---
slug: every-plus-or-minus
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: 131
confirmed: 99
wrong: 20            # 6 WRONG + 14 MINOR
unverifiable: 1
observed: 11
stale: 0
---

This was the claude-1018b decide-and-fix pass of the assay line on Every Plus-or-Minus Is a Decision, dated 2026-10-18. It carries forward the 131 claim lines of research/assay-line/claude-1018b/every-plus-or-minus/N2-claude.md, the checker's record against the N1 source record at 0a15703d9e. The fixer re-read the evidence for all six WRONG findings before deciding (a numpy recomputation of the 1024 correlation corners, the verifier's imports, a run of cross-check.py, the half-life census's published uncertainties, and facts.md K1 and L11), confirmed all six, and fixed them on the page, the relates notes, the engine (both byte-identical copies) and the verifier, with one dated Corrected line on the page and its mirror in the markdown. The fourteen MINOR findings are tier B and sit outside the sentences being fixed, so they are listed here and declined, not edited. The MINOR lines are counted under wrong.

## Claims

- CONFIRMED | The 1963 recommended value of the electron charge appears twice in this archive, with the same number and two different signs after it, because two layers read two different documents. | research/the-charge-that-crept/data.mjs and research/error-bar-coverage/data/anchor-rows.json | [dek-two-layers-same-number] both ship 4.80298 for 1963; the first with err_esu 0.00020 tagged 3sd, the second with 0.00006 under a review recorded as "standard deviation"
- CONFIRMED | transcribed it from Cohen and DuMond's 1965 review in Reviews of Modern Physics | research/error-bar-coverage/data/anchor-rows.json | [anchor-ships-0.00006] reviews[6] cites Cohen and DuMond, Rev. Mod. Phys. 37, 537 (1965) for the 1963 row
- UNVERIFIABLE | where it is a standard deviation | tried https://journals.aps.org/rmp/abstract/10.1103/RevModPhys.37.537 (403 on 2026-10-19), OpenAlex (closed access, abstract only, no uncertainty convention), Internet Archive advanced search and scrape (0 hits; control: a generic RMP 1965 query returns 25) | [1965-review-sd] the page itself says this rests on the anchor layer transcription and not on a reading here; indirect support only: 1973 Table 34.2 prints the 1963 e as 1.60210(2), 12 ppm
- CONFIRMED | Cohen and DuMond's 1965 review | https://api.crossref.org/works/10.1103/RevModPhys.37.537 | [crossref-1965] Crossref title "Our Knowledge of the Fundamental Constants of Physics and Chemistry in 1965", Cohen and DuMond
- CONFIRMED | In 1963 a committee recommended a value for the charge on the electron. | https://www.govinfo.gov/content/pkg/GOVPUB-C13-dc85722872ab7945773bde84551dd8af/pdf/GOVPUB-C13-dc85722872ab7945773bde84551dd8af.pdf (PDF p. 1, rendered at 300 dpi) | [committee-1963] card headed "GENERAL PHYSICAL CONSTANTS RECOMMENDED BY NAS-NRC Adopted by NBS", "Reprinted from NBS Technical News Bulletin, Oct. 1963"
- CONFIRMED | read the contemporaneous sheet from the National Bureau of Standards, which prints a wider figure | research/the-charge-that-crept/facts.md | [charge-layer-reads-nbs-sheet] E3: the sheet's ±0.00020 is a three-standard-deviation error limit; the card itself (PDF p. 1) prints 20 under Est.* error limit
- CONFIRMED | Based on 3 std. dev. | https://www.govinfo.gov/content/pkg/GOVPUB-C13-dc85722872ab7945773bde84551dd8af/pdf/GOVPUB-C13-dc85722872ab7945773bde84551dd8af.pdf (PDF p. 1 and p. 2, rendered) | [nbs-3-std-dev-quote] footnote reads "*Based on 3 std. dev., applies to last digits in preceding col."
- CONFIRMED | The card is NBS Misc. Publ. 253, and its uncertainty column is headed Est. error limit over a footnote reading, in full, Based on 3 std. dev., applies to last digits in preceding col. | https://www.govinfo.gov/content/pkg/GOVPUB-C13-dc85722872ab7945773bde84551dd8af/pdf/GOVPUB-C13-dc85722872ab7945773bde84551dd8af.pdf (PDF p. 1 and p. 2, rendered) | [card-misc-253-header-and-footnote] "NBS Misc. Publ. 253" on the card face; column headed "Est.* error limit"; footnote text matches the page verbatim
- CONFIRMED | They agree on the number, 4.80298 in units of ten to the minus ten electrostatic units | https://www.govinfo.gov/content/pkg/GOVPUB-C13-dc85722872ab7945773bde84551dd8af/pdf/GOVPUB-C13-dc85722872ab7945773bde84551dd8af.pdf (PDF p. 1, rendered) | [card-rows-e-esu] Elementary charge row: 1.60210, 7, 10^-19 C and 4.80298, 20, 10^-10 electrostatic
- CONFIRMED | The adjustment's own uncertainty is printed in the comparison table of the 1973 adjustment as 1.60210(2) in units of ten to the minus nineteen coulombs, 12 ppm | https://physics.nist.gov/cuu/pdf/1973JPCRD.pdf (p. 724, Table 34.2, rendered) | [1973-table342-1963-row] "1.60210(2) 12" in the column "Value, 1963 adjustment, and ppm uncertainty"; footnote b gives units of 10^-19 C
- CONFIRMED | and rounds to the 0.00006 it ships. | research/every-plus-or-minus/facts.md | [0.00006-rounds] 12 ppm of 4.80298 is 0.0000576, which rounds to 0.00006 (recomputed)
- CONFIRMED | taking instead the 0.00006 the anchor layer ships, it is 3.33. | research/every-plus-or-minus/facts.md | [ratio-3.33] 0.00020 / 0.00006 = 3.33 (recomputed)
- CONFIRMED | the honest statement is that the printed limit is about three and a half times the adjustment's standard deviation | research/every-plus-or-minus/facts.md | [three-and-a-half-times] 0.00020 / 0.0000576 = 3.47 (recomputed); about three and a half is a fair statement
- CONFIRMED | The coulomb column of the same card shows the same thing and not more: a sigma of 2 in the last digits against a printed limit of 7, a ratio of 3.5 that three-times-then-round-up cannot produce on its own, since three twos are six. | https://www.govinfo.gov/content/pkg/GOVPUB-C13-dc85722872ab7945773bde84551dd8af/pdf/GOVPUB-C13-dc85722872ab7945773bde84551dd8af.pdf and https://physics.nist.gov/cuu/pdf/1973JPCRD.pdf p. 724 | [coulomb-column] card prints 7 for e in coulombs; the 2 is the 1973 table sigma, not the card; 12 ppm gives sigma 1.92, three of which is 5.77, rounding up to 6, so 7 is not three-times-then-round-up
- CONFIRMED | The elementary charge has been exact since the 2019 revision of the SI, so we know what the committee was aiming at. | https://en.wikipedia.org/wiki/2019_revision_of_the_SI | [exact-since-2019] "The elementary charge e is exactly 1.602176634×10−19 coulomb (C)", effective 20 May 2019
- MINOR | Against today's exact value the same row is 3.7 sigma out or 1.1 sigma out, depending which one you opened. | research/the-charge-that-crept/data.mjs; public/strata/the-charge-that-crept/index.html; public/strata/every-plus-or-minus/engine/findings.mjs | [dek-3.7-1.1-sigma] the 1.1 is the charge layer figure read as one sigma, which that layer says it is not ("the ±0.00020 is a THREE-σ error limit (often misread as 1σ)"); honouring its tag gives 3.37 sigma, the page own table row 2 ("discrepant") and the relates note ("the 1963 row" crosses the line when conventions are honoured). Which layer you open does not give 1.1 sigma; dropping the footnote does [declined: tier B, listed by the fixer]
