TernionAG: Quality Engineering (Step 4b)

QE run 2026-10-03 (Opus). Inputs read: _text/manifest.json (118 documents), all 149 digest files, the deep/ reads I needed to check (GB417501, US2089966 and others), facts.json (118 entries), claims.md, physicist.md, mathematician.md, contrarian.md, arbiter.md, primer.md, and example.py, which I ran. I also checked these against the extracted text and against three page images: GB417501 pp.10-11, the two TIFFs inside the skipped .docx, and the Guardian PDF.

Tags: [read] from the extracted text or a page image, [calc] my own arithmetic, [web] with URL, [memory] with confidence.

Verdict: MINOR

The central claims, the verdicts and the primer's worked example stand. None of the defects below would change a final verdict in arbiter.md. The fixes are a few targeted deep reads, additions to facts.json and claims.md, a "QE revision" section in arbiter.md, and small additions to primer.md. The main open process gap is web verification, which failed again in Steps 2 and 3 because the search quota was used up. QE could not close it either: WebSearch was still out of quota, and one WebFetch attempt failed.

---

1. Coverage

• Manifest against digests. The manifest lists 118 documents and 149 chunks, and there are 149 digest files. No digest is missing and there are no extra digest files [checked by script].
• Skipped files:
• Ternion-AG/press/JMPEE-Vol22-Pg49-Voss.pdf: byte-identical duplicate. Acceptable.
• audio/fwoimperialwarmuseum/notes.txt: 34 characters, just the IWM collections URL. Acceptable.
• ternion-docs-request/20160228/Coorespondance-20170228.docx: defect (small). It was skipped as "no readable text (0 characters)", but the file holds two embedded TIFF screenshots (about 2 MB each) with plenty of readable text. I OCR'd them. They are a 28 Feb 2017 follow-up letter saying the register copies had not arrived, and the Glarus register's reply (Silvia Aschwanden, deputy head) saying the documents were posted on 30 Jan 2017 and would be re-sent by e-mail and post [read: OCR of word/media/image1-2.tiff]. The content is only logistics, so no claim is affected. The skip reason in facts.json is still wrong, and Step 0 should OCR images embedded in .docx files.
• Partly extracted files:
• related-news/guardian-sonic-attack.pdf: defect (small). The facts entry says "body text not extracted". The PDF has a header-only text layer, so Step 0 never OCR'd the body. Rendering at 150 dpi and running tesseract recovers about 976 words: Hannah Devlin, The Guardian, 25 Aug 2017, "How could the 'sonic attack' on US diplomats in Cuba have been carried out?". It covers ultrasonic-weapon speculation and says chemical causes must first be ruled out [read: OCR p.1]. This is peripheral (it bears only on claim 12) but should be digested. Step 0 should OCR any page whose text layer is only print headers and footers.

