Brown-Paul: Step 4b Quality Engineering (QE)

QE agent inputs: _text/manifest.json, both digests, facts.json, claims.md, physicist.md, mathematician.md, contrarian.md, arbiter.md, primer.md, example.py, and the extracted full text in _text/. For spot checks I also rendered PDF pages 4 and 5 of the Tesla Symposium bundle at 150 to 200 dpi into the session scratchpad (not the source folder) and read the images [read: page image p.4, p.5].

Note on process history: this project was reviewed before digests existed, so Step 1 worked from the raw text, not from digests, and no deep/ folder exists. The project is very small (two PDFs, 11 pages and about 42,000 characters in all), so every reviewer read the whole text, and all three rendered the patent and figure pages themselves.

Verdict: MINOR. The central verdicts are sound and well supported. The defects are in the Step 1 files (facts.json and claims.md were never corrected, though the reviewers and the arbiter found the errors), plus a few small missed points and one date slip. None of them would change a central verdict.

---

1. Coverage

• Every manifest document was digested, and nothing was skipped. No coverage defect.
• The patent body is not in the folder; only the front page is on PDF p.4. The project's one quantitative performance claim ("23 amperes at 400 volts") comes from the full patent text on Google Patents [web]. Three agents fetched it, and the arbiter confirmed it verbatim. This is a source gap, not a coverage failure. All later files label it [web].
• The digest of the bundle calls p.4 "primarily U.S. Patent 4,835,433 documentation" and says p.5 contains "Figure 6". Both are slightly off: p.4 holds the patent front page plus Figs 1 to 6, and p.5 holds Figs 7 to 12. The reviewers corrected this, so nothing downstream depends on the digest's error.

2. Shortcuts

• Step 1 did not render the patent and figure pages. facts.json says the patent text is "mostly unreadable OCR noise; I did not render page images" and that the figures are "largely garbled ... not rendered". That shortcut was a defect at Step 1, because Fig. 7 ("β Flux ≥ i²R") is the load-bearing equation of the whole review. The reviewers repaired it in Step 2: all three rendered pp.4-5, and the arbiter and primer use their readings. So it is no longer a defect in the verdicts. It is still a defect in facts.json (see section 6).
• No deep reads were run. For two short documents that every reviewer read end to end, this is acceptable. No central claim rests on a 1-3 sentence digest alone.
• No key document was only sampled. Both documents were read in full.
• Web shortcuts. The arbiter used its search budget up before it could verify the giant-dipole cross-sections, the (γ,n) thresholds, the Hanford inventory or accelerator costs. It deliberately built the Claim 4/5 verdicts so that they do not depend on those numbers (the S_n energy-balance argument). That is an acceptable way to handle a failed check, and it is disclosed.
• The "75 W from Sr-90 / 100,000×" figure and the Hanford "80 MCi" came from search summaries only. Both are flagged as unverified and are not used for any verdict. Acceptable.

3. Missed material

These are things in the documents that facts.json, claims.md, the reviews, the arbiter or the primer missed or under-used. Items 1 to 4 are worth adding. The rest are small.