- CONFIRMED | The committee is either three and three quarters of its own error away from the truth, or one and a tenth, depending on which of two honest layers of this archive you happened to open | research/every-plus-or-minus/facts.md | [three-and-three-quarters] 3.745 and 1.124 recomputed; "of its own error" here means the printed figure, so 1.1 of the printed limit is literally true
- MINOR | the entire difference is a presentation decision taken in 1963 and announced in a five-word footnote that then travelled separately from the number. | https://www.govinfo.gov/content/pkg/GOVPUB-C13-dc85722872ab7945773bde84551dd8af/pdf/GOVPUB-C13-dc85722872ab7945773bde84551dd8af.pdf | [five-word-footnote] the page itself quotes the footnote "in full" as "Based on 3 std. dev., applies to last digits in preceding col.", which is eleven words; only its first five carry the three-sigma statement [declined: tier B, listed by the fixer]
- CONFIRMED | Its 1971 card for the following adjustment carries the footnote Based on 1 std. dev.; applies to last digits in preceding column, and immediately beneath it: | https://www.govinfo.gov/content/pkg/GOVPUB-C13-a2fb26e45504787736a93a7f20d21bc7/pdf/GOVPUB-C13-a2fb26e45504787736a93a7f20d21bc7.pdf (PDF p. 2, rendered) | [sp344-1std-dev-footnote] "*Based on 1 std. dev.; applies to last digits in preceding column." with the ** footnote immediately beneath; values "taken from Taylor, Parker, and Langenberg, Rev. Mod. Phys. 41, p. 375 (1969)"
- CONFIRMED | These values may be in conflict with data available since the Taylor, Parker, Langenberg review. Pending a complete new readjustment of the constants, it would be prudent to multiply the above uncertainties by 3. | https://www.govinfo.gov/content/pkg/GOVPUB-C13-a2fb26e45504787736a93a7f20d21bc7/pdf/GOVPUB-C13-a2fb26e45504787736a93a7f20d21bc7.pdf (PDF p. 2, rendered) | [sp344-triple-quote] verbatim match of the ** footnote
- CONFIRMED | NBS Special Publication 344 (1971), footnote to the table of recommended values | https://www.govinfo.gov/content/pkg/GOVPUB-C13-a2fb26e45504787736a93a7f20d21bc7/pdf/GOVPUB-C13-a2fb26e45504787736a93a7f20d21bc7.pdf (PDF p. 1) | [sp344-citation] "NBS Spec. Publ. 344 Issued March 1971"
- MINOR | A standards body telling its readers, in print, to triple its own published uncertainties. | https://www.govinfo.gov/content/pkg/GOVPUB-C13-a2fb26e45504787736a93a7f20d21bc7/pdf/GOVPUB-C13-a2fb26e45504787736a93a7f20d21bc7.pdf (PDF p. 2, rendered) | [triple-own-uncertainties] the ** marker that carries "multiply the above uncertainties by 3" is printed against four of the card's 21 uncertainties only (N_A 40**, u 11**, m_p 11**, F 54**), so the card did not tell readers to triple its published uncertainties in general [declined: tier B, listed by the fixer]
- CONFIRMED | they disagree about the 1952 review: the JSON records its uncertainties as standard errors, the Python tags them as probable errors and converts them at 1.48. | research/error-bar-coverage/anchor_reproduction.py and research/error-bar-coverage/data/anchor-rows.json | [json-vs-python-1952] five Datum(1952, ..., "pe") rows and summarize(pe_to_sd=1.48); JSON reviews[4] errorMeasure "standard error"
- CONFIRMED | the JSON records its uncertainties as standard errors | research/error-bar-coverage/data/anchor-rows.json | [json-1952-standard-error] reviews[4] (1952) "errorMeasure": "standard error"
- CONFIRMED | keeps its forty rows in two artefacts, a JSON data file and a standalone Python reproduction | research/error-bar-coverage/anchor_reproduction.py | [two-artefacts] 40 Datum rows beside data/anchor-rows.json
- CONFIRMED | This one turns out to be harmless: at the published threshold both labellings give the same count, because none of the five 1952 rows sits near enough to the line for a factor of 1.48 to carry it across. | research/error-bar-coverage/data/anchor-rows.json (recomputed) | [1952-harmless] 1952 rows against 1973 targets: 1.04, 1.72, 1.95, 1.72, 7.43 sigma as printed and 0.70, 1.16, 1.32, 1.16, 5.02 converted; none crosses 2.33 either way
- CONFIRMED | A probable error is 0.6745 sigma. | https://en.wikipedia.org/wiki/Probable_error | [probable-error-0.6745] "For a normal distribution this is γ = 0.6745 × σ"
- CONFIRMED | estimated limits of uncertainty rather than the so-called probable errors | https://archive.org/download/sim_physical-review_1913-08_2_2/sim_physical-review_1913-08_2_2_djvu.txt | [millikan-limits-quote] "the numbers in the error column represent in the case of the first six numbers estimated limits of uncertainty rather than the so-called probable errors which would be much smaller"
- CONFIRMED | Millikan's 1913 figure | https://api.crossref.org/works/10.1103/PhysRev.2.109 | [millikan-cite] Millikan, On the Elementary Electrical Charge and the Avogadro Constant, Phys. Rev. 2, 109 (1913)
- CONFIRMED | Twentieth-century atomic weights ran on three incompatible scales, and the Avogadro constant moves with them. | https://en.wikipedia.org/wiki/Dalton_(unit) | [three-scales] chemists 1/16 of natural O, physicists 1/16 of oxygen-16, unified carbon-12 scale adopted 1960-61
- MINOR | The conversions are small, exact, and run in opposite directions. | https://en.wikipedia.org/wiki/Dalton_(unit); research/every-plus-or-minus/decide.mjs | [conversions-small-exact] small and opposite (recomputed: 22 of 40 with opposite directions, 23 or 24 with the same) but not both exact: the chemical scale rested on natural oxygen, whose isotopic abundance varies ("about 1.000282", "isotopes of oxygen had different natural abundances in water and in air"), and decide.mjs line 29 itself says CHEMICAL_TO_PHYSICAL = 1.000275 is "inherited, not derived" [declined: tier B, listed by the fixer]
- CONFIRMED | Did the Error Bars Hold?, whose eight Avogadro rows sit on all three | research/error-bar-coverage/README.md | [avogadro-eight-rows-three-scales] "chemical through 1941, physical from 1947 to 1955, unified afterwards" across eight reviews, one N_A row each
- CONFIRMED | One layer here ships Birge's 1941 recommendation | research/error-bar-coverage/data/anchor-rows.json | [birge-1941-in-ebc] reviews[1] is Birge 1941, Rev. Mod. Phys. 13, 233
- CONFIRMED | another deliberately omits Birge's 1941 and 1944 rows because it could not verify them from a primary source. | research/the-charge-that-crept/data.mjs | [birge-omitted-in-charge] line 60 "Birge's 1941/1944 rows were EXCLUDED, could not be verified"; facts.md E4 gives the reason (poor-provenance secondary, conflicting figures).
