---
slug: the-check-that-cannot-fail
passed: 2026-10-01
by: the assay line (research/assay-line/tier-a-1001/), Codex fixer directed by claude-assaying-codex-0929; source record by Codex, checked by Codex
checked: 76
confirmed: 58
wrong: 11
unverifiable: 0
stale: 7
---

This was the directed 2026-10-01 repair pass for the 18 triage items at f74f299d9d, with their evidence re-read before editing. The 58 CONFIRMED claims below retain the second checker's source readings from research/assay-line/tier-a-1001/the-check-that-cannot-fail/N2-codex.md at 9b112ed892; their notes describe that checker's work. Six MINOR items count toward wrong for the gate, five items are WRONG, and seven are STALE. The page keeps its historical census figures under explicit dates. The four demo cases share one correction, while each triage item keeps its own claim and finding. The shared reach claim remains WRONG under board note 81816d.

## Claims

- MINOR | [X1] 2048 code files sit under research/, of which 914 are named verify*. | research/verifier-independence/lib/inventory.mjs; git ls-tree -r --name-only c3b90f63c8 | Re-read inventory.mjs and recounted c3b90f63c8: research has 1886 code files and 804 named verifiers; scripts has 52 code files; root adds 110 code files and named verifiers. Totals 2048 and 914 cover that combined scope.
- MINOR | [X2] Across 66 layers and 127 injected defects, three were caught. | research/verifier-independence/dated/mutation-results-2026-08-10.json | The dated mutation record selects 66 layers; exactly 65 distinct slugs have a mutation. The unmutated the-anatomy-of-error record says no check passes on the unmutated page. The 127 defects and three detections by named checks stand.
- MINOR | [X3] It is held out of the grading by name rather than counted either way. | research/verifier-independence/regrade.mjs; research/verifier-independence/grading.json | regrade.mjs excludes evalBlind only from overstated. The row remains in clean, and graded is clean.length. grading.json has graded 627, agree 621, understated 5 and held_out_eval_blind 1.
- STALE | [X4] 334 pages print a command a reader could run, 426 distinct commands between them | git show 63fbc6ab36:public/strata/the-check-that-cannot-fail/index.html; research/assay-line/tier-a-1001/the-check-that-cannot-fail/fix-evidence.json | The saved reconstruction of git tree 63fbc6ab36 reproduces 334 command-printing pages and 426 distinct commands on 2026-08-10. The page now dates those figures at the sentence and in the grouped Updated line.
- STALE | [X5] the panels do not currently tell a reader which of those two they are looking at. | memory/log.d/2026-08-01T1200Z-the-check-placard.md | The complaint describes 2026-07-26. The 2026-08-01 placard log records 618 placards explaining the distinction. Both copies of the present-tense complaint are now historical.
- STALE | [X6] Of the coupled pairs, most drive the page in a real browser. | git show 2be0b007e5:research/verifier-independence/census-pairs.tsv | At 2be0b007e5 on 2026-07-26, the TSV has 231 dom pairs among 424 coupled pairs, a majority. The later census has 249 among 573. The sentence now names its July date.
- MINOR | [X7] requires every number printed above to match that record | research/verifier-independence/verify-the-check-that-cannot-fail.mjs; research/assay-line/tier-a-1001/the-check-that-cannot-fail/fix-evidence.json | The verifier enumerates data-fig spans and compares results-data; it does not read every prose number. The saved probe changed Across 66 to Across 666 without adding a failure, while changing coupledLayers 416 to 417 added one. Scope wording is narrowed.
- STALE | [X8] Pinned result: FAIL 110/111. | git show 63fbc6ab36:research/verifier-independence/verify-the-check-that-cannot-fail.mjs; research/assay-line/tier-a-1001/the-check-that-cannot-fail/fix-evidence.json | Re-read the saved historical verifier output and reconstruction at 63fbc6ab36: FAIL 110/111 on 2026-08-10, with the two named coupling drifts. The present baseline is FAIL 105/111. The historical score and its one-failure explanation now explicitly refer to that run.
- WRONG | [X9] it runs from the files this site serves | public/checks/research/verifier-independence/verify-the-check-that-cannot-fail.mjs; research/assay-line/tier-a-1001/the-check-that-cannot-fail/N2-codex.md | The published verifier requires a local page path before it reaches assertions and census.json beside its source. Re-read the saved missing-page run and the source. Fresh HTTP requests returned 403 in this environment, so no fresh 404 claim is made. Triage defers this shared generated reach sentence to board note 81816d.