2. Shortcuts

3. Missed material (in the digests or source, absent from facts, claims, reviews, arbiter or primer)

1. Kassner's "term differences" are the fine-structure splittings of the hydrogen atom. GB417501 p.11 l.80-84 says reactions between hydrogen, chlorine and similar substances may emit radiation at "the term differences of hydrogen, which correspond to wave-lengths of 2.74, 9.25 and 27.75 cm". It gives the most effective reaction bands as 2.5-3.5, 8.6-9.5 and 26.5-30.5 cm [read: page image p.11]. These are exactly the Sommerfeld/Dirac fine-structure intervals for hydrogen, ΔE = α²·Ry/(n³·l(l+1)): n=2, l=1 gives 0.3652 cm⁻¹ (2.74 cm); n=3, l=1 gives 9.24 cm; n=3, l=2 gives 27.72 cm [calc; formula consistent with https://en.wikipedia.org/wiki/Fine_structure [web]]. Each of Kassner's three reaction bands brackets one of these wavelengths. So the "most effective frequencies" for chemistry were most likely taken from atomic theory, not measured, and the reviews treat type-(3) frequencies as having "no example". This is new evidence on method: Kassner's numbers can be traced to textbook calculations (as with the λ ≈ 3 × diameter rule). It also gives the primer a good learner exercise. It does not change a verdict, but it should go into facts.json, the arbiter's notes on claims 2 and 3, and primer section 4 or 7.
2. The 1937 shareholder list implies other shareholders. At the 20 Dec 1937 meeting only 54 of the 100 shares were represented: Stoffel 48 (capped at 10 votes) and Hoffmann 6 (OCR "AS" is 48, since 54 − 6 = 48) [read: ternion-artifacts-original-Ternion-AG-2 l.326-328]. Hoffmann held 40 shares at founding, so at least 34 of his shares had moved, and 46 shares belonged to people who were not present. The minutes also refer to general meetings 1-4 on 17 Jul 1936. facts.json, claim 8 and the arbiter do not mention the change of ownership. It is an ownership-records lead for the Investigator, and it bears on who financed the Fr. 1.12 million.
3. Bauch sold a radiation-therapy patent to Marconi. The 1996 memoriam says that while at GEC (1934-35) Bauch sold "An Improved Method of Radiation Therapy" to the Marconi Company for £80 [read: memoriam digest p.1]. facts.json has it, but claims.md, the arbiter (claim 9) and the primer do not. It is the only documented technical output of Bauch's GEC stay, and its subject (medical radiation) overlaps Kassner's therapeutic applicators. It is not proof of a Ternion link, but it is the best concrete lead for claim 9 (search the British patent records for a Bauch application of 1934-36).
4. Morray note 11 argues that "coarse sweep steps" explain why the resonance was missed. The note says standard frequency sweeps use steps wider than a 2 % line. None of the reviews answers this, and the contrarian's strongest case should have. It is easy to refute by calculation: a 2 % line at 333 MHz is 6.7 MHz wide, and a 1601-point network-analyser sweep from 0.1 to 3 GHz has 1.8 MHz steps, so about 4 points would fall across the line [calc]. The 1930s Drude-type and modern broadband measurements do not resolve it this way either. Add this to the arbiter's claim 1.
5. Details of the GB417501 food experiments. The milk warmed only from 20 to 23 °C; the veal from 20 to 24.5 °C; the juice from 22 to 24.5 °C. The milk and veal were kept fresh "still longer" only by re-exposing them for 5 minutes every hour (3.2 W and 3 W) [read: page image p.10]. So the 28-day and 2-week results include continuing treatment. The arbiter's "2-8 °C" range should give these exact figures. The ongoing hourly exposure, and the fact that the storage temperature is never stated, belong in the claim 3 evidence. The measured 3 K rise in the milk fits 574 W × 120 s into 2 kg (8.2 K upper bound, with losses) [calc].
6. Two values for water's band. GB417501 p.10 l.52-54 calls 1.28 cm "characteristic of water". The same patent family gives water's friction band as "about 1.5 cm". The arbiter's "the patent attributes 1.28 cm to serum (an oddity)" reads US2089966 p.8 only. The GB text instead shows two different wavelengths for water in one specification. This is a small internal inconsistency to note under claim 1.
7. Kassner's places of residence. He gave his address as Berlin (GB383508, 1931-32), Arbon (US2089966, 1932), and London as "a German subject resident in London" (GB448417, GB448494, US2109843, 1934; US2104697, 1936) [read: digests]. facts.json has these, but the arbiter and primer do not use them. A London residence in 1934, the year of Bauch's GEC secondment, is relevant to claim 9.
8. The images folder includes an "Elektro Spezial GmbH, Berlin" catalogue page. This is the Philips subsidiary Bauch founded after 1935 [read: images digest p.1-2; Swiss Sound p.4]. facts.json covers it, so this is only a primer or arbiter context gap, not a defect.