- CONFIRMED | The Charge That Crept, which carries five conventions in one column and converts between them never | research/the-charge-that-crept/data.mjs | [five-conventions-never-converted] errType values lim, pe, q, sd, 3sd; header "the convention THE SOURCE ITSELF used (never silently converted)"
- CONFIRMED | The Half-Lives That Would Not Agree, which watches five evaluators refuse an average at a numeric threshold | src/content/strata/halflife-disagreement.md | [five-evaluators] relates note says it "watches five evaluators formally decide, at a numeric threshold"
- CONFIRMED | Here are forty rows: five constants, as recommended by eight reviews between 1929 and 1969, hand-transcribed from the primary papers by Did the Error Bars Hold? and shipped beside that layer. | research/error-bar-coverage/README.md | [forty-rows-five-constants-eight-reviews] "We transcribed all forty values and uncertainties by hand from the eight primary reviews"
- CONFIRMED | In 1986 Henrion and Fischhoff reported that 57% of them fell more than 2.33 of their own standard uncertainties from the then-current value, where a calibrated set would put 2% outside. | https://www.Gwern.net/doc/statistics/bias/1986-henrion.pdf (p. 794 Table I and p. 795 text) | [hf-57-2.33-2] Table I row "Recommended values 1928-1973 40 7.42 (0) 22% 57%"; "*SI Surprise index: Percent that fall outside assessed 98% confidence interval, or outside 2.33"; "Well-calibrated normal distribution 1.00 50% 2%". The record cites p. 794 text and p. 793 table; they are pp. 795 and 794
- CONFIRMED | Henrion and Fischhoff | https://api.crossref.org/works/10.1119/1.14447 | [hf-crossref] Henrion and Fischhoff, Assessing uncertainty in physical constants, Am. J. Phys. 54, 791 (1986)
- CONFIRMED | They never printed the rows. | https://www.Gwern.net/doc/statistics/bias/1986-henrion.pdf | [hf-never-printed-rows] no table of the 40 values; Fig. 3 plots deviations for 1952-1973 only (positive control: text search finds the Table I row)
- MINOR | Their paper names the five constants and the eight reviews and gives the aggregate | https://www.Gwern.net/doc/statistics/bias/1986-henrion.pdf (p. 795 and reference list p. 797) | [hf-names-constants-and-reviews] the paper does not name the eight reviews: it says "eight reviews from 1929 to 1969" with superscripts 3,17,19,28,29, whose entries are Cohen-Crowe-DuMond 1957, Taylor-Parker-Langenberg, Birge Rep. Prog. Phys. 8, 90 (1942), DuMond-Cohen 1953, Cohen et al. 1955, Cohen-DuMond 1965, Cohen-Taylor 1973 and Birge 1929; the 1947 and 1950 reviews the anchor transcribed are not among them. The five constants are named only as Fig. 3 panel labels [declined: tier B, listed by the fixer]
- MINOR | That number is the reason anybody says physicists underestimate their error bars. | https://www.Gwern.net/doc/statistics/bias/1986-henrion.pdf (p. 795) | [physicists-underestimate] an uncountable causal superlative, and the claim predates the number: the same paper quotes Taylor et al. (1969) that investigators "tend to underestimate the sources of systematic error in their experiments" [declined: tier B, listed by the fixer]
- CONFIRMED | Eight rules do, five of them landing on 57.50% and three on 57.14%, and the naive one, values and uncertainties exactly as printed, is among them; that is the rule the member layer says on its own face it used. | recreated: node -e anchorRun over ANCHOR_AXES in public/strata/every-plus-or-minus/engine/findings.mjs | [eight-rules-reproduce] 8 rules within one point at 2.33: five at 57.50, three at 57.14; the naive rule (as printed, 1965 review, as printed, all eight, 1973) is one of them
- CONFIRMED | that is the rule the member layer says on its own face it used | research/error-bar-coverage/README.md | [member-says-naive] "Values and uncertainties are used exactly as each review printed them"; "The naive reading ... is the one that returns their published number"
- CONFIRMED | Across the whole grid of 144 rules the surprise rate runs from 42.5% to 80.0%. | recreated: public/strata/every-plus-or-minus/engine/findings.mjs | [grid-range] 144 rules enumerated: min 42.5%, max 80.0%
- CONFIRMED | by a factor of at least 11.5× | recreated: public/strata/every-plus-or-minus/engine/findings.mjs | [conclusion-floor] smallest ratio to calibrated expectation 11.538
- CONFIRMED | which rows you admit moves the answer more than what you take the sign to mean, and both move it more than which epoch you measure against. | recreated: public/strata/every-plus-or-minus/engine/findings.mjs | [main-effect-order] swings: threshold 13.4, include 7.8, convention 6.3, mass 4.0, charge1963 2.7, reference 1.3
- CONFIRMED | The forty rows are scored against the 1973 adjustment of the constants, which is what Henrion and Fischhoff used. | https://www.Gwern.net/doc/statistics/bias/1986-henrion.pdf (p. 795) | [reference-1973-hf] "with residuals calculated relative to 1973 estimates"; ref 28 includes Cohen and Taylor 1973
- CONFIRMED | Cohen and Taylor, J. Phys. Chem. Ref. Data 2, 663 (1973), footnote 8, on the page where section 11 begins | https://physics.nist.gov/cuu/pdf/1973JPCRD.pdf (p. 677, rendered) | [cohen-taylor-cite] footnote 8 at the foot of the column where "11. Summary of the More Precise Data" begins
- CONFIRMED | Three of those five constants have since been fixed exactly by the 2019 SI, and the other two are known far better than they were. | https://physics.nist.gov/cuu/Constants/Table/allascii.txt (refetched 2026-10-19, same SHA-256) | [three-fixed-two-better] N_A, e, h "(exact)"; alpha^-1 uncertainty 0.000000021 against 0.00011 in 1973; m_e 0.0000000028e-31 against 0.000047e-31
- CONFIRMED | The 1973 recommended values, scored by the 1986 rule | https://physics.nist.gov/cuu/pdf/1973JPCRD.pdf (pp. 717-718, rendered) | [table-1973-values] alpha^-1 137.03604(11); e 1.6021892(46) and 4.803242(14); h 6.626176(36); N_A 6.022045(31); m_e 9.109534(47)
- CONFIRMED | Four of the five are more than 2.33 of their own standard deviations from the value we now hold. | recreated: public/strata/every-plus-or-minus/engine/findings.mjs | [four-of-five] z = 0.37, 2.66 (esu; 2.73 in coulombs), 2.94, 3.20, -3.09
- OBSERVED | and the modern values are fetched from the NIST table rather than recalled. | research/every-plus-or-minus/build-data.mjs; recreated: curl allascii.txt, sha256sum | [modern-values-fetched] 2026-10-19 refetch: 40801 bytes, sha256 77fb90e6...67e67, the digest the build script records
- CONFIRMED | Table 33.4, the same paper, page 722 | https://physics.nist.gov/cuu/pdf/1973JPCRD.pdf (p. 722, rendered) | [table-334-heading] "TABLE 33.4. Expanded and combined variance-covariance and correlation coefficient matrix for our final recommended set of constants"
- CONFIRMED | the uncertainties of these constants are correlated, and therefore the general law of error propagation must be used in calculating additional quantities requiring two or more of these constants. | https://physics.nist.gov/cuu/pdf/1973JPCRD.pdf (p. 719, footnote a, rendered) | [footnote-a-quote] verbatim match