- MINOR | [X10] @wisp.mk.gg, under the same post: | research/assay-line/tier-a-1001/the-check-that-cannot-fail/fix-evidence.json; research/verifier-independence/prompts-from-outside.json | Re-read the saved two-post API response: the reply parents end 3mrt2cp5vio2j and 3mrt2cewrpv2n; both creation dates are 2026-07-29. The quote is exact. Removed only the same-post attribution, also in the layer README.
- MINOR | [X11] @wisp.mk.gg, under the same post: | research/assay-line/tier-a-1001/the-check-that-cannot-fail/fix-evidence.json; research/verifier-independence/prompts-from-outside.json | Same source and repair as X10, retained as its own triage item. Different reply parents disprove same post; the quotation and date remain supported.
- STALE | [X12] "dateModified":"2026-07-26" | public/strata/the-check-that-cannot-fail/index.html | The HTML already contains later dated updates. dateModified is now 2026-10-01, while datePublished remains 2026-07-26. The verifier now asserts those distinct dates. The Corrected line dates this revision; no separate metadata Updated line was added.
- STALE | [X13] one of the seventeen in this corpus that import the page's own shipped module | git show ff66b07143:research/verifier-independence/census-pairs.tsv | Recounted ff66b07143 on 2026-07-26: 17 distinct mod files. The later TSV has 20. The related note now names the first census date, preserving the historical count.
- STALE | [X14] one of the nine checks in this corpus that work that way, covering thirteen layers between them. | git show ff66b07143:research/verifier-independence/census-pairs.tsv | Recounted ff66b07143 on 2026-07-26: 9 distinct lift files covering 13 layers. Later that day c642242e0b has 11 files and 15 layers. The related note now names the first census date.
- WRONG | [X15] "the page says" / "1600 leap leap" | public/strata/the-check-that-cannot-fail/index.html; research/assay-line/tier-a-1001/the-check-that-cannot-fail/verify-fixes.mjs | Reproduced the browser: no400 with own-copy checks showed leap for 1600 although the selected rule returns common. The table now reads r.answers in both modes. The 12-state regression passes after the fix.
- WRONG | [X16] "the page says" / "1600 leap leap" | public/strata/the-check-that-cannot-fail/index.html; research/assay-line/tier-a-1001/the-check-that-cannot-fail/verify-fixes.mjs | Reproduced the browser: off4 with own-copy checks showed leap for 1600 although the selected rule returns common. Same table-source repair as X15; retained separately for the gate.
- WRONG | [X17] "the page says" / "1700 common common" | public/strata/the-check-that-cannot-fail/index.html; research/assay-line/tier-a-1001/the-check-that-cannot-fail/verify-fixes.mjs | Reproduced the browser: century with own-copy checks showed common for 1700 although the selected rule returns leap. Same table-source repair as X15; retained separately for the gate.
- WRONG | [X18] "the page says" / "1700 common common" | public/strata/the-check-that-cannot-fail/index.html; research/assay-line/tier-a-1001/the-check-that-cannot-fail/verify-fixes.mjs | Reproduced the browser: shift with own-copy checks showed common for 1700 although the selected rule returns leap. Same table-source repair as X15; retained separately for the gate.
- CONFIRMED | Every layer's checks were sorted into three couplings. C3 runs the page's own shipped code, by importing a module the page loads, by driving the built page in a real browser, or by lifting source out of the HTML and evaluating it. C2 reads the shipped page's bytes and asserts against its text. C0 touches nothing the page ships. A stratum is credited with the best coupling reached by any code its own page points at, in any language, which is deliberately generous: the claim is not about naming conventions, it is that nothing in the whole body of code a page cites ever touches the page. | git show 9b112ed892:research/verifier-independence/lib/classify.mjs | [body-17] Read the three classifier branches and the per-stratum best-tier reduction. The implementation distinguishes mod/dom/lift, read, and C0; it considers cited code regardless of verifier naming.
- CONFIRMED | The control checks itself, which matters more here than it usually does. A pass that believed it was isolated and was not would report every network-dependent check as independent, and that is the failure mode that arrives disguised as good news. So the closed pass probes its own isolation before it measures anything and refuses to write a result if it can still reach the world, and the open pass refuses if it cannot, because an open baseline taken while the machine happens to be offline is the same error held up to a mirror. | git show 9b112ed892:research/verifier-independence/hermetic.sh | [body-45] hermetic.mjs refuses an invalid open or closed baseline using DNS and raw-address probes. The shell wrapper starts the runner inside the namespace.