1. Brown's own example of long-lived waste is plutonium, and plutonium is where his gamma idea works worst. In Wired he says "Some waste products have half-lives of 24,000 years" [read: Wired p.3]. That is the half-life of Pu-239 (about 24,100 y [memory, high]). A (γ,n) reaction on Pu-239 gives Pu-238 (about 88 y [memory, high]), not something "safe within a few months", and the competing photofission makes new fission products. The reviewers note that "all waste" fails for transuranics in general, but nobody points out that Brown's own headline example is a transuranic. This belongs in arbiter Claim 4 and primer Section 7.
2. The patent abstract makes a separate claim that nobody checked: the core is "a mixture of three radioactive materials ... [which] provides a greater flux of radioactive decay products than the equivalent amount of a single radioactive nuclide" [read: page image p.4]. Activities simply add, and mixing nuclides cannot raise the flux above the sum of their separate activities. Brown's own recipe also makes the point: 300 g of U and Th add only about 4% to a 1 mCi radium needle (example.py). This is a small, easy "Refuted by calculation" item. It also suits the primer, because the reader can check it.
3. Wired says "Brown showed that when nuclear waste is showered with gamma rays, it's transformed..." [read: Wired p.3]. The verb "showed" implies a demonstration. Only the contrarian quotes it. The arbiter and primer should say plainly that no demonstration, data or publication is on record (the same tests were only to be run "soon").
4. Output network in Figs 9 and 11. The rendered figures show the transformer secondary feeding a filter network of capacitors and inductors, with a voltmeter across the load and an ammeter in series [read: page image p.5]. A symbol in the middle of that network looks like a battery or a ground; I could not tell which at 150 dpi. The 10 kV start-up supply enters through a diode. The primer rightly warns that "volts × amps" without phase is not real power. A Sonnet deep read of p.5 should describe this network exactly, because it bears on how "23 A at 400 V" might have been mismeasured (open question 1 in the arbiter).
5. The paper's claim that Kr-85 is "environmentally the most acceptable radioisotope available for power production" [read p.3] was never examined. The primer's own budget needs about 113 atm of pure Kr-85, which is a pressurised radioactive gas. Small, but it would be instructive in the primer.
6. The economic claim (radioisotope batteries "competitive with chemical batteries for applications requiring lifetimes of over two years at remote locations") and the 10 to 5,000 mW target range [read p.3] are not assessed anywhere. These are mainstream-compatible statements, and the contrarian's point that the paper's own market needs only milliwatts touches on them. Peripheral.
7. Brown's 1987 reference title, "The Moray Device and the Hubbard Coil were Nuclear Batteries" [read p.3], is listed as a key claim in facts.json but dropped from claims.md and the reviews. It appears only as a citation in primer Section 9. For this reader it is useful context, because it ties Brown to the Moray and Hubbard free-energy lore. It deserves one sentence in primer Section 1 or 7, and no verdict.
8. Figure numbering: the paper's text calls Fig. 12 an economic "comparative analysis", but the sheet shows a photograph of a "RADIOACTIVE MATERIALS" container on a bench [read: page image p.5]. The physicist noted this. No action is needed beyond facts.json.

4. Accuracy spot checks

5. Process

• Source tags. Tags are present and honest throughout the reviews, the arbiter and the primer. [read: page image] is used correctly, and the primer marks figure readings it took second-hand from the reviewers ("as read by the reviewers"). [web] tags carry URLs, and search-summary-only items are labelled as such.
• Web checks. All three reviewers and the arbiter made web checks: the patent text twice and confirmed verbatim, plus Kr-85, Y-90, Cs-137, betavoltaics, giant resonance, Deseret 1988 and power-eng 2002. Failed checks are disclosed (cross-sections, S_n, Hanford, accelerator cost). The Moseley 1913 reference sits at [memory, high] in the primer. That is acceptable, since no verdict depends on it.
• Verdict scale. The arbiter normalises the reviewers' labels onto the Step 2 scale. The contrarian's off-scale "Not supported" (Claim 4) was absorbed into "Never properly tested". Split verdicts are clearly separated. One soft spot: Claim 3's thermal reading is labelled "Plausible, untested", although it is pure arithmetic about the isotope rather than a claim anyone could test. That is defensible, but the primer could say "arithmetic, not a device claim". No change is required.
• Tested / never tested / dismissed. The arbiter (Section 5) and the primer keep the contrarian's distinction visible. No suppression story is asserted without a source. The Idaho lawsuit and the INEL opinion are tagged as secondary [web].
• Primer against STEP 4 and the reader profile. All nine sections are present in order, with answers at the end. Symbols are defined (λ, A, Q, S_n, δ, μ₀, η). Acronyms (RTG, INEL, IECEC) are expanded. Numbers are worked step by step, and textbooks are named without guessed chapter numbers. Check-yourself answers 2, 3 and 5 were recomputed: f = 86.3 kHz, Q ≈ 108 ("≈110" is fine), 2 W, E_γ < 6.5 MeV. All correct.
• example.py. Runs cleanly with .venv/bin/python, and its assertions pass. Output: ceiling 1.776e-4 W, ratio 5.18e7, 54,000 Ci (54.7 kg) of Ra-226 needed, R_max 3.4e-7 Ω. These match primer Section 5.
• Step 1 files never corrected. The arbiter (Section 6) gives a corrected claim list and corrections to Step 1, but facts.json and claims.md still carry the old errors. This is the main process defect (section 6, actions A1-A2).