- CONFIRMED | The three constants that move up together, e, h and m_e, are correlated with one another between 0.904 and 0.991, and the one that moves the other way, N_A, is anti-correlated with all three | https://physics.nist.gov/cuu/pdf/1973JPCRD.pdf (p. 722, rendered) | [correlations-e-h-me] e-h 0.991, e-m_e 0.904, h-m_e 0.953; N_A-e -0.902, N_A-h -0.951, N_A-m_e -0.997
- CONFIRMED | Every coefficient is given to three decimals, so each is known only to within five ten-thousandths. | https://physics.nist.gov/cuu/pdf/1973JPCRD.pdf (p. 722, rendered) | [three-decimals] every coefficient printed to three decimals
- WRONG | Take the box that describes and evaluate the joint statistic at all 1024 of its corners | recreated: numpy eigvalsh and cholesky over the 1024 corners of Table 33.4 (https://physics.nist.gov/cuu/pdf/1973JPCRD.pdf p. 722) with the engine z vector; public/strata/every-plus-or-minus/engine/findings.mjs line 141 | [1024-corners] the page says the statistic "comes back between 0.34 and 1684, a factor of 5,014, and 36 of those corners are not a valid correlation matrix at all". 837 of the 1024 corners have a negative eigenvalue and are not valid correlation matrices; the engine discards only the 36 where z'R^-1z is not positive. The 0.34 comes from such a matrix (min eigenvalue -0.000144). Over the 187 valid corners the statistic runs 14.9 to 1684, a factor of 113 [fixed 2026-10-18]
- CONFIRMED | Two further things fall straight out of it, and both hold. | https://physics.nist.gov/cuu/pdf/1973JPCRD.pdf (p. 722) | [pairs-above-0.9] six of the ten pairs among the five exceed 0.9 in magnitude (recounted)
- CONFIRMED | Recall that σ_I, the uncertainty determined by internal consistency, is the expected uncertainty in the mean as determined by the a priori uncertainties, σ_i, assigned each individual measurement | https://physics.nist.gov/cuu/pdf/1973JPCRD.pdf (p. 677 footnote 8, rendered at 300 dpi) | [birge-footnote-quote] verbatim match of the first sentence
- CONFIRMED | R_B>1 generally implies that either the σ_i have been underestimated or that some or all of the data contain systematic errors. Wherever applicable, we shall quote the larger uncertainty. | https://physics.nist.gov/cuu/pdf/1973JPCRD.pdf (p. 677 footnote 8, rendered) | [birge-footnote-rest] verbatim match
- CONFIRMED | Our final recommended set of constants based on adjustment No. 41 of table 31.2. | https://physics.nist.gov/cuu/pdf/1973JPCRD.pdf (p. 717, rendered) | [table-331-caption] "TABLE 33.1. Our final recommended set of constants based on adjustment No. 41 of table 31.2. χ2 = 14.50 for 27 − 6 = 21 degrees of freedom; RB = 0.83a"
- CONFIRMED | The square root of 14.50 over 21 is 0.8309, so the caption's arithmetic checks. | research/every-plus-or-minus/facts.md C5; recreated with python3 sqrt(14.5/21) | [sqrt-14.5-21] 0.83095
- CONFIRMED | The paper says so itself in the footnote above, where a ratio above one is explained by underestimated errors or systematic error, and where nothing at all is offered for a ratio below one. | https://physics.nist.gov/cuu/pdf/1973JPCRD.pdf (p. 677 footnote 8) | [nothing-offered-below-one] the footnote explains RB > 1 only; nothing is said of RB < 1
- MINOR | the oil-drop measurements of the electron charge agreed beautifully with each other for twenty years, because they shared a wrong value for the viscosity of air. | public/strata/the-charge-that-crept/index.html; research/the-charge-that-crept/data.mjs | [oil-drop-viscosity] the member's finding is that Millikan's five VISCOSITY determinations agreed "to better than one part in two thousand" and shared a systematic error, and that the raw measurements "never crept. They split, oil-drop values pinned low by one shared wrong viscosity"; it does not say the oil-drop measurements agreed beautifully for twenty years, and its own data carries Bäcklin and Flemberg 1936, an oil-drop value of 4.7517 below Millikan's 4.774 ("the method still scattered low") [declined: tier B, listed by the fixer]
- CONFIRMED | The half-life census gives 45 sets of measurements that a real evaluation flagged as contested. | research/halflife-disagreement/README.md | [census-45] 46 comments with a numeric Birge quantity, 45 print complete inputs; the member dek: NUBASE2020 "flags a contested half-life by printing a Birge ratio"
- CONFIRMED | Run each set through six combination rules, each crossed with three values of each of the two thresholds the evaluators use, 54 settings in all. | public/strata/every-plus-or-minus/engine/findings.mjs | [54-settings] HALFLIFE_AXES: six combine rules, t1 [1.5, 2.5, 3.5], t2 [3, 4, 6]; 6 x 3 x 3 = 54; NUBASE2020 uses 2.5 and 4
- CONFIRMED | across all 54 settings any one entry yields at most 5 distinct answers, because the threshold rule can only ever hand back one of the estimators the other five already give. | recreated: node over HALFLIFE.rules and combine() in research/every-plus-or-minus/decide.mjs | [at-most-5-answers] max distinct (value, sigma) outcomes per entry across 54 rules: 5
- CONFIRMED | NUBASE's fixed threshold of 2.5 | https://www-nds.iaea.org/amdc/ame2020/NUBASE2020.pdf | [nubase-threshold-2.5] "It is a policy of NUBASE 2020 to use the weighted average result (equation 3) when χn is smaller or equal to 2.5"
- CONFIRMED | CODATA's 1 + sqrt(2/ν) call the same | https://srd.nist.gov/jpcrdreprint/1.556049.pdf | [codata-threshold] "Thus a value of the Birge ratio greater than 1 + sqrt(2/ν) would suggest a possible inconsistency in the data" (CODATA 1998)
- CONFIRMED | NUBASE's fixed threshold of 2.5 and CODATA's 1 + sqrt(2/ν) call the same 42 of 45 sets inconsistent and disagree on none. | recreated: python3 over research/halflife-disagreement/data/birge-halflife-census.json | [nubase-codata-agree-45] independent recount: 42 and 42 of 45, disagree on 0
- CONFIRMED | its published rule excludes measurements whose error exceeds three root-N times the error of the mean, and the least precise of the three falls outside that ceiling. | https://pdg.lbl.gov/2024/reviews/rpp2024-rev-rpp-intro.pdf | [pdg-ceiling] "We evaluate S using only the experiments with smaller errors. Our cutoff or ceiling on δxi is arbitrarily chosen to be δ0 = 3N1/2 δx"; the page says "computed from", so excluding from S is what it claims
- CONFIRMED | computed from 2 of the 3 | https://www-nds.iaea.org/amdc/ame2020/NUBASE2020.pdf (196Au comment); recreated with python3 | [gold196-ceiling-excludes] delta0 = 0.00283; 63Ik01 sigma 0.010 exceeds it; S from the other two = 15.23
- CONFIRMED | the three measurements give a Birge ratio of [figure] against the [figure] the evaluation prints. | https://www-nds.iaea.org/amdc/ame2020/NUBASE2020.pdf | [gold196-printed-birge] "196Au T : unweighted average 11Hi19=6.1451(0.0013) 01Li17=6.1669(0.0006) 63Ik01=6.183(0.010); Birge ratio=10.86"; recomputed 10.8578
- CONFIRMED | the median-with-scaled-deviation estimator returns an uncertainty of exactly zero, because more than half the measurements share a value and the median absolute deviation is then zero. | recreated: public/strata/every-plus-or-minus/engine/findings.mjs | [zero-sigma-defect] zeroSigmaEntries = 1; MAD is zero when more than half the values coincide
