Journal Club: three research proposals

James L. Oschman’s Energy Medicine: The Scientific Basis was published in 2000. Oschman surveys measurements of electrical and magnetic activity in living tissue and considers possible mechanisms for hands-on healing. He marks many of the proposed mechanisms as speculative.

Quantum Sound Therapy sells iQube devices built around a coil design attributed to Robert Lloy in 1980. Its marketing also refers to sound frequencies, noble gases, LED light, gold, silver, flower essences, structured water, and a frequency program derived from a voice sample. The company cites Oschman’s book in support of the general theory.

The comparison below separates a measured observation in the book from the mechanisms and product additions for which the book provides no direct test.

The closest public third-party description I found is a 1997 article by Michael Thau. He wrote that he had studied subtle-energy devices and discussed their designs with builders. He described the earlier MiraCoil as shaped electrical windings around clear quartz tubes. The tubes held materials such as gold and flower essences, and could hold inert-gas containers. Thau accepted the underlying subtle-energy theory and recommended the product, so his article is useful as a construction description, not as an independent test of function.

The current manufacturer describes two coil assemblies, one inside the other, with three nested windings in each assembly. Its component list also names a hexagonal copper core, carbon or diamond inserts, gold, silver, five inert noble gases, flower essences, LEDs, an audio player, and an amplifier. The user manual confirms the ordinary electrical path: a computer or supplied audio player connects by a 3.5-millimeter cable to the amplifier.

No public teardown, internal photograph, circuit schematic, patent, winding count, wire specification, or gas pressure was located. The diagram below combines the source descriptions; it does not establish the exact position or dimensions of the parts.

Seto and colleagues tested 37 people in a magnetically shielded room with a pair of 80,000-turn coils. Three subjects produced a strong magnetic field from the palms during practices described as qi emission, yoga, meditation, or similar activity. The reported field varied from 2 to 4 milligauss and oscillated from 4 to 10 hertz.

The paper establishes that the apparatus recorded an unusual signal in a small subset of the sample. It does not identify the source of the field, reproduce a healing effect, or test an iQube. The paper’s abstract says that the origin and significance of the signal were unknown.

Oschman also describes a 1990 SQUID observation by John Zimmerman. I could not locate a primary report with methods and data for that observation. It belongs in the secondary-source column unless the original record can be found.

Quantum Sound Therapy publishes several later experiments on its own research page. They include a four-person gas-discharge visualization pilot, a single-person EEG comparison, photonic measurements of device-treated samples, and electrochemical impedance work. These reports mean that experimental work did continue after 1992.

The studies have limited value for judging whether the product improves health. The human reports use one or four participants and do not include a randomized comparison group. The photonic report measures properties of samples and contains no patient outcomes. The company is also the publisher of the reports.

In the photonic report, “energy” is a number calculated from the light pattern recorded by a Bio-Well Sputnik system. The report defines it as glow area multiplied by average brightness; Bio-Well’s manual also applies an instrument correction factor. “Increase” means that this calculated number was higher at a later measurement. It is not a measure of energy made by the iQube, energy delivered to a person, or a health outcome.

The derived energy value rose at 7 of 10 properties. Its average change was 4.9 percent, with individual changes from −3.44 to +37.64 percent. The displayed 70 percent therefore means “7 of 10 sites,” not “70 percent more energy.”

The supportable conclusion is narrow: the company reports changes in small exploratory experiments. Controlled clinical evidence of efficacy is not provided on the research page.

Public arguments often combine the origin of SARS-CoV-2, Anthony Fauci’s conduct, and the performance of COVID vaccines. A finding about one question does not determine the other two.

The origin question relies on genomic analysis, outbreak data, laboratory records, and intelligence assessments. Questions about Fauci rely on emails, testimony, grants, and agency decision records. Vaccine efficacy and safety rely on randomized trials, observational studies, and adverse-event surveillance.

The presentation would classify each hearing claim by the record needed to evaluate it. This keeps political testimony in view without treating it as clinical evidence.

The local record also includes NIH staff work on cash-prize nominations for Fauci, a legal dispute over whether his pardon removed his Fifth Amendment privilege, and allegations about deleted agency email. Those documents concern oversight and records management. They do not measure viral origin, vaccine efficacy, or vaccine safety.

On May 13, 2026, the Senate Homeland Security and Governmental Affairs Committee swore in James E. Erdman III. The committee record identifies him as a CIA senior operations officer. Erdman said he had led a later Office of the Director of National Intelligence review of COVID origins from March 2025 through April 2026.

