Kilner-Walter: Quality Engineering (Step 4b)

QE run: 2026-10-04. Inputs checked: _text/manifest.json, all six digests, deep/humanatmosphere00kilngoog-d2c3dd.md, facts.json, claims.md, sites.md, all six research/ notes (the saved files listed only), physicist.md, mathematician.md, contrarian.md, arbiter.md, primer.md and example.py. I checked against the extracted text _text/humanatmosphere00kilngoog-d2c3dd.txt. I also rendered PDF pp.80-81 (Tables I and II) to /tmp/qe-kilner and read them.

Page tags below are PDF pages. The printed page number is the PDF page minus 19.

Verdict: MINOR. The central verdicts are unlikely to change: every claim still rests on one unblinded observer. But the case material behind claims 3 and 4 (Chapters V-VII) was never deep-read. This review also found several accuracy errors: OCR has misread case numbers, the sibling count is incomplete, and two menstrual "predictions" have been mis-scored. All of these can be fixed in facts.json, an arbiter "QE revision" section and the primer.

---

1. Coverage

2. Shortcuts

• Chapter IV and the Appendix (claims 5 and 6) had a thorough deep read. Acceptable.
• Chapter III (claim 2) was not deep-read. The Physicist (PDF pp.109-142) and the Mathematician (PDF pp.109-132) each read it themselves. Acceptable.
• Chapters V-VII, PDF pp.166-325: claims 3 and 4 rest on digests alone. Defect. These chapters contain the book's experimental results for the disease, epilepsy, hysteria, menstruation and pregnancy claims. The reviewers each read only a few pages: the Physicist PDF pp.314-318 and the Mathematician PDF pp.321-323. The Physicist says, "I did not read the dead-foetus and cancer sections in detail". As a result, the reviews never collected the cases where Kilner gives a reading before the outcome was known (see section 3). For claims 3 and 4 those cases are the only evidence that resembles a test.
• Chapters I-II, PDF pp.20-102 (claim 1: the protocol and the healthy-aura descriptions) are digest-only, apart from the reviewers' own reading of pp.24-27. The digest is adequate for the claim. Two observer details were missed, though (section 3).
• The Investigator's run of about 6 minutes was within its cap. Unreached sites are logged with "Recheck? yes". Acceptable.

3. Missed material (from the digests and my own reading of the text)

None of the following appears in facts.json, claims.md, the reviews, the arbiter or the primer, unless a role is named.