4. Accuracy spot checks (5)

5. Process

• Source tags. All review files carry tags. Tag counts ([read]/[memory]/[web]): physicist 23/28/1, mathematician 21/18/1, contrarian 33/19/2, arbiter 3/12/6, primer 24/13/3. Every file says honestly that the web checks failed. No [memory] fact is presented as [read]. Untagged textbook statements in the primer are covered by its blanket note ("textbook material [memory, high]"), which is acceptable.
• Web checks. Not done (process defect, carried over). Step 1 did no web verification. Steps 2 and 3 tried and failed: WebSearch was at 200 of 200, and only two Wikipedia pages were fetched. Still unverified [memory] items: water's relaxation time τ = 8.3 ps (high confidence; also derived independently in [calc]), serum conductivity of about 1.5 S/m (medium), the non-thermal microwave and Grundler/Keilmann literature (medium), electret charge ranges (low), and the 1945 Goodrich/Firestone purchase beyond the clippings. QE's own attempt also failed (WebSearch quota still used up; the LSBU water page timed out). The arbiter rightly says no verdict depends on any of these alone. The serum σ figure does underlie the "Q of order 1" argument in primer answer 3. Retry once the quota is raised.
• Verdict scale. Applied consistently, with three small issues: 1. Claim 12's "BAE statement Refuted (a factual error)" uses a word that is not on the scale. Call it a factual error outside the physics scale, as was done for claim 8. 2. Claim 9's "Ternion project: Never properly tested" applies a test-based verdict to a historical claim. Better wording: "Uncorroborated (single source), history; outside the physics scale". [Superseded 2026-10-03: that label breaks the "No suppression of evidence" rule; the wording now used is "documented in one published source (1996 JAES memoriam); no second source found yet; the other accounts are silent on it, not contrary" (arbiter.md, Evidence-rule revision).] 3. Claim 7's "purchase reported: Established" is fine as a press record.
• Primer. It follows the STEP 4 order (sections 1 to 9, with answers). Bessel functions, Q, Debye relaxation and state functions are all introduced from scratch. Symbols are defined, and the worked example uses real numbers. It fits the reader profile.
• example.py. It runs with the project venv and prints exactly the numbers in primer section 5 (32.68 cm, 100.8 Ω, 9.01 µJ, x = 1.047, ε_eff 3.09 / 69 %, Debye Q 0.29, 3879 × kT).
• Secondary sources. The three Morray-TH notes are labelled secondary everywhere and none is used as evidence. Good.

6. Remediation actions

Deep reads (most important first; 4 requested)

1. GB417501 complete specification, pp.9-12 (page images), Sonnet. Transcribe every quantitative example from the page images: the gold sol (18.5/18.6 cm, 0.1 W); agar (10 W); milk, veal and juice wavelengths, powers, temperature rises and the hourly re-exposure; beans; the three electret recipes and the pastilles; the hydrogen term differences and reaction bands; and the claims beyond claim 4. Also: does any example name a storage temperature, a sample count or a control? (This closes the shortcut on central claim 3.)
2. GB483062 pp.8-12, Sonnet. Do the later embodiments or claims 16-35 give any measured disc or plate dimensions with their wavelengths, or any statement that revises or explains the 1932 "λ ≈ 3 × diameter" figures? Are there any efficiency or power figures? (Bears on claim 4.)
3. Ternion AG register dossier: ternion-artifacts/original/Ternion AG 2.pdf pp.6-11 and Ternion AG 3.pdf pp.6-9 (page images), Sonnet. Confirm the 1937 attendance figures (48 + 6 = 54) and whether any proxy holders or other shareholders are named. Find out what the minutes say about the 17 Jul 1936 meetings 1-4. Determine who made up the 100 % of capital represented in the 1952 universal assembly. List every name for the Investigator's ownership search. (Bears on claim 8.)
4. OCR and digest the two files Step 0 under-extracted, Sonnet (mechanical). These are Coorespondance-20170228.docx (embedded TIFFs) and related-news/guardian-sonic-attack.pdf (rendered pages). Write the digests and correct the facts entries.