- CONFIRMED | A printed figure of 0.124 is a claim about the interval from 0.1235 to 0.1245, whose endpoints are exact rationals and whose width the printer chose. | src/content/strata/the-room-inside-a-number.md | [room-interval] "0.124 could have been anything from 0.1235 to 0.1245 before it was rounded"
- CONFIRMED | whose endpoints are exact rationals | public/strata/the-room-inside-a-number/index.html | [room-rationals] "Every value is a ratio of two arbitrary-precision integers"
- CONFIRMED | On the same 2,842 bullet comparisons, four conventions that are all in print give false-positive rates of 0.70%, 2.04%, 0.70% and 66.19%. | https://www.ebi.ac.uk/europepmc/webservices/rest/PMC9254335/fullTextXML | [ames-four-rates] Table 2 different-source bullets total 2842; Table 3 different source 2.04%, 0.70%, 66.19%, 0.70%
- CONFIRMED | Its uncertainty is a Clopper-Pearson interval on a count, not a symmetric sigma | public/strata/the-answer-that-cant-be-wrong/index.html | [clopper-pearson] "reproduces PCAST's three confidence bounds with an independently implemented Clopper-Pearson interval"
- CONFIRMED | Gold-196's three measurements really do not overlap, under every rule on the grid. | research/halflife-disagreement/README.md (recomputed) | [gold196-no-overlap] gaps 0.0218, 0.0161, 0.0379 d against summed errors 0.0019, 0.0106, 0.0113
- CONFIRMED | Every recommended value of the electron charge from 1929 to 1963 really does sit below the exact value, under every reading of every sign. | public/strata/the-charge-that-crept/index.html | [charge-below-exact] 1929 4.770, 1939 4.8025, 1947 4.80214, 1950 4.80223, 1952 4.80288, 1955 4.80286, 1963 4.80298 all below 4.80320471
- CONFIRMED | The historical intervals really did fail to cover, by a factor of at least 11.5× under all 144 rules. | recreated: public/strata/every-plus-or-minus/engine/findings.mjs | [historical-intervals-fail] 11.54
- WRONG | Everything above is recomputed in your browser from data files committed beside the member layers, and independently in a verifier that does not import the page's engine. | research/every-plus-or-minus/verify.mjs lines 120 and 432 | [check-recomputed-in-browser] the verifier does import the page's engine: `const F = await import('./findings.mjs')` and `await import('./decide.mjs')`, and cmp shows both byte-identical to public/strata/every-plus-or-minus/engine/. Several sections assert the engine's own outputs rather than recomputing them (e.g. `ok(CO.probeCorners < 1024, ...)`, `near(F.HALFLIFE.driftMedian, 3.83, ...)`), which is how the corner miscount (1024-corners) passed [fixed 2026-10-18]
- CONFIRMED | reads the members' own committed files and shipped bytes; the verifier re-runs that assembly and requires the result to come back byte-identical, so a member editing its data turns this layer red instead of leaving it stale. | research/every-plus-or-minus/build-data.mjs; research/every-plus-or-minus/verify.mjs | [build-data-byte-identical] §1 re-runs assemble() and compares to data.mjs; the run at 0a15703d9e passed
- OBSERVED | The modern values were fetched, not recalled, from the NIST CODATA table, whose digest and fetch time are recorded, and the five lines used are quoted verbatim in the build script and re-parsed from that quotation by the verifier. | research/every-plus-or-minus/build-data.mjs; recreated: curl allascii.txt, sha256sum | [nist-fetched-digest] same digest on 2026-10-19; build-data.mjs quotes the five lines and verify.mjs §1 re-parses them
- OBSERVED | The PDF returned the same SHA-256 that the member layer's provenance record wrote down a week earlier. | research/error-bar-coverage/data/PROVENANCE.md; recreated: curl 1973JPCRD.pdf, sha256sum | [1973-pdf-sha] 2026-10-19: 7,051,710 bytes, c2f646c2...68b17, matching the record of 2026-08-01
- OBSERVED | The 1973 quotations were read from page renders, not from a text layer. | research/every-plus-or-minus/facts.md | [quotations-read-from-renders] C2-C4 record reads from 300 dpi renders; the PDF also has an OCR text layer, so this is a statement of method, not of the file
- OBSERVED | After this page was built, an independent agent was told to refute it rather than confirm it. It rebuilt every central statistic in Python before reading the engine and reproduced all of them, and it verified all four quotations against the source scans. | git show 0a15703d9e:memory/log.d/2026-08-09T0000Z-every-plus-or-minus.md | [adversarial-pass] "It reproduced every central statistic in Python before reading the engine, verified all four quotations against the source scans, and then found sixteen things"
- MINOR | It also found that the first draft's headline was wrong, that the check count had a unit slip in it, that the verifier could be left green while the page was falsified, and five smaller overclaims. | research/every-plus-or-minus/facts.md section L | [adversarial-found] the record does not say the check count had a unit slip: L10 is a unit slip in "1.60210(2) coulombs" and L6 is an off-by-one in the portal count. And it found sixteen things; besides the three named, at least eight are overclaims (L2, L7, L8, L11, L12, L13, L14, L4), not five [declined: tier B, listed by the fixer]
- WRONG | cross-check.py reads only the raw member data files, uses Python's own error function rather than any approximation the JavaScript carries, and re-derives every number above independently. | research/every-plus-or-minus/cross-check.py; recreated: python3 research/every-plus-or-minus/cross-check.py | [cross-check-py] it does not re-derive every number above: its output covers the 1963 readings, the three sensitivities, the 144-rule grid, the five reference z values, the 0.83 caption and gold-196's weighted mean and Birge ratio; it computes nothing for the correlation section (participation ratio, condition number, the 1024 corners), the half-life drift, the threshold null, the PDG factor, Hubble or muon [fixed 2026-10-18]
- CONFIRMED | The probable-error factor is recomputed from the inverse normal distribution, and the physical-to-unified mass-scale factor is recomputed from the mass of oxygen-16. | research/every-plus-or-minus/verify.mjs lines 61 and 166 | [conversion-constants-derived] ownProbit(0.75) and 16 / 15.9949146196; the heading "The conversion constants are derived" covers two of three (decide.mjs line 29 says 1.000275 is inherited)
- OBSERVED | With the directions derived rather than assumed, the record reproduces exactly. | recreated: python3 recount over research/error-bar-coverage/data/anchor-rows.json | [correction-before-shipped] opposite directions give 22 of 40
- OBSERVED | applying them the same way gives 23 or 24 instead of the 22 that record states. | recreated: python3 recount over research/error-bar-coverage/data/anchor-rows.json | [mass-scale-23-or-24] both multiplied gives 24, both divided gives 23
- OBSERVED | All 59 portal layers that existed before this one, 125 files, swept with word-boundary matching for the seven member slugs and for the vocabulary of this join. | recreated: git ls-tree at dfa8dc92223 (the commit that added the layer) | [unwalked-sweep] 59 portal-tagged strata besides this one; their .md plus .html/.json/.js/.mjs under public/strata come to 126, 125 without the-number-nobody-is-holding/data.json, which facts.md J3 says was excluded