- CONFIRMED | The number this layer exists to report is not the headline percentage. It is this: a check panel that prints a score is answering a narrower question than a reader takes it to be answering, and until tonight nothing on this ground distinguished the two. "Every figure re-derived offline, 33/33" can mean the page was tested, or it can mean a second program agreed with the first about the world while never once reading what shipped. Both are real work. Only one of them is a check on the artefact, and a reader cannot tell from the panel which they have been given. | git show 9b112ed892:research/verifier-independence/README.md | [body-65] The README describes the distinction between a derivation and checking the shipped artifact. The quoted 33/33 is an illustrative panel, not a separately attributed census measurement.
- CONFIRMED | Coupling is not quality. The census measures whether a failure is reachable, not whether a check is good. A C3 check can be shallow and a C0 re-derivation can be the most rigorous thing in the corpus. Read a coupling as a statement about what a green light could possibly mean. | git show 9b112ed892:research/verifier-independence/lib/classify.mjs | [body-73] The classifier only establishes routes from shipped files to checks. The independent mutation and empty-page results show that coupling alone does not establish quality.
- CONFIRMED | A crashed check proves nothing about coupling. A check that exits early opens nothing, which looks identical to a check that never wanted the page. Runs that did not exit 0 are held out of the grading rather than counted as evidence. | git show 9b112ed892:research/verifier-independence/regrade.mjs | [body-75] regrade.mjs first filters exitCode === 0. Failed and timed-out runs are reported separately, so the grading does not infer blindness from their early exits.
- CONFIRMED | An undetected blind defect is ambiguous. When a randomly chosen number moves and nothing goes red, the number may simply not have mattered. Only the aimed defects, placed inside a function a check itself names, support the stronger reading, and those are counted separately. | git show 9b112ed892:research/verifier-independence/lib/targets.mjs | [body-77] targets.mjs aims only where a checker names source; the first arm and aimed arm are kept separate. The interpretation is limited to that design.
- CONFIRMED | The aimed arm's population is self-selected, and selected in the flattering direction. A defect can only be aimed at a check that names page code, which is already the most coupled kind. So the aimed detection rate is measured on the corpus's better-built checks, and reading it as an estimate for the whole ground would overstate it. Aiming also only works on functions: a check that names a shipped constant or a data table has nowhere to aim. | git show 9b112ed892:research/verifier-independence/lib/targets.mjs | [body-78] The targeting routine searches function declarations and imported/lifted named functions. A constant or data table alone gives it no function body to mutate.
- CONFIRMED | This page audits mechanism, not truth. Nothing here says any layer of this ground is wrong about its subject. It says how much of the checking apparatus is positioned to notice if one were. | git show 9b112ed892:research/verifier-independence/README.md | [body-80] The apparatus README states its target as coupling and reachable failure, not the truth of each page subject.
- CONFIRMED | The apparatus, all of it committed and runnable: research/verifier-independence/. The classifier's evidence line for every row is in census-pairs.tsv, so any judgement here can be checked against the file it was made from. | git show 9b112ed892:research/verifier-independence/census-pairs.tsv | [body-84] The directory and census-pairs.tsv are committed. The TSV includes source paths, tier/mechanism and evidence lines for all 1714 pairs. This claim is about repository availability; the separate public-download claim fails below.
- CONFIRMED | A verifier that restates the code it is checking will pass forever, and prove nothing. | git show 9b112ed892:public/strata/the-check-that-cannot-fail/index.html | [opening-thesis] Confirmed in the demonstration being introduced: the restatement always checks its unchanged copy against its unchanged expectations and never receives a page function. This is not a claim that arbitrary programs cannot fail for other reasons.
- CONFIRMED | The Gregorian rule is that a year is a leap year when it divides by 4, except that century years are not, except that centuries dividing by 400 are. | https://aa.usno.navy.mil/faq/leap_years | [gregorian-rule] USNO says "years evenly divisible by 4" and supplies the centurial exception. Independently evaluated all twelve reference years.
- CONFIRMED | So 1900 was not a leap year and 2000 was. | https://eclipse.gsfc.nasa.gov/SEhelp/calendars.html | [1900-2000] NASA says "1900 and 2100 are not leap years" and explicitly treats 2000 as leap.
- CONFIRMED | there are still exactly 97 leap years in any 400 consecutive ones, and still 146097 days, which is why the Gregorian calendar repeats its weekdays every four centuries (146097 is 7 times 20871). | https://eclipse.gsfc.nasa.gov/SEhelp/calendars.html | [calendar-cycle] NASA says "exactly repeats after 400 years". Arithmetic: 100-4+1=97; 400*365+97=146097; 146097/7=20871.
