QE: Ainslie-Rosemary-Ann

Step 4b quality check. Inputs read: _text/manifest.json, all 7 digests in digest/, facts.json, claims.md, physicist.md, mathematician.md, contrarian.md, arbiter.md, primer.md, example.py. Spot checks were made against the extracted text in _text/. There is no deep/ folder: Steps 1-4 ran before digests and deep reads existed.

Verdict: MINOR. The central experiment (the 2002 Quantum article and its 2009 reprint) was read in full. The reviewers and the arbiter checked its numbers carefully, and example.py runs and reproduces them. No verdict on claims 1-6 is likely to change. The main defect is that the 43-page patent bundle was treated as unreadable. Step 0 has since OCR'd it fully. Its body holds three of Ainslie's own experiments (Examples 1-3) with measurement methods that the review never looked at. They should go into facts, the arbiter and the primer as extra, instructive evidence. The Source gap note also needs correcting: two of the three "no text" PDFs now have OCR text.

---

1. Coverage

All six manifest documents were digested and none was skipped. Every manifest document has a facts.json entry.

2. Shortcuts

• Patent bundle (defect). facts.json, claims.md, the arbiter and the primer all say the patent body could not be read and that COP 1.086 and 2.45 are "known only via Kelly's digest ... not used". The OCR text (_text/harnessing-a-back-emf-...txt, pp.19-32) is legible and gives the full method and numbers. These are Ainslie's earliest experimental over-unity claims. A central experimental claim of the author was left resting on a secondary digest when the primary text was available. This needs a deep read.
• Magnetic field model (acceptable after digests). Claim 5's verdicts (charge accounting, mass-ratio arithmetic, untestable zipons) rest on pp.1-20, and those pages were read. Pages 21-37 (gravity, Earth's core, fire, black holes, 10 dimensions) are peripheral. A short deep read of the mass-ratio passage is still worthwhile, because the author's own arithmetic differs from the reviewers' summary (see spot check 5).
• Kelly pp.7-21 (acceptable). These pages are mostly Kelly's digest of the patent plus his own "self-powering" lamp and transformer variant and a build guide. They are peripheral, but see missed items 3 and 6.
• Central experiment. No shortcut was taken. The physicist, mathematician and contrarian all worked from the full reprint text and the Quantum page images.

3. Missed material (from the digests, confirmed in the text)

1. Patent Example 1 compares two different resistors [read: patent OCR pp.19-22]. The control circuit uses a 68 ohm resistor at 95 C. The test circuit uses a 22 ohm resistor (with an inductor and MOSFET) at 93 C. Equal temperatures are taken to mean equal power. Battery power in the test is (average current) x (battery voltage), with average current 0.1718 A and power 2.1097 W. The claimed gain is 8.6%, and that already includes a 5% error allowance (2.406 / 2.215 = 1.086). Nobody reviewed this method.
2. Patent Example 2 also compares different resistors and drops a battery [read: patent OCR pp.23-30]. The control is a 100 ohm resistor and the test a 47 ohm resistor ("identical surface area", per the text at OCR line ~1018). The test circuit has two 6 V batteries in parallel. The battery-2 shunt shows ±0.0055 V peaks across 0.05 ohm, which is about ±0.11 A of peak current, six times battery 1's 0.0175 A on-peak. The author "discounted" that whole waveform as a "small margin of gain". COP 2.45 comes from counting battery 1 only (0.025 A x 6.12 V = 0.153 W against 0.374 W). This is a clear and teachable bookkeeping error, and the review missed it.
3. Patent Example 3 mixes average and RMS [read: patent OCR pp.31-32 and Kelly p.14]. Battery power = V x average current. Load power = RMS voltage x RMS current (OCR lines ~1093-1104). For any non-constant current, I_rms ≥ I_avg, so this mixed method can show "COP > 1" with no anomaly. The table gives 0.976 at 50% duty, rising to 1.176 at 90%. A simple rectangular pulse would have I_rms / I_avg = 1/√D, which is largest at low duty, so the trend is not that simple form factor. The load current also includes the freewheel path through the diode, which does not pass through S1. The deep read must decide which artefact applies. Either way, power was computed in two inconsistent ways on the two sides of the comparison. This would make a good worked example for the primer, and the review never mentions it.
4. Patent dates. The priority application is ZA 98/0541, filed 22 January 1998, and the bundle is dated 19 January 1999 [read: patent OCR p.1, line ~25 and ~1394]. facts.json and the primer call it "the 1999 patent". Kelly gives WO 99/38247 filed 22 Jan 1999. Both dates should be stated.
5. Patent duty-cycle claims are 50-99% on (preferably 75-95%) at 40 Hz or more. This is the opposite regime from the 3.7% / 1.3% heater circuit, and nobody noted it. Kelly's 55% "runaway" figure is closer to the patent regime.
6. Kelly's other claims [read: Kelly pp.2, 15, 18-19]: "prototype outputs exceeding four times input", "COP 4 to 17" for Ainslie techniques, and a self-powered lamp/transformer/battery variant. These appear in facts only in part. They are advocacy, but claims.md should list them as third-party claims.
7. People. Donovan Martin (2009 co-author) and Brian Buckley (2002 co-author and technician) appear unevenly. John Marriott (Sasol) and Prof Mike Bruton are in facts. J De Bruto (Power Engineers) and Jonathan Green (BP) are named in the reprint but only loosely in the reviews. This is minor.
8. Where the 2009 appendix places the energy. It says the load energy "comes from the load" (bound fields in the resistor) and that delivered energy returns to the battery [read: reprint p.18]. The arbiter caught this. The primer should keep it.