- CONFIRMED | the-answer-that-cant-be-wrong appears in the 90-member list of The Anatomy of Error | git show dfa8dc92223:public/strata/the-anatomy-of-error/data.json | [anatomy-of-error-90] members has 90 entries and contains the-answer-that-cant-be-wrong
- MINOR | error-bar-coverage's own research record already varies three of these choices, one at a time, and reports that each moves the count by one row. | research/error-bar-coverage/README.md lines 25-29; recreated: anchorRun in public/strata/every-plus-or-minus/engine/findings.mjs | [credit-ebc-one-at-a-time] the README does say "Each of those choices changes the count by one row", but its own figures are 24, 21 and 22 against a base of 23 (recreated: 23, 21, 22), so the probable-error conversion moves two rows; and the first of its three choices (the 1929 sub-line padding digit) is not one of this page's six decisions and is not on the 144-rule grid [declined: tier B, listed by the fixer]
- CONFIRMED | halflife-disagreement's card already names What the Average Throws Away as the layer that distrusts an average and reads what it discards. | src/content/strata/halflife-disagreement.md | [halflife-card-names-average] "That page distrusts the average and reads what it discards"
- OBSERVED | Offline: 143 checks. | recreated: node research/every-plus-or-minus/verify.mjs in a git archive of 0a15703d9e | [143-checks] "ASSERTIONS 143 FAILURES 0 MUTATIONS CAUGHT 7/7"
- OBSERVED | returned 403 on every route tried | research/every-plus-or-minus/facts.md K1; recreated: curl https://journals.aps.org/rmp/abstract/10.1103/RevModPhys.37.537 | [403-every-route] 403 on 2026-10-19
- CONFIRMED | Six decisions stand between a set of measurements and the sentence that they agree, and here they are on real data from five layers, with 144 rules crossed and the answer recomputed live. | public/strata/every-plus-or-minus/engine/findings.mjs | [dek-144-rules] ANCHOR_AXES 2 x 3 x 2 x 2 x 2 x 3 = 144; five datasets with a sigma
- CONFIRMED | The published 57 percent moves between 42 and 80. | recreated: public/strata/every-plus-or-minus/engine/findings.mjs | [dek-57-moves-42-80] 42.5 to 80.0
- CONFIRMED | That page reproduces a published surprise rate of 57 percent and lists three one-at-a-time sensitivities in its research record | research/error-bar-coverage/README.md | [rel-three-one-at-a-time] three one-at-a-time sensitivities listed under "Where the reconstruction is fragile"
- MINOR | this one crosses them into 144 rules, adds two more decisions, and separates the reproduction, which moves between 42 and 80 percent, from the conclusion, which every rule agrees with. | public/strata/every-plus-or-minus/engine/findings.mjs ANCHOR_AXES; research/error-bar-coverage/README.md | [rel-crosses-144] the grid crosses two of the three (probable-error conversion, mass scale), not the 1929 padding digit, and adds three decisions the member never varied (the 1963 row, which reviews are in, the reference epoch), not two; the 42 to 80 range and the conclusion are right [declined: tier B, listed by the fixer]
- MINOR | It also turns the surprise test on that page's own 1973 reference values, where four of five fail it. | research/every-plus-or-minus/facts.md L1; public/strata/every-plus-or-minus/index.html | [rel-four-of-five-fail] the body corrected exactly this framing ("Those are not four failures. They are one failure, counted four times") and facts.md L1 calls the earlier headline wrong; the relates card still says "four of five fail it" [declined: tier B, listed by the fixer]
- CONFIRMED | finds that of 23 rows exactly one crosses the line when the stated conventions are honoured, and it is the 1963 row. | recreated: public/strata/every-plus-or-minus/engine/findings.mjs CHARGE | [rel-23-rows-one-crosses] 23 rows, 10 misses as printed and 11 converted, moved = [1963]
- CONFIRMED | That page tags the 1963 uncertainty as a three-sigma limit and refuses ever to convert a convention silently | research/the-charge-that-crept/data.mjs | [rel-charge-3sigma-tag] "'3sd' = three-standard-deviation error limit" and "never silently converted"
- WRONG | Forty of the 45 move further than the uncertainty they are published with. | research/halflife-disagreement/data/birge-halflife-census.json; recreated: node over HALFLIFE.per in public/strata/every-plus-or-minus/engine/findings.mjs | [rel-forty-of-45] 40 of 45 is drift measured in the weighted mean's internal sigma ("the error you would have printed ... had you never thought about the rule", as the body correctly says), not the uncertainty NUBASE2020 publishes. Against the published (adopted) uncertainties 23 of 44 comparable entries travel further (10Li-n not comparable on units); gold-196 itself moves 0.0038 d against a published ±0.011 d. Same wording in the SUMMARY row on the page [fixed 2026-10-18]
- CONFIRMED | The Particle Data Group's own scale factor on its worked case, gold-196, comes out at 15.2 against the evaluation's 10.86, computed from two of the three measurements because PDG's published ceiling excludes the third. | https://www-nds.iaea.org/amdc/ame2020/NUBASE2020.pdf (196Au comment); recreated with python3 | [rel-gold196-15.2] S = 15.23 from 2 of 3; NUBASE prints 10.86
- CONFIRMED | The Particle Data Group's own scale factor on its worked case, gold-196 | src/content/strata/halflife-disagreement.md | [rel-pdg-worked-case] "its worked case" reads as the half-life layer's, which opens on gold-196 ("Rebuild gold-196's published 6.165(11) d"); PDG does not treat gold-196, and the page does not say it does
- CONFIRMED | crossing its four late-universe samples with its three early-universe values and five correlation assumptions gives 60 cells running from 1.1 to 7.7 sigma, and all three verdicts occur. | recreated: public/strata/every-plus-or-minus/engine/findings.mjs HUBBLE | [rel-60-cells] 4 x 3 x 5 = 60 cells, 1.13 to 7.71, discrepant 19, consistent 16, tension 25
- CONFIRMED | That page states that its truncated uncertainties make every tension it reports an upper bound rather than the significance. | public/strata/two-rulers-will-not-agree/index.html | [rel-upper-bound] "every tension this instrument reports against the early-universe value is an upper bound on the real significance, not the significance itself"
- CONFIRMED | Hold the experiment still and change only the theory reference, and the muon anomaly runs from 5.7 sigma to 0.6. | recreated: public/strata/every-plus-or-minus/engine/findings.mjs MUON; https://arxiv.org/abs/2505.21476 | [rel-muon-5.7-0.6] at zero correlation 5.73 (Fermilab) or 5.76 (world) against WP20 and 0.59 or 0.60 against 2025; the 2025 paper gives a_exp - a_SM = 38(63)e-11
- CONFIRMED | That layer also declines to publish a combined estimate at all | https://arxiv.org/abs/2505.21476; public/strata/the-gap-that-changed-sides/index.html | [rel-muon-refuses-to-combine] abstract "impossible to combine the results in a meaningful way"; the layer shows "NO CONSENSUS VALUE" and "will not manufacture a current z score"