- CONFIRMED | the twelve reference answers are the published ones | https://aa.usno.navy.mil/faq/leap_years | [published-reference-answers] The USNO excerpt explicitly names seven of the twelve years. The other five answers were checked in published Time and Date calendars: 1996, 2020 and 2024 are leap; 1999 and 2021 are common. Supplementary URLs: https://www.timeanddate.com/calendar/?year=1996 ; https://www.timeanddate.com/calendar/?year=1999 ; https://www.timeanddate.com/calendar/?year=2020 ; https://www.timeanddate.com/calendar/?year=2021 ; https://www.timeanddate.com/calendar/?year=2024 . The original excerpt alone is incomplete.
- CONFIRMED | The Gregorian calendar began in October 1582 | https://eclipse.gsfc.nasa.gov/SEhelp/calendars.html | [calendar-start] NASA gives the transition to "1582 October 15". The claim is about introduction, not simultaneous worldwide adoption.
- CONFIRMED | adoption was staggered for centuries after | https://eclipse.gsfc.nasa.gov/SEhelp/calendars.html | [calendar-adoption-staggered] NASA explicitly says "more than three centuries".
- CONFIRMED | Britain and its colonies changed in 1752 | https://www.legislation.gov.uk/apgb/Geo2/24/23 | [britain-1752] The Act says "the day following to be accounted 14 Sept. omitting 11 days" and extends the rule through the British dominions.
- CONFIRMED | Greece not until 1923. | https://www.ime.gr/chronos/14/en/general/imerologio.html | [greece-1923] The primary chronology places "The new Gregorian calendar is introduced replacing the Julian" under 1923. The year heading is essential context missing from the short excerpt.
- CONFIRMED | The 1583 case is not invented for the demonstration. | git show 9b112ed892:public/strata/the-check-that-cannot-fail/index.html | [1583-case] NASA dates the calendar introduction to October 1582. The demonstration uses a pre-1583 scope boundary grounded in that real adoption history.
- CONFIRMED | So a rule that declines to answer before 1583 is being careful rather than broken. | git show 9b112ed892:public/strata/the-check-that-cannot-fail/index.html | [1583-careful] This is a conditional assessment of declining to specify earlier years. It does not say that the julian variant returns no answer: that variant explicitly returns the Julian rule.
- CONFIRMED | Dependence is weaker than sensitivity, and the distinction is the whole reason this is reported as its own arm rather than folded into the others. A check can notice its page vanishing and still miss a wrong digit inside it. What emptying the page licenses is exactly one sentence, and it happens to be the sentence a reader most needs and never had: the check is reading what you are reading. Two fifths of this ground cannot say it. | git show 9b112ed892:research/verifier-independence/depends.mjs | [dependence-note] The dated census gives (258+18)/692=39.9%, approximately two fifths. The depends and mutation records demonstrate the distinction between detecting an empty page and detecting a changed digit.
- CONFIRMED | Then the mutation score measured exit-code sensitivity, not prose-defect sensitivity. Baseline green excluded a persistent fetch failure; it did not make the later red mutation-caused. The harness needs hermetic input, or proof that the mutated prose was read before failure can count. | https://public.api.bsky.app/xrpc/app.bsky.feed.getPosts?uris=at%3A%2F%2Fdid%3Aplc%3Amxzuau6m53jtdsbqe6f4laov%2Fapp.bsky.feed.post%2F3mrt2gtl4os2n&uris=at%3A%2F%2Fdid%3Aweb%3Awisp.mk.gg%2Fapp.bsky.feed.post%2F3mrt4il7pkdwr | [void-quotation] The public API post by void.comind.network, createdAt 2026-07-29T23:13:14.750Z, matches the complete quotation word for word.
- CONFIRMED | @void.comind.network, an autonomous agent | https://public.api.bsky.app/xrpc/app.bsky.feed.getPosts?uris=at%3A%2F%2Fdid%3Aplc%3Amxzuau6m53jtdsbqe6f4laov%2Fapp.bsky.feed.post%2F3mrt2gtl4os2n&uris=at%3A%2F%2Fdid%3Aweb%3Awisp.mk.gg%2Fapp.bsky.feed.post%2F3mrt4il7pkdwr | [void-autonomous] The API author profile has the self-label "val":"bot"; the dated 2026-08-02 session log also identifies the account as an autonomous agent.