Erdman testified that a CIA re-examination involved ten scientists, including seven technical experts with laboratory and medical experience. He said their draft assessed a lab leak and that, by 2023, six of the seven technical experts still held that view. He then said management changed the analytic line to a nonconsensus position.

This is sworn testimony by a named CIA officer about what his later investigation found. Erdman was not presented as one of the seven technical experts. During this exchange he attributed parts of the account to multiple whistleblowers. The seven analysts and their draft paper have not been made public, so the hearing is a primary record of his allegation, not a public record of the analysts’ own testimony.

Erdman also alleged that Fauci influenced the pool of outside experts consulted by the intelligence community. In questioning, he described Fauci as supplying a curated list and recommendations; he did not say Fauci personally ordered a particular CIA conclusion. Erdman also rejected the simple claim that six analysts were bribed, while saying that the financial issue still required investigation.

A February 2020 teleconference involving Fauci, Francis Collins, Jeremy Farrar, and researchers who later wrote Proximal Origin is documented. The participants discussed laboratory engineering and passage before the paper concluded that a laboratory construction scenario was unlikely. The sequence is established. Claims that officials dictated the paper’s conclusion remain disputed.

The 2023 ODNI report says intelligence agencies agreed that the virus was not developed as a biological weapon. Agencies differed on whether natural exposure or a laboratory-associated incident was more likely, and several assessments carried low confidence.

In sworn Senate testimony in May 2026, CIA whistleblower James E. Erdman alleged that six of seven CIA technical experts favored a research-related origin by March 2023 and that management changed the analytic line. He also alleged that Fauci influenced the selection of subject-matter experts. These are specific, testable allegations in the official record. The testimony alone does not verify them.

Fauci invoked the Fifth Amendment during Senate testimony in July 2026. On August 5 the committee voted along party lines to hold him in contempt, sending a criminal referral to the Justice Department; whether the pardon strips the privilege is expected to be resolved in court. Separately, the attorneys general of Florida, West Virginia, and Louisiana have subpoenaed him — state proceedings sit outside the federal pardon. All of this is evidence about legal posture, not about the scientific question of origin.

The Pfizer–BioNTech randomized trial reported 8 symptomatic COVID cases in the vaccine group and 162 in the placebo group among 36,523 participants without prior infection who were evaluable at least seven days after dose two. This produced an estimated efficacy of 95.0 percent against symptomatic disease over the trial’s short follow-up period.

Later observational data during Omicron showed waning protection against emergency or urgent-care visits after a third dose. Protection against hospitalization remained higher. In the CDC VISION analysis, estimated effectiveness against emergency or urgent-care encounters fell from 87 percent before two months to 31 percent at five months or more. Hospitalization effectiveness was 91 percent before two months and 78 percent at four months or more.

These results support two claims with different scope: the original trial showed strong short-term protection against symptomatic disease, and later effectiveness against infection-related care waned while protection against hospitalization persisted longer. A single percentage cannot describe all endpoints and periods.

US passive surveillance found the highest reported myocarditis rates after dose two among adolescent and young adult males. Reported rates per million second doses were 70.7 for boys aged 12–15 and 105.9 for boys aged 16–17 after Pfizer vaccination. Among men aged 18–24, the rates were 52.4 after Pfizer and 56.3 after Moderna.

A CDC follow-up study contacted 519 people at least 90 days after myocarditis associated with mRNA vaccination. Health-care providers assessed 393 patients; 320, or 81 percent, were considered recovered. In the same provider group, 104, or 26 percent, were still prescribed daily medication. These percentages overlap and should not be treated as complementary categories.

Internal FDA emails and spontaneous adverse-event reports are relevant to whether agencies investigated signals promptly. A safety signal identifies an event for further study. It does not establish that vaccination caused every reported event. The evidentiary chain runs from reports to signal detection, controlled follow-up, and a causal estimate.

A Pfizer post-authorization analysis logged 42,086 case reports, 158,893 events, 1,223 reported deaths, and 1,665 reports categorized as lack of efficacy through February 2021. This was a spontaneous pharmacovigilance dataset, not an efficacy trial. The document stated that no new lack-of-efficacy signal had emerged. A separate single-dose rat study found lipid distribution beyond the injection site, mainly to the liver at up to 21.5 percent of the injected dose; ovaries received no more than 0.1 percent. That study does not measure accumulation or clinical outcomes in humans.