- CONFIRMED | Its intervals come from rounding rather than scatter, with endpoints that are exact rationals, so there is no sigma to decide about and none is invented. | public/strata/the-room-inside-a-number/index.html | [rel-no-sigma-rounding] rational arithmetic, no sigma
- CONFIRMED | Integer counts and a category that has to be scored before a rate exists, with four conventions in print giving 0.70, 2.04, 0.70 and 66.19 percent on the same comparisons. | https://www.ebi.ac.uk/europepmc/webservices/rest/PMC9254335/fullTextXML | [rel-answer-four-conventions] Table 3 different-source bullets
- CONFIRMED | Its participation ratio is 1.52 of five. | recreated: numpy eigvalsh on Table 33.4 (https://physics.nist.gov/cuu/pdf/1973JPCRD.pdf p. 722) | [participation-ratio] participation ratio 1.517
- CONFIRMED | And the constant that passes the test is alpha-1, whose largest correlation with any of the other four is 0.457, the only one of the five that is not mostly the same number again. | https://physics.nist.gov/cuu/pdf/1973JPCRD.pdf (p. 722) | [alpha-passes-0.457] alpha^-1 correlations with N_A, e, h, m_e: 0.034, -0.457, -0.334, -0.034
- CONFIRMED | one shared shift of roughly three of those sigmas is enough to produce all four. | recreated: public/strata/every-plus-or-minus/engine/findings.mjs | [one-shared-shift-three-sigma] z 2.66, 2.94, 3.20 and -3.09
- CONFIRMED | The 1973 recommended values did move by more than their own quoted uncertainties | recreated: public/strata/every-plus-or-minus/engine/findings.mjs | [moved-more-than-quoted] four of five beyond 2.33 of their own 1973 sigma
- WRONG | The honest output is the range, and the range is four orders of magnitude wide. | research/every-plus-or-minus/facts.md L1; recreated with numpy over the 1024 corners of https://physics.nist.gov/cuu/pdf/1973JPCRD.pdf p. 722 | [four-orders-of-magnitude] over the 187 corners that are valid correlation matrices the range is a factor of 113, about two orders of magnitude; even on the page's own 5,014 it is 3.7 [fixed 2026-10-18]
- OBSERVED | The sweep was positive-controlled on slugs known to be present. | research/every-plus-or-minus/facts.md J1 | [sweep-positive-control] record states the sweep was positive-controlled on four slugs; not re-run
- CONFIRMED | all three verdicts occur on the grid: | recreated: public/strata/every-plus-or-minus/engine/findings.mjs HUBBLE | [rel-all-three-verdicts] discrepant 19, consistent 16, tension 25
- CONFIRMED | sigma with the experiment held still | recreated: public/strata/every-plus-or-minus/engine/findings.mjs MUON | [rel-muon-range] 0.59 to 6.90 across 12 cells
- CONFIRMED | (added) The same bureau, on the same kind of card, then published the factor of three twice more, once inside the number and once outside it. | https://www.govinfo.gov/content/pkg/GOVPUB-C13-dc85722872ab7945773bde84551dd8af/pdf/GOVPUB-C13-dc85722872ab7945773bde84551dd8af.pdf (PDF p. 2, revised May 1969) and https://www.govinfo.gov/content/pkg/GOVPUB-C13-a2fb26e45504787736a93a7f20d21bc7/pdf/GOVPUB-C13-a2fb26e45504787736a93a7f20d21bc7.pdf (PDF p. 2) | the May 1969 revised printing of Misc. Publ. 253 keeps "Based on 3 std. dev."; SP 344 (1971) prints 1 std. dev. with "multiply the above uncertainties by 3" beneath
- WRONG | (added) Both documents behind the disagreement are opened here and both layers turn out to be right. (relates note to the-charge-that-crept) | research/every-plus-or-minus/facts.md K1 and L11 | [added-both-documents-opened] the 1965 review behind the anchor's ±0.00006 was not opened: "Cohen and DuMond's 1965 paper itself was not opened: every route to it returned 403"; L11 corrected the same overclaim in the body, and the body now says so, but the relates card still says both were opened [fixed 2026-10-18]
- MINOR | (added) this one takes its 45 flagged sets through 54 published combination rules (relates note to halflife-disagreement) | research/every-plus-or-minus/facts.md L8 | [added-54-published-rules] L8: "54 published combination rules" was "inflated and published was unearned for the threshold levels. Corrected on the page"; the body now says six rules and 54 settings, "sensitivity probes rather than published rules", but the relates card keeps the uncorrected wording [declined: tier B, listed by the fixer]
- MINOR | (added) This one measures how wide the bound is (relates note to two-rulers-will-not-agree) | public/strata/two-rulers-will-not-agree/index.html; public/strata/every-plus-or-minus/engine/findings.mjs HUBBLE | [added-how-wide-the-bound] the 60 cells vary the sample and the correlation but all use the same truncated uncertainties (the NGC 4258 maser term is not added), so each cell is still an upper bound; the grid measures how far the bound moves across choices, not how far it sits above the true significance [declined: tier B, listed by the fixer]
- CONFIRMED | (added) Across the 54 rules its value runs 6.163131 to 6.166900 days and its uncertainty runs 0.000544 to 0.017272, a factor of 31.8. | recreated: public/strata/every-plus-or-minus/engine/findings.mjs HALFLIFE.gold; python3 | weighted mean 6.163131, median 6.1669, internal 0.000544, scaled-MAD median 0.017272, ratio 31.75
- CONFIRMED | (added) median drift 3.83, largest drift 17.52, further than their own sigma 40 of 45 (body readout, in internal-sigma units) | recreated: public/strata/every-plus-or-minus/engine/findings.mjs HALFLIFE | 3.83, 17.52 (194Ir), 40 of 45 in the units the body defines; see rel-forty-of-45 for the relates card
- CONFIRMED | (added) The condition number is 3.6e+5 | recreated: numpy eigvalsh on Table 33.4 (https://physics.nist.gov/cuu/pdf/1973JPCRD.pdf p. 722) | 3.555e5
- MINOR | (added) the verifier could be left green while the page was falsified ... the verifier now covers what it did not. | recreated: edit public/strata/every-plus-or-minus/index.html in a scratch copy, then node research/every-plus-or-minus/verify.mjs | [added-verifier-coverage] changing "Eight rules do" to "Nine rules do", or "between 0.904 and 0.991" to "between 0.914 and 0.991", leaves it at ASSERTIONS 143 FAILURES 0; §8c gates six prose literals. Not the step-6 self-reading case (the page prints no verifier output beyond the check count) [declined: tier B, listed by the fixer]

## Findings

- [fixed] 1024-corners (WRONG): the engine kept a corner whenever z'R^-1z > 0, so the page said 36 of the 1024 corners were invalid and the statistic ran 0.34 to 1684 (factor 5,014). research/every-plus-or-minus/findings.mjs and its byte-identical copy public/strata/every-plus-or-minus/engine/findings.mjs now keep a corner only when R is positive definite (eigenvaluesSym(Rp)[0] > 0): 837 invalid, 187 valid, 14.88 to 1684, factor 113, matching an independent numpy recomputation. index.html prose reworded so the range is stated over the valid corners (new readout c-good set in engine/page.mjs); verify.mjs's probeCorners < 1024 replaced by its own Cholesky count of valid corners (187) required equal to the engine's, shown to go red with the old filter; facts.md L1 given a dated correction. Corrected line added.