6. Defects summary

7. Remediation actions

Deep read (one; Sonnet). - telsasymposium1990paperbypaulbrown-74ab16, pages 4-5, rendered at 200 dpi or more. Question: transcribe the patent front page in full (bibliographic data, all references cited, abstract), and describe each of Figs 1-12: labels, equations, component values, and especially the output network and metering in Figs 9 and 11 (what the symbol between the filter capacitors is, where the meters sit, how the start-up supply connects and disconnects). Note where the paper's text and the figure numbering disagree. Output: deep/telsasymposium1990paperbypaulbrown-74ab16.md, tagged [read: page image p.N].

File changes. - A1. facts.json: fix D1 and D2 as listed above. Set how_read for the bundle to "full text layer + page images pp.4-5 (Step 2 reviewers and QE deep read)". Add the patent's 23 A × 400 V statement as [web], plus the "greater flux from a mixture" claim and the Fig. 7 inequality. Add the Wired byline, "Nuclear Solutions", the planned tests, the DOE talks and the 24,000-year quote. Mark 75 W/100,000× as unverified secondary. - A2. claims.md: replace the claims with the arbiter's corrected claim list (arbiter.md Section 6), keeping the Source gap note. Update that note to say the figures and patent front page are legible as images, and that the patent body is available only on the web. - A3. arbiter.md: append a "QE revision" section that (a) adds the Pu-239 point to Claim 4 ("all waste": still Never properly tested in general, and Refuted by calculation for Brown's own 24,000-year example if it is Pu-239, since (γ,n) gives Pu-238 at about 88 y; the half-lives need a web check); (b) gives the "mixture gives greater flux" patent claim a verdict of Refuted by calculation (activities add); (c) says that Wired's "Brown showed" is unsupported by any recorded demonstration; (d) records the deep-read description of the Figs 9/11 output network under open question 1. No headline verdict changes are expected. - A4. primer.md: in Section 4, add a row for the patent abstract's "mixture" claim. In Section 6, Claim 4, add the Pu-239 note. In Section 7, add the Kr-85 "environmentally acceptable" versus 113 atm point and one line on the 1987 Moray/Hubbard reference as context. If the deep read clarifies the output network, update the "Measurement method" row in Section 4. A short optional check-yourself question on adding activities in a mixture would fit the reader profile. - A5. physicist.md: no rewrite needed. The arbiter's QE revision can note the 23 Apr 1986 date. - A6. example.py: no change needed (runs, matches primer). An optional extra print line could show the U/Th contribution of about 4% to support A3(b).

After remediation, a Sonnet recheck should append a "Recheck" section here listing which of D1-D7 are fixed.

---

Recheck

Rechecked against the current files (facts.json, claims.md, arbiter.md, primer.md, physicist.md, example.py, deep/).

Defects

Remediation actions

• Deep read of pp.4-5 (Sonnet): done (deep/telsasymposium1990paperbypaulbrown-74ab16.md).
• A1 facts.json: done. The 75 W / 100,000x unverified item is not carried in facts.json. It is flagged in arbiter.md line 23 and is used for no verdict. Trivial.
• A2 claims.md: done.
• A3 arbiter.md QE revision (a) to (d): done. No headline verdict changes.
• A4 primer.md: done. This covers the mixture row, the Pu-239 note, the Kr-85 "most acceptable" versus 113 atm point, the Moray/Hubbard context in Section 1 and the Figs 9/11 measurement row.
• A5 physicist.md: no rewrite required. The correction sits in the arbiter.
• A6 example.py: runs cleanly (exit 0).

Remaining items are trivial and need no action (D7, and the 75 W note in facts.json).

Final: PASS