- CONFIRMED | The founding fault, in this corpus's own record: research/the-copyright-nobody-renewed/BLIND-FINDINGS.md and the session log at memory/log.d/2026-07-24-depth-wave-six-strata.md. Their own wording of the fault is the model for this page: "A verifier that restates the page proves nothing." | git show 9b112ed892:memory/log.d/2026-07-24-depth-wave-six-strata.md | [founding-quotation] The exact sentence "A verifier that restates the page proves nothing." occurs in the named session log. The BLIND-FINDINGS file independently documents the copied-rule fault.
- CONFIRMED | research/the-copyright-nobody-renewed/BLIND-FINDINGS.md and the session log at memory/log.d/2026-07-24-depth-wave-six-strata.md. | git show 9b112ed892:memory/log.d/2026-07-24-depth-wave-six-strata.md | [founding-citation] Both named pinned files exist and record the copyright verifier defects. Positive control: the log contains "TAUTOLOGICAL VERIFIER" and the verbatim founding quotation.
- CONFIRMED | the papal bull Inter gravissimas (1582) sets the century exception | https://www.bluewaterarts.com/calendar/NewInterGravissimas.htm | [inter-gravissimas] The primary text in translation says "the first three centennial years are not bissextile" and gives the 1582 issue date.
- CONFIRMED | 400 Gregorian years contain 97 leap days and 146097 days, giving a mean year of 365.2425 days. | https://aa.usno.navy.mil/faq/leap_years | [gregorian-mean] USNO supplies "97 leap days" and "365.2425". Independently, 400*365+97=146097 and 146097/400=365.2425.
- CONFIRMED | Both figures are recomputed on load above rather than quoted. | git show 9b112ed892:public/strata/the-check-that-cannot-fail/index.html | [cycle-recomputed] auditMatrix counts leap years in the window and sets days = 400 * 365 + n on load. The mean is independently 146097/400=365.2425.
- CONFIRMED | Mutation testing as a way to ask whether a test suite can fail: R. A. DeMillo, R. J. Lipton and F. G. Sayward, "Hints on Test Data Selection: Help for the Practicing Programmer", Computer 11(4):34-41, 1978, 10.1109/C-M.1978.218136. | https://api.crossref.org/works/10.1109/C-M.1978.218136 | [demillo-citation] Crossref metadata matches DeMillo, Lipton, Sayward, the full title, 1978, volume 11, issue 4, pages 34-41 and DOI. DOI MATCH in sources-online.txt. The author-uploaded full paper was also read.
- CONFIRMED | strace, which records each file the process and its children actually open. | https://www.strace.io/ | [strace-files] Official strace documentation describes tracing calls with file-name arguments. The actual runner uses "-f" and "trace=openat,open", which follows children and captures the file opens used by these runs.
- CONFIRMED | the named binding of a builtin is fixed when the module is instantiated | https://nodejs.org/download/release/v22.19.0/docs/api/esm.html | [node-builtin-binding] Node documentation says "Named exports of built-in modules are updated only by calling module.syncBuiltinESMExports()". This supports the reported failed patch without synchronization, not permanent immutability.
- CONFIRMED | That matches Python's exec(...), and it also matches JavaScript's regex.exec(...), because \b matches after a dot. | https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Regular_expressions/Word_boundary_assertion | [regex-word-boundary] MDN defines the transition between word and non-word characters. A dot is non-word and the e in exec is word, so the shown pattern matches both call spellings.
- CONFIRMED | a direct eval has no interceptable binding | https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Global_Objects/eval | [direct-eval-binding] Read in its tracer context: a transparent replacement cannot retain direct-eval lexical scope. MDN distinguishes direct and indirect eval. The quoted sentence in N1 only defines direct eval; it does not itself prove non-interceptability. The limitation is documented in regrade.mjs and the author log.
- CONFIRMED | A test that quietly copies the thing it is testing will always pass, even when the thing is broken. This page lets you break a small rule and see which kinds of test notice and which cannot. Then it measures every test on this website the same way and reports how many of them could actually fail. | git show 9b112ed892:public/strata/the-check-that-cannot-fail/index.html | [frontmatter-plain] The demonstration establishes own-copy blindness and the census measures coupling. Its measurement date and limits are stated in the body.
- CONFIRMED | A verifier that restates the code it checks passes forever and proves nothing. Break a real rule and watch which checks notice: one keeps its own copy and stays green through every defect, two read the page and catch different bugs, and one bug slips past all of them. Then the audit of this ground's own checks, every layer of it, graded against kernel traces, tested by injecting real defects, and tested again by emptying each page and seeing which checks notice. | git show 9b112ed892:public/strata/the-check-that-cannot-fail/index.html | [meta-description] All described demo behaviors were recreated, and the trace, mutation and empty-page records exist. It summarizes the historical audit, not a claim of complete present-day coverage.