facts.json and claims.md

• Correct the how_read and text_available fields for the two under-extracted files (see deep read 4).
• Add to the Jacobsen and micro-records entries: "full-text grep of the extracted text by QE: no Kassner, Ternion or Bauch".
• GB417501 entry: add the hydrogen fine-structure identification (item 3.1), the food-experiment temperature rises and hourly re-exposure (item 3.5), and the 1.28 cm water value (item 3.6).
• Ternion AG 2 entry and claim 8: add that by 1937 Hoffmann held only 6 shares and 46 shares were unrepresented, so ownership had changed (item 3.2).
• Claim 9: add Bauch's £80 "Improved Method of Radiation Therapy" sale to Marconi and Kassner's London residence in 1934-36 as context (items 3.3 and 3.7), clearly marked as not evidence of a Ternion project.

arbiter.md: add a "QE revision" section

• Claim 1: answer the coarse-sweep argument with the calculation in item 3.4, and note the two water wavelengths (1.5 and 1.28 cm). The verdict does not change.
• Claims 2 and 3: the type-(3) frequencies and reaction bands are the hydrogen fine-structure intervals (calculated, not measured). Give the exact temperature rises and the hourly re-exposure in the food tests. The verdicts do not change. "Never properly tested" stands, with a note that the claim 3 examples involve continuing treatment.
• Claim 9: reword the verdict to "uncorroborated single-source history (outside the physics scale)", and list the radiation-therapy patent sale as a lead. [Superseded 2026-10-03: see the note under Section 5, Verdict scale, item 2.]
• Claim 12: replace "Refuted" for the BAE statement with "factual error (outside the scale)".
• Web verification: still not done. List the [memory] items to retry (Section 5).

primer.md

• Section 4 or 7: add a short box, "Where Kassner's numbers come from". His chemistry wavelengths 2.74, 9.25 and 27.75 cm are the hydrogen fine-structure intervals α²·Ry/(n³·l(l+1)), just as his 1932 disc figures follow λ = 3D. This is a useful second lesson on checking whether a number was measured or derived.
• Section 6, row 3b: note "temperature rises of 2.5-4.5 °C measured; milk and veal kept fresh only with hourly re-exposure".
• Check-yourself (optional): add a question asking the reader to compute 2.74 cm from the fine-structure formula.

Step 0 tooling (for future projects)

• OCR images embedded in .docx files when the text layer is empty.
• OCR PDF pages whose only text is print headers and footers.

Leads for the Investigator (sites.md)

• British patent records 1934-36 for an F. W. O. Bauch application on radiation therapy (sold to Marconi).
• Swiss Official Gazette of Commerce (SHAB) 1932-1952 notices for Ternion A.-G. and Dipol A.-G. Arbon.
• Shareholder names from deep read 3.
• Kassner's London address and any British company records from 1934-36 (for example, Universal Inventions Ltd).

Recheck

Rechecked 2026-10-03 against the current arbiter.md (with its new "QE revision" section), claims.md, facts.json, primer.md, example.py and the new deep reads (deep/ocr-underextracted-files.md, GB417501 pp.9-12, GB483062 addendum, Ternion AG 2/3 ownership section).

Open

1. GB417501 pp.9 and 12 have not been checked on page images. Low risk, because the print is clean and no value on them feeds a verdict.
2. Web verification of the [memory] items listed above, which needs the search quota to be raised.

Neither open item changes a verdict, so the QE verdict stays MINOR.

