MINOR, K20): two-words-in-every-five's relates note, written 39 minutes earlier by the same session, named the gap; loose, tier B (triage K20). - [declined] borrows-two-functions (MINOR, K22): tanh and expm1 are from the parent's clause, not among its eight measured functions; loose, tier B (triage K22). - [declined] sending-nothing (MINOR, K23): the battery's answers are not sent; the served page's site-wide page-view beacon is; loose, tier B (triage K23).