Compton: quality engineering (Step 4b)

QE agent: Opus. Date: 2026-10-03. Inputs checked: _text/manifest.json, digest/Compton-1923-fb4d86.01.md, facts.json, claims.md, physicist.md, mathematician.md, contrarian.md, arbiter.md, primer.md, example.py. Spot checks were made against the full OCR text _text/Compton-1923-fb4d86.txt (20 pages, all OCR'd by Step 0) [read: OCR text p.N]. My own numbers are tagged [calc] (project venv, numpy).

Verdict: MINOR

The paper's central claims and the arbiter's verdicts stand. The defects are stale or wrong statements in facts.json and claims.md that the reviewers corrected but nobody went back to fix, three small accuracy slips in the arbiter and primer, one quotation that is not in the paper, and two open figure-reading questions that one short page-image read would settle. None of these would change a final verdict.

1. Coverage

• The manifest has one document and it was digested in full. Nothing was skipped. No coverage defect.
• The OCR text is good for prose and numbers. It garbles the equations (as expected), and the digest copied some of that garbling (see Section 4).

2. Shortcuts

• Step 1 sampled the key document. It read 9 of the 20 page images and wrote facts.json and claims.md from those. The prompt counts sampling part of a key document as a defect. In practice the defect was fixed downstream: every reviewer rendered and read most pages, and each says so with page tags. The verdicts do not rest on the sampled reading.
• The stale record was never updated. facts.json still says how_read: "page images", text_available: "none", and lists pp. 5, 7, 8, 10-12, 14-16, 18, 19 as "not individually viewed". claims.md opens with the same out-of-date source-gap note. The primer (line 8) also still says "image-only scan with no text layer". Step 0 has since produced a full OCR text, so all three statements are now out of date. Defect (minor).
• No deep read was done. For a single 20-page paper read page by page by four agents, I accept this as a shortcut. The one place where it shows is the figures (Fig. 1B, Fig. 4, Fig. 7). Reviewers disagree about these or left them unresolved, and a single focused page-image read would settle them (see the remediation list).

3. Missed material

Below are items in the paper (checked against the OCR text) that facts.json, claims.md, the arbiter or the primer missed or under-used.

1. Compton reversed his own published position. In the footnote on p.3 [read: OCR text p.3], Compton says that in 1921 (Phil. Mag. 41, 749; Phys. Rev. 18, 96) he had "defended the view" that the softening of secondary X-rays came from an admixture of fluorescent radiation. He notes that J. A. Gray (1913, 1920) and Florance (1914) had argued instead for true scattering, and he concedes "the latter view is the correct one". The Discussion (p.19) also drops his own earlier "large electron" hypothesis as "superfluous". Only the contrarian mentions the reversal. The arbiter's "contrarian lens" and the primer leave it out. For this collection it matters: the researcher followed his data against his own earlier theory, and two earlier workers (Gray, Florance) were partly right before him. Add to facts.json, the arbiter's contrarian section and primer Section 7.
2. The recoil electrons from ordinary X-rays had not been observed. Compton admits this on p.14 [read: OCR text p.14]. The contrarian and the primer have it, but facts.json does not.
3. Other evidence Compton used to anchor the 0.022 A primary gamma wavelength (p.14) [read: OCR text p.14]: his own value of 0.025 A from small-angle scattering of gamma-rays by lead (Phil. Mag. 41, 777, 1921), and Ellis's beta-ray line spectra (0.045, 0.025, 0.021, 0.020 A, with 0.020 the strongest). The physicist and contrarian mention the Ellis lines. The primer's Step 6 calls the 0.022 A value simply "self-calibrated". That is fair, but the primer should add that Compton gave two partly independent cross-checks.
4. The citation of de Broglie (J. de Phys. 3, 33, 1922) for the lambda^3 fluorescent-absorption law (p.14 footnote), and Livens's moving-radiator result (p.8 footnote, which Compton says neglects the concentration of quanta). These are minor history and fine to omit. Listed for completeness.
5. The circle on Fig. 6 (total gamma-ray absorption in carbon, attributed wholly to scattering, p.16). Not discussed by anyone. It is a minor data point.
6. Debye's independent 1923 derivation of the shift is not in the paper. The reviewers mention it as [memory: medium] or as a search summary. The primer does not name Debye. Optional: one line in primer Section 7 or 9, tagged.