Final: OPEN (2 issues)

Evidence-rule check (2026-10-03)

Checked arbiter.md (all sections, including the Evidence-rule revision and "The record goes dark") and primer.md against the three SOURCE HONESTY rules in prompt.txt, then grepped claims.md and qe.md for the dismissal labels the rules forbid. arbiter.md and primer.md are compliant on the main points: the Bauch/Ternion project is quoted and weighed as one published source with silence elsewhere not read as contradiction (arbiter.md:101-105; primer.md:150); the Goodrich description is reported as an interested party's, with its stake, beside Kassner's own words (arbiter.md:81-91; primer.md:121-135); the silent record is analysed with six explanations, evidence for and against each, and deciding records (arbiter.md:128-154; primer.md:139-156); evidence against suppression and evidence not excluding it are both stated (arbiter.md:117-119). Remaining violations are stale wording in files the revision did not reach, plus one overstated absence:

1. Dismissive label, claims.md:28. Claim 9 still ends "Status: uncorroborated single-source history, outside the physics verdict scale", and calls the London-residence and Marconi-sale context "not evidence of a Ternion project". Rule 1 forbids reducing the memoriam sentence to "uncorroborated"/"single-source". Replace with the arbiter.md:104 wording ("documented in one published source (1996 JAES memoriam); no second source found yet; the other accounts are silent on it, not contrary") and describe the context as consistency, not a second source (as arbiter.md:103 does).
2. Stale dismissive label, qe.md:60, 87 and 128. The QE's own suggestion and Recheck table still prescribe and record claim 9 as "Uncorroborated (single source)" / "uncorroborated single-source history", which arbiter.md:263 shows was superseded. Line 128 reports that wording as the current state of the verdict scale, which is now false. (Historical lines; they need a superseded note, not deletion.)
3. Interested party and label, claims.md:23. Claim 7 states that Goodrich and Firestone "bought basic electronic-vulcanisation patents" with no mention that the buyers' 1945 and 1950 descriptions come from parties with a stake (arbiter.md:81-91, primer.md:121-135), and it reduces Voss to "Voss's comment is hedged", a label that reads as dismissal; arbiter.md:78 quotes both the hedge ("sketched") and the plain statement ("foresaw"). Add the stake and quote both.
4. Absence treated as neutral, arbiter.md:125 and primer.md:117. "No replication attempt exists in the folders, so there are no failures to explain" (arbiter) and "no replication attempt of any kind is on record, so there are no failures to explain" (primer). The primer's "on record" is broader than the searches done (Internet Archive, Europe PMC, UK National Archives, SHAB; sites.md), and the rule asks that a patented process cited in 1935-1955 literature with no recorded test or replication be treated as an anomaly with competing explanations (never attempted; attempted and unpublished; trade secrecy; the patentee's instruments not reproducible), not as an empty gap. Scope the primer sentence to "in the folders and searches so far" and add one line listing the explanations and the records that would decide (replication attempts in Goodrich or Firestone files, Arnold v. Watson exhibits).

Final: OPEN (4 issues)

Evidence-rule recheck (2026-10-03)

Rechecked claims.md, arbiter.md, primer.md and this file against the four items of the Evidence-rule check above, then grepped all four files again for "uncorroborated", "single source", "hedged", "no failures to explain" and "replication attempts exist". The only hits left are historical before/after rows and change-log lines in arbiter.md (216, 263, 277) and the superseded QE lines here, each marked superseded.

Not in the list, noted for the next pass and not changed here: - arbiter.md explanation 4 under "The record goes dark" still has "Against: one source only so far". It reads close to the label the rule forbids, although "so far" and the next clause ("silent on it (not contrary)") keep it within the weighing. - arbiter.md claim 9 dates GB448419 "filed 31 Aug 1934" and then calls GB448418/GB448419 "British filings of 31 Aug 1933". The page-1 header shows Convention Date (Germany) 31 Aug 1933 and UK Application Date 31 Aug 1934 [read: _text GB448419 p.1]. "Filings of 31 Aug 1933" should read "priority 31 Aug 1933".

No numerical value was corrected in this recheck.

Final: PASS

Investigator update check (2026-10-04)

Checked the "Investigator findings (2026-10-04)" sections of physicist.md, mathematician.md, contrarian.md and arbiter.md, the in-place updates in arbiter.md, primer.md, sites.md (leads 1-29) and research/, against STEP 4b (Retention, Process) and SOURCE HONESTY. Method: read the three review sections and the arbiter section; keyword-traced every lead and named item into arbiter.md and primer.md; grepped the new sections for dismissal labels ("uncorroborated", "single source", "hearsay", "fringe" and similar); recomputed the load-bearing numbers; ran example.py with the project venv.