1. Predictions against the patient's own expectation (claim 4). Two cases are mis-scored, and one is missing. - The 38-year-old woman. Kilner saw an "incipient darkening" above the pubes. He told her to expect her period in about six or seven days. She said it was not due for a fortnight, and Kilner "noted [it] down as mistaken prognosis". Two months later she told him it had come "seven days afterwards, being a week before the proper time" [read: PDF p.253-254]. - The young woman. She expected her period "in three or four days". Kilner saw no darkening, and "her period was a week later than the right date" [read: PDF p.254]. Kilner reads this as the band correctly showing that the period was not imminent. The arbiter (section 4, claim 4) and the primer (evidence table) call this case "a miss". The Physicist says the timing was "off by days in two of the two outcomes". Those readings are wrong. By Kilner's own logic both cases are hits, and both went against what the patient herself expected. - Case 21. She believed she was pregnant. Kilner ruled pregnancy out and said menstruation would not come within 4-5 days. It came 7 days later [read: PDF p.317-318]. This is also against the patient's belief. - Why it matters. Predictions that run against the subject's own expectation bear directly on the cueing objection. They should be stated fairly beside the "no denominators" point, which still stands.
2. Other readings made before the history was known (claims 3 and 4). These are all in the digests and none is in any review. - Case 17 (boy, 13). "No questions before inspection." Kilner judged the aura "distinctly epileptic". His father later confirmed the boy was being treated for petit mal [read: PDF p.188-194]. - Case 18 (I.X., 58). "Before we inspected him, we had not the slightest idea what shape the Aura would take." Kilner concludes "masked epilepsy", a reading that cannot be checked [read: PDF p.194-195]. - Case 16. Kilner diagnosed epilepsy from the aura, and a fit followed days later. The Physicist cites this case once. - Case 37 (printed in the OCR as "87"). The patient coughed up blood with no organic cause found. A light C.C. spot appeared in the left second intercostal space, and the stethoscope over that spot found fine crepitation [read: PDF p.295-296]. - An unnumbered woman. A light spot appeared over the left scapula where she had back pain, with a breast "subsequently found malignant" [read: digest 05, PDF ~p.279-280]. - The one patient "said never to have had a fit" but with an epileptic-type aura: Kilner held to his opinion [read: PDF p.194].
3. Countable data the reviews missed. The Mathematician and the primer say Tables I-II are "the one place in the book with counts". That is inaccurate. - Colour series, "the first hundred cases inspected" [read: PDF p.247-248]:
   • Blue: 40 cases ("no individual below ... the average in mental power").
   • Blue with grey: 36 cases (1 hemiplegia, 2 epileptics, 1 meningitis).
   • Grey: 17 cases (6 epileptics, 2 eccentric, 1 general paralysis, 3 mentally dull).
   • Not recorded: 7.
   • Kilner calls this table "conclusive" for colour and mental fitness, and notes that the proportion of healthy people was larger than in his later inspections. It is a real count for claim 3 (colour and temperament), and nobody has analysed it. The classifier knew the diagnoses.
   • Epilepsy: "in all ten" cases the left side was diminished (4 male, 6 female, ages 12-45) [read: digest 03, PDF ~pp.173-176]. Kilner adds that in none of these cases did the C.C. band show a change typical of the complaint (PDF p.194), a negative result he reports himself.
4. Other observers. - Case 1: the infant's "mother and nurse" could also see the cloud through the screen [read: PDF p.41]. - Kilner's claim that if observers first look through the dark screen "all can see it", some within a minute [read: digest 02, PDF p.131-132]. This contrasts with Surya's "about 50% see nothing" and with the Paris failure. - Neither point is in the arbiter or the primer.
5. Investigator leads dropped from the arbiter and the primer (see section 5).
6. Case 30 (OCR "80"). Chapter IV cites this case to rule out aura thickness. It is a stout, non-pregnant 58-year-old woman whose abdomen made her look six months pregnant. Kilner said he could not explain her aura [read: PDF p.151-152, 235-238]. Digest 03 calls her "pregnant". The deep read cites "Fig. 21", but her figure is Fig. 24.

4. Accuracy spot checks

5. Retention check

The question here is whether pertinent material from the sources or the Investigator was left out of the arbiter or the primer because a reviewer judged it negatively. The supportive and contrary 1911-12 trials are all kept, each with its stake marked: O'Donnell, H. C. S. and Goodspeed, Surya, the Paris group, the Rebman journal and the BMJ via Wikipedia. So are Boddington's wartime and reputation statements and the absence-of-evidence analysis. No item was dropped because it was judged negatively. These pertinent items are missing, though:

• Missing from the arbiter and the primer (Investigator, priority 2: supporting information):
• Dr Drysdale Anderson bought Aurospecs in 1929, and a Daily Chronicle article of 14 March 1930 is said to confirm Kilner. This comes via Boddington, a seller.
• The Scatcherd and Baraduc "photographs" (Boddington, priority 3).
• The Rebman review's own caution about distinguishing "between sight and suggestion".
• The editorial's opera-glass and "no reading spectacles for 24 hours" anecdotes. These bear on the myopia paper used in the worked example.
• Boddington's "dicyanine cures short sight", which is the opposite direction to Kilner's myopia result. The Contrarian has it; the arbiter and primer do not.
• Missing from the arbiter and the primer (inventor's later work, 1920 edition):
• "Some women can change the colours of their auras at will (no man or boy)". This extends claim 6.
• The June 1915 repeat of experiments with the artist's model.
• The primer mentions only the ultraviolet hypothesis and the Wimshurst experiment.
• Source material favourable to Kilner that never reached the reviews. This is a coverage shortcut, not a negative judgement. It comprises the counter-expectation predictions, the pre-history readings (Cases 16 and 17), the hemoptysis-crepitation case and the "mother and nurse" observers (section 3). The Contrarian's "strongest honest case" is therefore weaker than the record allows. These items must be added beside the existing criticisms (no denominators, no blinding), not in place of them.