No claims, experiments or numbers of central importance were missed.

4. Accuracy spot checks (against the OCR text)

example.py ran cleanly with the project venv. SymPy derives h*(1 - cos(theta))/(c*m). Every number in primer Section 5 matches the script output: 0.02426 A, 0.7323 A, 580 eV, beta 0.0476, 44.03 deg, the momentum check 24.3587 keV/c both ways, and the Part C table 0.9359/0.9371/0.9058/0.9072 ... 0.3119/0.4132/0.2046/0.2650.

5. Process

• Source tags. Present throughout and honest. [read: page image p.N], [web] with URL, [memory] with confidence, and [calc] are all used consistently. The arbiter says plainly that its search budget ran out and lists what remained unverified (Duane, Debye, Ishino data quality, Compton's Phys. Rev. 22 follow-up, Ross). It correctly notes that no verdict depends on these.
• Web checks. Physicist 3, contrarian 3, arbiter 3 (budget exhausted). Each medium-confidence fact that a verdict depends on was checked: the 1925 Bothe-Geiger and Compton-Simon tests and the end of the Bohr-Kramers-Slater (BKS) proposal, the Klein-Nishina formula, NIST XCOM for aluminium, and the unshifted line. The 1927 Nobel Prize check failed at nobelprize.org (HTTP 403) but is supported by the Wikipedia page cited. Acceptable.
• Verdict scale. Mostly consistent. One awkward label: the arbiter gives Claim 3b (the RaC gamma Table I data as evidence) "Plausible, untested in this form", after saying itself that "untested misdescribes data that exist but are weakly probative". A cleaner split would be: Claim 3b is not a claim about physics at all but about the strength of 1923 evidence, so describe it in words ("weak indirect test; formula Established by later work") rather than force a scale label. Minor; fix it in the QE revision.
• Arbiter internal slip. In the Claim 5 arbitration, the arbiter writes "90 deg intensity 19% low in Compton's favour of lower, i.e. KN is 1.23 times larger". The sentence is garbled; the numbers (0.0697 vs 0.0859) are right. Rewrite it.
• The arbiter's Fig. 7 open question is real. The contrarian says the points lie above the curve near 60 deg and below it from 70 to 150 deg. The mathematician says they lie below over 60-150 deg. The arbiter left it unresolved, and the primer asserts only the single quoted 90 deg datum, which is safe. A deep read of the page image will settle it.
• Typo claims not yet verified on the page. Eq. 1's plus sign: the arbiter and the primer disagree on how Fig. 1B marks theta. The primer zoomed in and concluded the angle shown is the scattering angle. Eq. 23's R vs R^2: the OCR shows "2 pi R^2" in the first line of Eq. 23 and "16 pi R" in the last line, where a superscript may simply have been lost. So the "typo" may be an artefact of reading the scan. The primer does not repeat it, which is good. The arbiter accepted it without re-reading. Settle both with a page image.
• The primer follows Step 4 and suits the reader. All nine sections are present, in order. Symbols are defined (nu, beta, gamma, alpha, r_e, sigma, dsigma/dOmega, P). The units change (Gaussian to SI) is flagged. Klein-Nishina is introduced as a benchmark without requiring quantum field theory. The worked example has real numbers, and the answers are at the end. Accuracy slips: 1. Section 1 says Compton found the scattered X-rays' shift was larger "at larger angles". In this paper the X-ray measurement is at 90 deg only. The angle dependence is shown only for gamma-rays, by the indirect absorption method. Reword. 2. Section 7 says Compton's guess "fitted soft X-ray data to a fraction of a percent". That is not what the paper shows. It agrees with Klein-Nishina to a fraction of a percent at small alpha. The actual soft X-ray data fit only "perhaps within the experimental error" (Hewlett, p.16), and Barkla-Ayers disagrees beyond error (p.17). Reword as "agrees with the later exact theory to a fraction of a percent". 3. The source-gap paragraph (line 8) is out of date (see Section 2).
• example.py runs. Yes (see Section 4).