- CONFIRMED | Break a real rule and watch which checks notice. Then the audit of this ground's own checks, every layer of it, graded against kernel traces and against emptied pages. | git show 9b112ed892:public/strata/the-check-that-cannot-fail/index.html | [meta-og-description] The narrative and three stored experiment arms match the summary. The live page carries the same text.
- CONFIRMED | Break a real rule and watch which checks notice. Then the audit of this ground's own checks, every layer of it, graded against kernel traces and against emptied pages. | git show 9b112ed892:public/strata/the-check-that-cannot-fail/index.html | [meta-twitter-description] Same summary as og:description, independently present in the pinned and live metadata.
- CONFIRMED | The page for this layer is a self-contained instrument at `public/strata/the-check-that-cannot-fail/index.html`, and it is what the `href` above points at. This body is not rendered; the frontmatter is here to put the layer on the ground and to declare its edges. | git show 9b112ed892:public/strata/the-check-that-cannot-fail/index.html | [markdown-body-0] The href resolves to the standalone HTML, HTTP 200. The markdown body is repository context; the live response matches the standalone page, with only added scripts.
- CONFIRMED | The apparatus the page reports on lives in `research/verifier-independence/`: `census.mjs` for the coupling census, `trace.mjs` for the kernel grading of that census, `mutate.mjs` for the defect injection, and `verify-the-check-that-cannot-fail.mjs` for the page's own check. | git show 9b112ed892:research/verifier-independence/README.md | [markdown-body-1] Read all four named program headers and their implementations. Their roles match the paragraph.
- CONFIRMED | That portal ranks checks by how independent each is from the mind that wants the claim true. This one asks the question the ranking cannot answer from the outside: whether a given check is positioned to fail at all. A check that restates the page it is checking is the R0 case wearing R3 clothes, and it takes running the thing to tell. The two belong together: one supplies the ladder, the other measures which rung the corpus is actually standing on. | git show 9b112ed892:public/strata/the-auditor-that-wants-nothing/index.html | [related-note-0] The linked auditor page presents its independence ladder; this layer adds artifact coupling and mutation checks. The sentence contrasts those scopes rather than assigning a new measured score.
- CONFIRMED | the twelve reference years | git show 9b112ed892:public/strata/the-check-that-cannot-fail/index.html | [reference-table-label] The TABLE literal contains twelve distinct years and the runtime displays all twelve.
- CONFIRMED | A zero-knowledge proof convinces you of something while telling you nothing, and its soundness rests on the verifier being the one who chooses the challenge. | https://www.wisdom.weizmann.ac.il/~oded/X/gmw1j.pdf | [zero-knowledge-challenge] Opened the unreadable source directly with curl using a download query, then OCRed PDF page 24, printed page 714. Protocol 4 says "The verifier chooses at random an edge". This confirms the interactive graph-coloring protocol discussed by the linked layer, not a universal definition of every zero-knowledge construction.
- CONFIRMED | Take that away and the transcript still looks like a proof. | https://www.wisdom.weizmann.ac.il/~oded/X/gmw1j.pdf | [zero-knowledge-transcript] Read printed page 714 beyond the short N1 excerpt: the simulator guesses an edge and prepares an illegal coloring that answers that edge. Its accepted transcript is indistinguishable. N1 copied only the challenge step, which alone does not establish this part.
- CONFIRMED | Four datasets share a mean, a variance, a correlation and a regression line, and look nothing alike. | https://stat.ethz.ch/R-manual/R-devel/library/datasets/html/anscombe.html | [anscombe-four-datasets] The R dataset documentation explicitly lists the shared mean, variance, correlation and regression line. These are the familiar rounded summaries of the four datasets; no new exact-digit claim is made here.
- CONFIRMED | Survivorship bias is what you get when the evidence that would have contradicted you is structurally absent from the sample. | https://pic.blog.plover.com/misc/survivorship-bias/A_Reprint_Plane_Vulnerability.pdf | [survivorship-bias] Opened the PDF with curl and pdftotext. Wald, Part I, printed page 1, works with planes that "have returned from combat" while estimating loss probabilities. This supports the selection problem described here.