6. Process

• Source tags. Present and honest throughout. [memory] confidence is marked. The arbiter's web table is good.
• Page conventions are mixed.
• The Physicist uses printed pages.
• The Mathematician mixes the two conventions. For example, the cloud-banner and vapour argument is cited as "p.91-92" but is at PDF p.110-111, while "p.109" is a PDF page.
• The arbiter writes "p.131/PDF p.112" style double tags.
• The primer declares PDF numbering but carries "[read: p.91-92 as cited by the Mathematician]" (section 3.6). It should be PDF p.110-111.
• Web checks.
• Reviewers: the Physicist made about 2 web checks, the Mathematician 2 and the Contrarian about 3, all under budget. The medium-confidence facts that verdicts depend on (uranyl fluorescence, the lens UV cut-off) were web-checked by the arbiter.
• The BMJ review was attempted by three roles and stayed blocked.
• The dicyanin absorption band is still not web-confirmed. That is acceptable, because the verdicts rest on Kilner's own spectroscope reading.
• Verdict scale. The Mathematician uses an off-scale label, "Not supported by any controlled test" (claim 6). The arbiter's final verdicts are all on the scale and consistent across claims.
• Primer (STEP 4).
• Sections 1-9 are in order, with answers at the end.
• Symbols are defined: λ, D, vergence, M, f_e, NA, τ, ρ, the von Kries gains.
• The worked example and textbook suggestions suit the reader profile.
• The primer adds no verdicts.
• example.py. Runs cleanly in the venv. Its output matches the primer's tables (6.3 D per division at M = 25, f_e = 50 mm; Fisher p = 0.0119 and 0.00089). It needs the sibling recount added (section 4, item 4).

7. Remediation

Deep reads (most important first)

Changes to later files

• facts.json
• Note the OCR 3-to-8 case-number error and give the correct numbers.
• Add the colour-series counts (40, 36, 17, 7 of 100) and the epilepsy counts (10 of 10 left-sided).
• Add the pre-knowledge and counter-expectation predictions from deep read 1.
• Add the 1920 edition, Boddington and Annales as read entries after deep reads 2, 4 and 5 (marking Boddington as secondary and interested).
• Fix digest 03's page-label drift in any cited pages.
• claims.md
• Replace the "Case 88 / Case 21 appear twice" oddity with the OCR explanation.
• Mark "dead foetus" in claim 4 as attributed by Kilner to texture (Case 40), as the arbiter says.
• Add to claim 6 the 1920 statement that some women change colour at will.
• arbiter.md: add a "QE revision" section. Expected verdict changes: none, unless deep reads 1-2 show blinding or counts.
• (a) Correct correction 1: the duplicate numbers are OCR errors, not a continuation.
• (b) Section 2d: add the Table I sibling sets (9 of 13, or 9 of 16) and soften "the asymmetry favours a method effect".
• (c) Claim 4: re-score the young-woman case as a hit by Kilner's logic. Record that two predictions and Case 21 ran against the patient's expectation, and weigh this against the cueing objection.
• (d) Claim 3: add the colour-series and epilepsy counts and Cases 16 and 17 (no history taken), each with the unblinded-classifier caveat.
• (e) Relabel the cases: Case 18 → 13, Case 28 → 23, Case 80 → 30 (not pregnant), Case 82 → 32.
• (f) Add the Investigator items listed in section 5.
• primer.md
• Section 5b: drop "the one place in the book with counts" and add the sibling recount.
• Evidence table, claim 4 row: re-score the menstrual cases and fix the page citations (PDF p.253-254, p.317-318).
• Evidence table, claim 3 rows: add Cases 16, 17 and 37 and the colour and epilepsy counts.
• Section 7 "what is solid / open": add the counter-expectation predictions as the strongest items for claim 4, with their limits.
• Section 3.6: change "p.91-92" to PDF p.110-111.
• Correct the case numbers.
• Add the 1920 colour-at-will statement, Drysdale Anderson (seller-reported) and the "sight and suggestion" caution to the further-reading notes.
• example.py: add the sibling recount: 9 of 13 (chance about 0.024; Fisher one-sided 0.035) and 9 of 16 (Fisher 0.004).
• Investigator, next run (not a deep read). Retry the BMJ review (PMC2344221) in Chrome. Save the Archives of the Roentgen Ray v.16 pp.344-352 table as page images. Arbiter 2a depends on that table.

---

Recheck (STEP 4b, 2026-10-04)

Checked against the current claims.md, facts.json, arbiter.md (new section 11), primer.md and example.py, and against the new deep/ files. example.py was syntax-checked only. It needs sympy, which the system python lacks, and I did not find the venv, so I did not re-run it.

No retention problem: nothing pertinent is dropped because of a negative judgement. The arbiter's stated limits (BMJ review unread, Archives table unsaved, 1920 tables garbled) are disclosed accurately.

Final: PASS