6. Remediation actions

Deep reads (most important first)

1. Compton-1923-fb4d86, pp. 18-19 (Sonnet, page image at 200 dpi or more). Fig. 7: at each plotted angle, does the experimental circle sit above or below the Eq. 27 curve and the classical curve? Is the 90 deg point about 0.037 I_0, against the curve's about 0.07 I_0? Is "beautiful agreement" a fair description of the plot? Report approximate read-off values.
2. Compton-1923-fb4d86, pp. 4-4 and 10-10 (Sonnet, page images, zoomed). (a) Fig. 1B: which angle is marked theta, the scattering angle or 180 deg minus it? Is the sign in Eq. 1 printed "+"? (b) Eq. 23: is the denominator in its final reduced line printed "16 pi R^2" or "16 pi R"? Settle whether either is a real print error.
3. Compton-1923-fb4d86, pp. 13-14 (Sonnet, page image of Fig. 4 plus OCR text of Table I). Does the Fig. 4 scattered spectrum show any residual peak at the primary 0.708 A position (an unshifted line)? This bears on the "only a small part, if any" statement. Also confirm the Table I column headings (mu/rho, tau/rho, lambda observed, lambda calculated) and values.

facts.json

• Fix the ratio of energy removed from 1/(1+2 alpha)^2 to 1/(1+2 alpha).
• Update how_read to "OCR full text (Step 0) + digest; all 20 page images read across Steps 1-3", and text_available to "yes (OCR; equations from page images)". Remove the "not individually viewed" list.
• Replace the quoted "little or none" with the paper's wording: "only a small part, if any".
• State that the gamma-ray wavelengths were inferred from lead absorption, not measured with a spectrometer, and include the full Table I observed/calculated pairs.
• Add: Compton's reversal of his own 1921 fluorescence view (citing Gray and Florance); his abandonment of the large-electron hypothesis; his admission that recoil electrons from ordinary X-rays had not been observed; and the 0.037 I_0 Fig. 7 datum.

claims.md

• Replace the source-gap note: the full OCR text now exists and all page images have been read. What is still not obtained is the cited papers (Compton's Phys. Rev. 22 follow-up, Ishino, Hewlett, Barkla-Ayers, Bothe-Geiger, Compton-Simon, Klein-Nishina).
• Claim 3: say that the gamma-ray values are inferred, and give 0.068 observed against 0.063 calculated.
• Claim 4: say that (1+2 alpha)^2/(1+alpha) applies to the truly scattered part, and that the total energy removed is lower by (1+2 alpha).

arbiter.md: add a "QE revision" section

• Correct the "little or none" quotation (Claim 3 and corrected claim 6).
• Rewrite the garbled Claim 5 sentence about the 90 deg intensity (KN is 1.23 times Eq. 27).
• Record the outcome of deep reads 1 and 2 (the Fig. 7 position of the points; whether the Eq. 1 sign and Eq. 23's R^2 are real print errors). Delete "typo" from the researcher-fault list if the page shows R^2.
• Reword the Claim 3b label: say "weak indirect 1923 test; formula Established by later work" rather than "Plausible, untested in this form". No other verdict changes.
• Add Compton's self-reversal (fluorescence, large electron) and Gray and Florance's priority to the "Contrarian lens" section.

primer.md

• Update the source-gap paragraph (line 8).
• Section 1: remove "more so at larger angles" for X-rays, or attribute the angle dependence to the gamma-ray absorption data.
• Section 7: change "fitted soft X-ray data to a fraction of a percent" to "agrees with the later exact theory to a fraction of a percent at small alpha; the 1923 soft X-ray data were only roughly matched (Hewlett within error; Barkla-Ayers off beyond error)".
• Section 7 (or a short box): add Compton's 1921-to-1923 reversal and the earlier Gray and Florance view as a learning point about following the data.
• Step 6: mention Compton's two cross-checks on 0.022 A (his own 0.025 A from lead, and Ellis's beta-ray lines).
• If deep read 1 finds the Fig. 7 points sit consistently below the curve, say so. If not, keep only the single-datum statement.

example.py

• No change needed.

Recheck

QE recheck (Sonnet), 2026-10-03. Checked the current facts.json, claims.md, arbiter.md, primer.md, deep/ and example.py against the defects and actions above.