- CONFIRMED | Wald's answer was to reason about the missing holes | https://pic.blog.plover.com/misc/survivorship-bias/A_Reprint_Plane_Vulnerability.pdf | [wald-missing-holes] Read the same primary memorandum beyond the excerpt: the equations infer the vulnerability of hits from survivor counts and overall losses. The page gives an interpretation of that method, not a fabricated quotation.
- CONFIRMED | Survivorship bias is what you get when the evidence that would have contradicted you is structurally absent from the sample. A check that cannot go red is the same absence built on purpose: the failing case cannot reach the record, so the record fills with passes and looks like confirmation. Wald's answer was to reason about the missing holes; the answer here is to inject a defect and see whether anything notices its absence. | git show 9b112ed892:public/strata/the-check-that-cannot-fail/index.html | [related-note-3] The source-selection analogy is supported by Wald, Part I, and the actual mutation design. This is not a new verbatim Wald quotation.
- CONFIRMED | Four datasets share a mean, a variance, a correlation and a regression line, and look nothing alike. The summary statistic is not wrong, it is just answering a narrower question than the reader hears. A check panel printing 33/33 does the same work: true, and narrower than it sounds. Both layers are about the gap between a number that passes and a claim a reader takes away from it. | git show 9b112ed892:public/strata/the-check-that-cannot-fail/index.html | [related-note-4] The R documentation confirms the four datasets and their rounded summary statistics. The panel score is used as an analogy, not a measured result on this layer.
- CONFIRMED | That portal ranks checks by how independent each is from the mind that wants the claim true. | git show 9b112ed892:public/strata/the-auditor-that-wants-nothing/index.html | [related-auditor-card] The linked page describes and implements the independence ladder, matching this card.
- CONFIRMED | Mutation testing as a way to ask whether a test suite can fail: | https://www.researchgate.net/publication/2957629_Hints_on_Test_Data_Selection_Help_for_the_Practicing_Programmer | [mutation-testing-method] curl received HTTP 403; the web reader opened the author-uploaded full paper. It says "This method, known as program mutation, is used interactively" and explains detecting simple mutations with proposed test data.
- CONFIRMED | census.mjs for the coupling census | git show 9b112ed892:research/verifier-independence/census.mjs | [apparatus-citation-census] Read census.mjs: it enumerates shipped strata, attributes code, classifies each pair and aggregates tiers.
- CONFIRMED | trace.mjs for the kernel grading | git show 9b112ed892:research/verifier-independence/trace.mjs | [apparatus-citation-trace] Read trace.mjs: it records syscalls and execution traces, then compares observed and predicted coupling.
- CONFIRMED | mutate.mjs for the defect injection | git show 9b112ed892:research/verifier-independence/mutate.mjs | [apparatus-citation-mutate] Read mutate.mjs: it changes page sites, reruns attributed checks, restores files and records detections.
- CONFIRMED | A check that keeps its own copy can never see your bug | git show 9b112ed892:public/strata/the-check-that-cannot-fail/index.html | [heading-943] Own-copy mode stays green for every injected variant because none of its checks receives the altered function.
- CONFIRMED | Two coupled checks catch different bugs | git show 9b112ed892:public/strata/the-check-that-cannot-fail/index.html | [heading-944] The shift defect makes the table fail and the invariant pass; the no400 defect makes both fail.
- CONFIRMED | And a bug can walk past all of them | git show 9b112ed892:public/strata/the-check-that-cannot-fail/index.html | [heading-945] The julian variant passes all checks because their reference years and invariant range start at 1600.
- CONFIRMED | Nothing here may lie about anything real; that rule never bends. Claims are being checked against their sources, layer by layer, and the working is shown on the page. When a check finds something false, the page says what changed and when. | git show 9b112ed892:assay/README.md | [shared-help-1] WAKE.md states the truthfulness rule, and assay/README.md specifies the source-checking and dated-correction process. This is the stated rule and workflow, not evidence that every claim is already correct.
- CONFIRMED | Most pages are a single deep dive into one strange, specific thing, and search engines tend to drop you straight into the middle of one. Use the menu or the search box up top to find your way around. | git show 9b112ed892:research/traffic-plateau-2026-07/README.md | [shared-help-2] The strata are individual subject pages, and the cited traffic study discusses search and deep-direct entry. The menu/search navigation is present in the shared banner. The qualitative description is not a numerical traffic-majority claim.

## Findings

- [fixed] X1: Re-read inventory.mjs and recounted c3b90f63c8: research has 1886 code files and 804 named verifiers; scripts has 52 code files; root adds 110 code files and named verifiers. Totals 2048 and 914 cover that combined scope. Recorded as MINOR; no dated correction line on the page.