4. Accuracy spot checks

5. Process

• Source tags. Present and mostly honest (physicist 22, mathematician 18, contrarian 26, primer 20 tags). The arbiter says plainly which web checks failed or were based on search summaries only (Poynt99, the Lettenmaier COP > 4). That is good practice.
• Page citations do not match the PDF pages. Several files cite "Quantum p.7" or "p.8", but the PDF has 5 pages. These are probably journal page numbers. Cite PDF pages, or say that journal pages are meant.
• Stale how_read. Quantum ("no OCR") and the patent ("no text layer") are now wrong because the manifest lists 5 and 43 OCR pages. The Source gap paragraph in claims.md, arbiter.md and primer.md repeats the stale statement.
• Web checks. Medium and low-confidence items were checked or flagged as failed (arbiter used 4 of its 8 web calls). The IRFPG50 Qg was taken from a search summary, and the arbiter says so. Acceptable.
• Verdict scale. Applied consistently, with split verdicts explained. The arbiter's reasoning on "Refuted by calculation" vs "Never properly tested" is sound and was applied the same way in the primer table.
• Primer. Follows STEP 4 sections 1-9 with answers. Symbols are defined, the worked example uses real numbers, and the reader profile is respected. The patent experiments are missing (see above).
• example.py. Runs cleanly (exit 0) with the venv. It reproduces the arbiter's steady-state table, the energy bound, the shunt sensitivity and the Qg x f check. The analytic and time-stepped results agree.

6. Remediation actions

Deep reads 1. Patent bundle harnessing-a-back-emf-id00000000130311-575f51, pp.19-32 (Sonnet). Extract Examples 1-3 exactly: circuits, resistor values, how each power was computed, and every number. Check the battery-2 (SR3) treatment in Example 2 and the average vs RMS method in Example 3. Read the figures on pp.40-42 as page images if the OCR is unclear. 2. Same document, pp.1-18 and 33-38 (Sonnet). Get the dates (priority ZA 98/0541, 22 Jan 1998), the claimed duty and frequency ranges, the "efficiency 90% or more" wording, and what the claims actually assert. 3. magnetic-field-model-4e21ac, pp.17-22 (Sonnet). Quote the author's mass-ratio calculation exactly, including the 1837.5 − 1.5 step and the electron and proton composition. 4. quantum-october-2002-ec98f3, pp.2-5 (Sonnet). Verify against the OCR: how the 0.039 A was obtained, the exact sample-range wording, and the journal-page vs PDF-page mapping for citations.

File changes (MINOR route) - facts.json: update how_read and text_available for the patent and Quantum (OCR now available). Rewrite the patent summary from the original text, covering Examples 1-3, the differing control and test resistors, the discounted battery 2, and average vs RMS. Add the 1998 priority date. Add Example 3's COP table. Note Kelly's "4x" and "COP 4-17" as third-party claims. - claims.md: correct the Source gap note: the patent and Quantum have OCR text, and the remaining gaps are the raw data, company reports and replications. Add a sub-claim to claim 1 or 2 for the patent bench tests (COP 1.086, 2.45, up to 1.18). - arbiter.md: add a "QE revision" section. It should give a verdict on the patent examples (likely Internally inconsistent: different resistors, an omitted battery, average-vs-RMS mixing; and net gain Refuted by calculation). It should change "0.039 A inferred" to "method not stated", note the author's 1837.5 − 1.5 arithmetic slip under claim 5, and fix page citations. No headline verdict should change. - primer.md: update the Source gap paragraph. Add a short subsection or check-yourself question on the average-vs-RMS mix in patent Example 3 (V·I_avg on one side, V_rms·I_rms on the other), using whichever mechanism the deep read confirms. Add the patent examples to the evidence table. Cite the patent as filed 1998 (ZA) / 1999 (WO). - example.py: optionally add a few lines comparing V·I_avg with V_rms·I_rms for a simulated RL-plus-freewheel waveform at 50-90% duty, next to the patent's Example 3 ratios.

---

Recheck

Recheck of the QE defects and remediation actions against the current files (deep/ now holds three reads; facts.json, claims.md and arbiter.md were modified after the QE; primer.md and example.py carry their pre-QE timestamps).

Defects and missed material

Remediation actions

Open issues

1. primer.md was not updated: stale Source gap paragraph, no patent Examples 1-3 content, evidence table and check-yourself questions unchanged, patent still described as unread and dated "1999".

Final: OPEN (1 issues)

---

Recheck 2

Re-examined the one issue left open by the first recheck, against the current primer.md (modified after Recheck 1).

Note, not counted as an open issue: the primer's header still says example.py "reproduces every number in Section 5", while the patent subsection says example.py does not reproduce those numbers. This is a small wording inconsistency worth fixing in a later pass.

Final: PASS