What holds: - Source tags are present and honest in the new sections. [web] items are marked "as the Investigator recorded them"; the arbiter's one new check (blood at 300 MHz, eps about 65, sigma about 1.3 S/m) is tagged as a snippet and as blood rather than serum; unverified [memory] items (water tau at 0 C, electret charges, Debye-Falkenhagen) are listed as such. Web use was Chrome, one search, new tab closed. - No dismissal labels in the new material. The remaining hits for "single-source" in arbiter.md (lines 234, 281) are the history of the superseded claim 9 label. - Interested parties are handled: Goodrich 1950 and the 1960 MRI/Navy survey are reported as parties with a stake, beside the inventor's own text. Evidence against is stated (Gandhi 1980, Gos 1997, Mittler 1976, the 1930s German dispute) as well as for (Heller, Arnold, Webb and Booth, Grundler and Keilmann). Silence in unread pages of FR 788,428 and FR 786,294 is not read as absence (arbiter ruling 1). - The record-goes-dark analysis was updated, not reduced to a label. - Recomputed and agree: 0.337 GHz crossover (89 cm), 0.36 GHz (83 cm) for blood, Debye half-height ratio 13.9 and Q 0.29, 2 percent band 173 times narrower, 1/5 us = 0.2 MHz = 0.74 percent of 27 MHz, 6.7e4 frequency range, 1.7e-4 eV at 41.7 GHz (150 times below kT), Heller heating 0.005 K at 1e-4 S/m, skin depth about 1 m at 27 MHz, 0.18 of the research account for Fr. 200,000. example.py runs with .venv/bin/python (system python3 lacks mpmath; the header gives the venv command). - Verdict tables agree between arbiter.md and primer.md for claims 1a-1c, 3, 4d, 7, 8, 12, 13, 14, 15.

Defects:

1. Wrong number in arbiter.md, and arbiter and primer disagree. arbiter.md line 339 (and mathematician.md line 232, the physicist's source for it) says a 10 K local temperature error near 100 C changes a rate "by a factor 1.3 for E_a = 80 kJ/mol [calc: exp(E_a/R(1/363 - 1/373)) = 1.30]". The expression equals 2.04. example.py step 8d prints 2.04 and its header says it corrects the 1.3; primer.md line 176 correctly says 2.04. Fix: arbiter.md line 339 (and the mathematician's line) to 2.04, with the 2026-10-04 correction noted. The point (small temperature errors make apparent non-thermal effects) is strengthened, not changed.
2. Retention: lead 14 left out of arbiter.md and primer.md. The physicist (D2), mathematician (4a) and contrarian (lead 1, 14) all record that Vibranic's FR 926,351 and CH 250,508 carry the same priority dates as Grell/Richards US 2,415,025 and CH 254,655 (assigned to Firestone and Goodrich), so the Fr. 200,000 Nov 1945 purchase from Grell included Grell's foreign heating-patent portfolio, and whether any Kassner patent was in it is still open (sites.md lead 14). arbiter.md claim 7 and "Investigator findings" item 4 give only "Fr. 200,000 ... 0.18 of the research account"; neither arbiter.md nor primer.md mentions FR 926,351, CH 250,508, US 2,415,025 or CH 254,655. It is pertinent to claim 7 and to "continuation under other names". Fix: add one paragraph to arbiter claim 7 (and its Investigator item 4) and one to primer Section 4.3, stating the matched priority dates, that they show Vibranic's purchase was at least Grell's heating rights, and that no document yet shows a Kassner patent in it.
3. Retention: reviewer-recorded items dropped by the arbiter. Not in arbiter.md or primer.md: ES 135,727 (Spanish member of the energy-content family; physicist and mathematician); Archangelsky (physicist B2, with Schliephake and Compere and Kowarschik) and Krasny-Ergen 1936-37 (physicist B1, beside Muth 1927 and the Maxwell-Wagner explanation, which the arbiter does keep as a bare term); and sites.md lead 13's other citers of GB417501, "Jonas (seed treatment by HF fields, 1950) and later cell- and productivity patents", which no reviewer weighed and the arbiter does not mention. The Jonas seed patent bears directly on claim 3 (seeds). Fix: add them to arbiter "Investigator findings" (one line each, stated as claims by others with no data seen), and have the arbiter note that no reviewer weighed Jonas.
4. Retention: primer.md omits items that arbiter.md keeps. Absent from primer.md: Weissberger, Thorpe and Ellis as citers of GB417501 (Chemical Abstracts is kept); the 1929 Elektro-Physikalische Gesellschaft "Ernst Kassner" and the 1923-28 "Ernst Kassner" adding-machine patents (identity unknown); the Bodleian SPSL file "W Kaeser - E Kassner"; Goodrich Series VII (the series to ask Akron for); the Swiss police dossier references (E4264); the Stadtarchiv Arbon tax rolls; the Proclaimed/Statutory List check; the closed items (1941 SHAB hit, US 2,190,668, "Vibranic sterilizing canned foods"). None is a negatively judged item, but the primer is the reader's document and the rule says condensing steps keep pertinent material. Fix: add a short "Unread or unidentified leads" list to primer Section 6 or the further-reading block (line 314 area), one line each.