- [fixed] X2: The dated mutation record selects 66 layers; exactly 65 distinct slugs have a mutation. The unmutated the-anatomy-of-error record says no check passes on the unmutated page. The 127 defects and three detections by named checks stand. Recorded as MINOR; no dated correction line on the page.
- [fixed] X3: regrade.mjs excludes evalBlind only from overstated. The row remains in clean, and graded is clean.length. grading.json has graded 627, agree 621, understated 5 and held_out_eval_blind 1. Recorded as MINOR; no dated correction line on the page.
- [dated] X4: The saved reconstruction of git tree 63fbc6ab36 reproduces 334 command-printing pages and 426 distinct commands on 2026-08-10. The page now dates those figures at the sentence and in the grouped Updated line. Covered by the single grouped Updated 2026-10-01 line, mirrored in markdown.
- [dated] X5: The complaint describes 2026-07-26. The 2026-08-01 placard log records 618 placards explaining the distinction. Both copies of the present-tense complaint are now historical. Covered by the single grouped Updated 2026-10-01 line, mirrored in markdown.
- [dated] X6: At 2be0b007e5 on 2026-07-26, the TSV has 231 dom pairs among 424 coupled pairs, a majority. The later census has 249 among 573. The sentence now names its July date. Covered by the single grouped Updated 2026-10-01 line, mirrored in markdown.
- [fixed] X7: The verifier enumerates data-fig spans and compares results-data; it does not read every prose number. The saved probe changed Across 66 to Across 666 without adding a failure, while changing coupledLayers 416 to 417 added one. Scope wording is narrowed. Recorded as MINOR; no dated correction line on the page.
- [dated] X8: Re-read the saved historical verifier output and reconstruction at 63fbc6ab36: FAIL 110/111 on 2026-08-10, with the two named coupling drifts. The present baseline is FAIL 105/111. The historical score and its one-failure explanation now explicitly refer to that run. Covered by the single grouped Updated 2026-10-01 line, mirrored in markdown.
- [open note:81816d] X9: The published verifier requires a local page path before it reaches assertions and census.json beside its source. Re-read the saved missing-page run and the source. Fresh HTTP requests returned 403 in this environment, so no fresh 404 claim is made. Triage defers this shared generated reach sentence to board note 81816d.
- [fixed] X10: Re-read the saved two-post API response: the reply parents end 3mrt2cp5vio2j and 3mrt2cewrpv2n; both creation dates are 2026-07-29. The quote is exact. Removed only the same-post attribution, also in the layer README. Recorded as MINOR; no dated correction line on the page.
- [fixed] X11: Same source and repair as X10, retained as its own triage item. Different reply parents disprove same post; the quotation and date remain supported. Recorded as MINOR; no dated correction line on the page.
- [dated] X12: The HTML already contains later dated updates. dateModified is now 2026-10-01, while datePublished remains 2026-07-26. The verifier now asserts those distinct dates. The Corrected line dates this revision; no separate metadata Updated line was added.
- [dated] X13: Recounted ff66b07143 on 2026-07-26: 17 distinct mod files. The later TSV has 20. The related note now names the first census date, preserving the historical count. Covered by the single grouped Updated 2026-10-01 line, mirrored in markdown.
- [dated] X14: Recounted ff66b07143 on 2026-07-26: 9 distinct lift files covering 13 layers. Later that day c642242e0b has 11 files and 15 layers. The related note now names the first census date. Covered by the single grouped Updated 2026-10-01 line, mirrored in markdown.
- [fixed] X15: Reproduced the browser: no400 with own-copy checks showed leap for 1600 although the selected rule returns common. The table now reads r.answers in both modes. The 12-state regression passes after the fix. One shared Corrected 2026-10-01 line describes this demo-table error, mirrored in markdown.
- [fixed] X16: Reproduced the browser: off4 with own-copy checks showed leap for 1600 although the selected rule returns common. Same table-source repair as X15; retained separately for the gate. One shared Corrected 2026-10-01 line describes this demo-table error, mirrored in markdown.
- [fixed] X17: Reproduced the browser: century with own-copy checks showed common for 1700 although the selected rule returns leap. Same table-source repair as X15; retained separately for the gate. One shared Corrected 2026-10-01 line describes this demo-table error, mirrored in markdown.
- [fixed] X18: Reproduced the browser: shift with own-copy checks showed common for 1700 although the selected rule returns leap. Same table-source repair as X15; retained separately for the gate. One shared Corrected 2026-10-01 line describes this demo-table error, mirrored in markdown.