- [fixed] check-recomputed-in-browser (WRONG): index.html 'independently in a verifier that does not import the page's engine' now 'again in a verifier that redoes part of the arithmetic with its own code and imports the page's engine for the rest' (verify.mjs lines 120 and 432 import findings.mjs and decide.mjs). Corrected line added.
- [fixed] cross-check-py (WRONG): index.html 're-derives every number above independently. It agrees on all of them' narrowed to what the script prints (the opening row, the board's sensitivities and grid, the reference test and its 0.83 caption, gold-196's mean and Birge ratio), stating it does not reach the correlation matrix, the rest of the half-life section or the two-number cases; the cross-check.py docstring narrowed the same way. Corrected line added.
- [fixed] rel-forty-of-45 (WRONG): the relates note to halflife-disagreement (src/content/strata/every-plus-or-minus.md) and the SUMMARY row (findings.mjs, both copies) said forty of the 45 move further than 'the uncertainty they are published with'; the drift is in units of the plain weighted mean's internal sigma, so both now say 'the uncertainty a plain weighted mean would print beside them'; the engine comment and the verify.mjs label reworded to match. Corrected line added.
- [fixed] four-orders-of-magnitude (WRONG): index.html 'the range is four orders of magnitude wide' now 'two orders of magnitude wide' (factor 113 over the valid corners). Corrected line added.
- [fixed] added-both-documents-opened (WRONG): the relates note to the-charge-that-crept said both documents behind the disagreement are opened here; it now says the card and the 1973 adjustment's comparison table are opened and Cohen and DuMond's 1965 review could not be reached (facts.md K1, L11). Corrected line added.
- [declined] dek-3.7-1.1-sigma (MINOR, tier B, listed): MINOR, tier B, not in a sentence being fixed: recorded and declined. The checker's note holds (honouring the charge layer's 3sd tag gives 3.37 sigma, verify.mjs zC; 1.1 is the card's limit taken at face value), but the dek's figures are the page's own computed readings and the body explains them.
- [declined] five-word-footnote (MINOR, tier B, listed): MINOR, tier B: the footnote the page quotes in full is eleven words, of which the first five carry the three-sigma statement; loose wording, not false. Declined.
- [declined] triple-own-uncertainties (MINOR, tier B, listed): MINOR, tier B: the ** marker attaches to four of the card's uncertainties, so 'its own published uncertainties' is broader than the card; the blockquote beside it quotes the exact sentence. Declined.
- [declined] conversions-small-exact (MINOR, tier B, listed): MINOR, tier B: physical-to-unified is fixed by the oxygen-16 mass but chemical-to-physical rests on natural oxygen and decide.mjs itself calls 1.000275 inherited; a precision slip. Declined.
- [declined] hf-names-constants-and-reviews (MINOR, tier B, listed): MINOR, tier B: per the checker's reading of Henrion and Fischhoff p. 795, the paper cites the eight reviews by reference number and names the constants as figure panels; loose wording. Declined.
- [declined] physicists-underestimate (MINOR, tier B, listed): MINOR, tier B: an uncountable causal superlative; the claim predates the number (Taylor et al. 1969, quoted in the same paper). Declined.
- [declined] oil-drop-viscosity (MINOR, tier B, listed): MINOR, tier B: the member's finding is about Millikan's viscosity determinations agreeing and being shared, with oil-drop values pinned low; the page's compression is loose. Declined.
- [declined] adversarial-found (MINOR, tier B, listed): MINOR, tier B: facts.md L10 is the unit slip (in '1.60210(2) coulombs', not in the check count) and the smaller overclaims number more than five; an overstatement-class slip about the page's own checking history. Declined.
- [declined] credit-ebc-one-at-a-time (MINOR, tier B, listed): MINOR, tier B: the member README does say each choice moves the count by one row while its own figures give a two-row move for the probable-error conversion; the page repeats the member's words. Declined.
- [declined] rel-crosses-144 (MINOR, tier B, listed): MINOR, tier B: the grid crosses two of the member's three sensitivities and adds three decisions, not two; the 42 to 80 range and conclusion hold. Not in a sentence being fixed. Declined.
- [declined] rel-four-of-five-fail (MINOR, tier B, listed): MINOR, tier B: four of five do fail the 2.33 test (cross-check.py prints 4 SURPRISE); the card keeps the framing the body retracts as one failure counted four times. Declined.
- [declined] added-54-published-rules (MINOR, tier B, listed): MINOR, tier B: facts.md L8 corrected 'published' in the body; the relates card keeps it. This sentence is not the one being fixed for rel-forty-of-45 (that is the next sentence of the note). Declined.
- [declined] added-how-wide-the-bound (MINOR, tier B, listed): MINOR, tier B: the 60 cells all keep the truncated uncertainties, so they show how far the bound moves rather than how far it sits above the true significance; loose wording. Declined.
- [declined] added-verifier-coverage (MINOR, tier B, listed): MINOR, tier B: an overstatement of the page's own checking (the verifier gates six prose literals; other prose figures can change with it green). Declined.