Deep reads

1. Pp. 18-19, Fig. 7: Fixed. deep/Compton-1923-fb4d86.md exists; the arbiter's QE revision records the result (points on or above Eq. 27 near 44-60 deg, 55-75% of it at 90-150 deg; 90 deg datum about 0.041 read, 0.037 in text, against about 0.07).
2. Pp. 4 and 10, Fig. 1B angle, Eq. 1 sign, Eq. 23 R vs R^2: Remains. No deep read was done. facts.json and claims.md still list it as "not yet done", and the arbiter calls both questions "unconfirmed". The arbiter's handling is honest (not counted as researcher faults), but the check was not performed.
3. Pp. 13-14, Fig. 4 unshifted line and Table I headings: Remains. Same status. The Table I values were checked against OCR text by QE, but the figure and column headings were not read from the page image.

facts.json

• Energy removed corrected to 1/(1+2 alpha): Fixed.
• how_read and text_available updated: Fixed.
• "little or none" replaced by "only a small part, if any": Fixed.
• Gamma wavelengths stated as inferred; full Table I pairs: Fixed.
• Self-reversal, large-electron drop, recoil electrons unobserved, 0.037 I_0 datum: Fixed.

claims.md

• Source-gap note replaced: Fixed.
• Claim 3 (inferred gamma values, 0.068 vs 0.063): Fixed.
• Claim 4 (factor applies to true scattering; energy removed lower by 1/(1+2 alpha)): Fixed.

arbiter.md

• "QE revision" section added: Fixed.
• "Little or none" quotation corrected: Fixed in the revision and the Claim 3 text. Remaining mentions are the correction notes.
• Garbled Claim 5 sentence: Fixed (0.0697 vs 0.0859, KN 1.23 times larger).
• Fig. 7 outcome recorded: Fixed.
• Eq. 1 / Eq. 23 outcome: Remains, depends on deep read 2 (the "typo" is already removed from the researcher-fault list, which is the conditional action QE asked for).
• Claim 3b relabelled in words: Fixed.
• Self-reversal and Gray/Florance priority in the contrarian lens: Fixed (added under "Added to the contrarian lens").
• Minor leftover: the arbiter's opening source-gap line still says "no text layer"; the revision notes it is out of date. Cosmetic, not counted.

primer.md

• Source-gap paragraph updated: Fixed.
• Section 1 angle-dependence wording: Fixed.
• Section 7 "fraction of a percent" rewording (agrees with Klein-Nishina; 1923 data only roughly matched): Fixed.
• Self-reversal and Gray/Florance box: Fixed.
• Step 6 cross-checks (0.025 A from lead, Ellis lines): Fixed.
• Fig. 7 statement: Fixed (the deep-read result is stated).

example.py

• No change needed; reruns cleanly with the project venv and reproduces the primer numbers: OK.

Final: OPEN (2 issues)

Recheck 2

QE recheck 2 (Sonnet), 2026-10-03. Checked the current facts.json, claims.md, arbiter.md, primer.md and deep/ against the two issues the first recheck marked as remaining.

1. Deep read 2 (pp. 4 and 10: Fig. 1B angle, Eq. 1 sign, Eq. 23 R vs R^2) and the arbiter's Eq. 1 / Eq. 23 outcome: Fixed. deep/Compton-1923-fb4d86.md now holds page-image reads of p.4 (220-500 dpi) and p.10 (300 dpi). Result: Fig. 1B's theta is the interior (scattering) angle, so Eq. 1's printed plus is a real print error that does not propagate; Eq. 23's final line prints 16 pi R, a real print slip, not an OCR artefact. arbiter.md records this in items 3-4, in the researcher-fault list (typos, not researcher faults) and in "Round 2". facts.json, claims.md and primer.md carry the same conclusions.
2. Deep read 3 (pp. 13-14: Fig. 4 unshifted line, Table I headings and values): Fixed. The deep file reads Fig. 4 and Table I from page images. Result: no resolved unshifted peak, but not excluded (poor resolution, no error bars); headings and values confirmed. These are reflected in facts.json, claims.md, arbiter.md ("not resolved", not "absent") and primer.md Step 6 and Section 5. The arbiter states that no QE deep reads remain outstanding.

Residual, not counted: facts.json and claims.md say Fig. 5 (p.15) and the p.17 forward/backward passage have had no dedicated page-image read. They were not among the actions QE required, and no verdict depends on them.

Final: PASS