Neither defect 3 nor 4 changes a verdict; defect 1 changes no verdict; defect 2 adds context to claim 7 (the purchase verdict, Established as a press and trade record, is unchanged).

Final: OPEN (4 issues)

Recheck

Rechecked 2026-10-04 against the current arbiter.md, the three reviews, primer.md and example.py.

No new dismissal labels were found in the added text. No verdict changed.

Final: PASS

Investigator update check (2026-10-08)

Checked 2026-10-08 against prompt.txt (STEP 4b, SOURCE HONESTY), the "Investigator findings (2026-10-08)" sections of physicist.md, mathematician.md and contrarian.md (read in full), arbiter.md (same section), primer.md, sites.md leads 30-40 and the follow-up notes' citations as quoted by the reviewers. No new research note dated 2026-10-08 exists in research/.

Checked and sound: - Every verdict change (claim 11 split in three or four parts, claim 10 wording, claims 1-4 notes) in the three reviews is carried into arbiter.md with a ruling where reviewers disagreed (Established vs "tested, not replicated" for the catalyst rate effect; the 35,000 and 0.2 vs 0.02 V/Angstrom errors). - Recomputed and agree: 1 V/Angstrom = 1e10 V/m, so 2e9 V/m = 0.20 V/Angstrom and Saitta's 0.35 V/Angstrom = 3.5e9 V/m. The arbiter is right and the mathematician's "correction" (D5, E2) is wrong; the primer carries the arbiter's version. IR photon ratio 3,407-4,749 and O-H 49,555 (example.py 9c); Q = 1.14 and 0.96; 1.28 cm = 23.4 GHz, tau 6.8 ps; Carnot 0.20; example.py runs (project venv) and step 9 prints the numbers the primer quotes. - Primer tables 11a-11e match the arbiter's verdict table; the 2026-10-07 review-direction wording of claim 11 is in both; interested parties (Serra patent, Roy/Rao/Kanzius as claimants, Meyer court finding as the court's characterization) are labelled; single-source Kanzius evidence is described, not dismissed.

Defects:

1. Retention: Puharich dropped; Meyer dropped from the primer. Physicist C8, mathematician C3, contrarian lead 35 and sites.md lead 35 all keep Puharich US 4,394,230 (1983, "complex electrical wave form" water splitting, not read) and the mathematician adds that Puharich's 1987 compilation names "22.3 GHz" as water's resonance (the vapour line). Neither arbiter.md nor primer.md mentions Puharich. Meyer's "water fuel cell" (Ohio court 1996) appears in arbiter.md only as a parenthesis in the verdict table, with no description and no source tag, and not at all in primer.md. Fix: one line each in arbiter "Investigator findings" and primer Section 4.6 "related work", tagged [web: Wikipedia snippet via Investigator], stated as claims by others (Puharich unread), Meyer as the court's characterization.
2. Retention: evidence for and against dropped from arbiter and primer. (a) Mathematician B2: the Institute of Science in Society author writes of the Kanzius effect "The phenomenon is real, but the mechanism is entirely unknown" and that the 13.56 MHz photon carries about 1e-8 of the energy needed per molecule; an anonymous reader comment calls the Raman spectra similar apart from "a minor difference in peak intensities". Single-source and from an advocacy outlet, but it is the only third-party description of the observation and it cuts both ways; neither arbiter nor primer quotes it. (b) Stovall et al., EES Catal. 2026 "questions the model" of the second Wien effect (evidence against) is in sites.md lead 37 and the reviews, but absent from arbiter.md and primer.md (the primer cites only Onsager and Bui). (c) Ambient radio as a second reservoir (0.1-10 microwatt/cm2; about 660 m2 of perfect collector for 6.6 W; 19-23 GHz tank does not collect broadcast bands) is in all three reviews but not in arbiter.md. Fix: add one sentence each, with sources and the stake of the outlet.
3. SOURCE HONESTY: priority claims overstated in primer.md and example.py. primer.md line 83 says the water-vapour line was "first measured in 1946" and line 334 calls Cleeton and Williams 1934 "the first microwave absorption spectrum"; example.py line 225 says "measured from 1946". The contrarian's web check says only that the measurements he could find date from 1946 and "I found no source saying these were the first observations, so priority is unverified"; the arbiter repeats "earliest line measurement found". Fix: "earliest measurement found: 1946 [web, priority unverified]" and drop "first" for Cleeton and Williams unless a source is cited.
4. Stale copies of corrected errors (CORRECTIONS ARE WELCOME). The errors the arbiter confirmed are still uncorrected or unannotated at their source: "35,000 times" in sites.md lead 34 and physicist.md C5 (the physicist's own "Corrections" list does not mention it, and its verdict summary repeats "35,000 times"); the mathematician's wrong 0.02 V/Angstrom correction (mathematician.md D5, E2) and its "17 times below" claim; sites.md lead 31 still presents the vapour line as an inference without the 1946 dating. The primer says openly that the old figure persists in those places. Also claims.md claim 11 lacks the 2026-10-07 review-direction wording (standing field by displacement current, radiated RF), which the arbiter left to the orchestrator. Fix: add a dated correction note at each place (or edit lead 34, C5, D5/E2, claim 11) and re-render.
5. Minor inconsistency. arbiter.md ruling 4 says "use about 3,500-4,700" and the corrected list "about 3,500-4,700"; the calculation (and the mathematician, primer, example.py) give 3,400-4,700 (3,407-4,749). Use one range.

None of these changes a verdict. Defects 1 and 2 are retention (small); 3 and 4 are source-honesty and stale-copy fixes; 5 is a nit.

Final: OPEN (5 issues)

Recheck (2026-10-08)

Rechecked against the current arbiter.md, physicist.md, mathematician.md, primer.md, example.py, sites.md and claims.md (file times: arbiter, mathematician, physicist and primer edited after the check; sites.md last edited 2026-10-07, claims.md 2026-10-03).

No verdict changed. No new dismissal labels found in the added text.

Final: OPEN (1 issue)

Orchestrator fix (2026-10-08)

• sites.md lead 34: correction note (3,400-4,700, not 35,000). Lead 31: correction note (earliest vapour-line measurement found is 1946, after Kassner's 1932 patent).
• claims.md claim 11: 2026-10-07 review-direction wording added (standing field by displacement current, or radiated RF through a wand applicator). The recheck's one open issue is closed.

Final: PASS