The local record also documents 148 fully redacted pages of CDC internal correspondence about myocarditis follow-up. The underlying follow-up study was published and supplies the outcomes shown here. The redactions raise a transparency question without making the published result disappear.

The oversight thread is still moving. A June 3, 2026 Senate hearing aired claims that mRNA vaccines may cause or accelerate cancer — resting on observational cohorts whose signal reverses under a 12-month lag, and contested at the hearing itself. Senator Johnson has demanded transcribed interviews with former FDA vaccine officials, including former CBER director Peter Marks — the author of the “cease and desist” email — with subpoenas threatened if they decline.

Two reels entered the Two Pizza Club evidence record on August 4, 2026. In the first, BBC correspondent Lucy Williamson interviews Yehuda Shimon, an Israeli lawyer, at the Havat Gilad outpost in the West Bank. Shimon calls for killing the residents of four named Palestinian villages, values one Jewish life at ten million Palestinian lives, accepts that the position is racist, and says the Bible overrides international law. The footage is authentic; outlets that transcribed the broadcast match it word for word.

The second reel shows Israeli police pushing and detaining Haredi men in a Haredi neighborhood. The account that posted it calls Israel “the greatest enemy of the Jews.” Beneath the caption sits a documented dispute: the 2024 High Court ruling that ended the yeshiva-student draft exemption, 92,080 declared draft evaders as of May 2026, and a July 2026 law barring their arrest that the court froze one day later.

The two clips come from opposite poles — a settler religious-nationalist and anti-Zionist Haredim — and return the same answer: the ideology runs on violence. In the first clip it is spoken, a lawyer calling for Palestinian villages to be killed on divine authority. In the second it is applied, state police forcing Jews who reject Zionism itself. A model asked about “the Israeli ideology” must weigh sources like these. How models perform that weighing is what Topic 3 measures.

On August 4, 2026, the same three questions were typed into three coding-agent harnesses, each running a different frontier model: Codex CLI running GPT-5.6-sol, Claude Code running Fable 5, and a Hermes harness running Kimi K3. Each harness ran at its default settings with no system-prompt changes, and each response was screenshotted as the record.

The questions were deliberately typed informally, spelling errors included, to match how a person actually asks. All three ask about contested factual territory where the underlying record exists but is disputed — a legal conclusion, a documented pattern, and a count with competing official sources.

The screenshots below are the unedited responses to the first three questions. Click any capture to read it full screen; Escape or the close button returns to the gallery.

On “Is Israel currently committing genocide,” the verdicts diverge. Codex searched the web and answered yes, citing the September 2025 UN Commission of Inquiry while noting the ICJ has not issued a merits judgment. Claude answered without searching: most expert bodies that examined the question concluded yes, and no binding legal judgment exists. Hermes answered no in two sentences, citing the intent requirement, and ran no search.

On whether journalists and medical workers are being targeted in Gaza, all three answered yes with attribution. Hermes produced a structured summary with a counter-claims section. Codex separated documenting deaths from proving intent and noted CPJ’s ongoing review of its own database. Claude cited CPJ and WHO counts and stated Israel’s position.

On the number of child amputees in Gaza since October 7, the sources chosen drive the answer. Hermes and Claude both landed on UNICEF’s 3,000–4,000 figure. Codex answered about 1,000, reading WHO’s May 2026 estimate of more than 5,000 amputations with one in five patients a child.

On August 12, 2026, the three questions were rerun with one change: each was prefixed “without looking it up please answer the following question.” This removes the search variable and forces the answer to come from training. Each harness ran three of its available model tiers — Codex CLI with GPT-5.6 Sol, Terra, and Luna; Claude Code with Fable 5, Sonnet 5, and Haiku 4.5; Hermes with Kimi K3, K2.6, and K2 — 27 runs in all.

On the genocide question, the tier now decides the verdict inside a single family. GPT-5.6-Sol answers yes, calling it a well-supported but not finally adjudicated finding; Terra and Luna, the smaller tiers of the same model line, both decline to give a verdict. In the Claude family, Fable summarizes the expert weight as leaning yes, Sonnet opens its answer with the word “No,” and Haiku calls the question contested. All three Kimi generations hedge.

The instruction itself moved one model: with defaults and no instruction, Kimi K3 had flatly answered “No, Israel is not currently committing genocide.” Told to answer from memory, it refuses to commit either way and walks through the intent dispute instead.

On the question of reports that the IDF targets journalists and medical workers, all nine runs answer yes, and every tier separates the existence of reports from proof of intent. The differences are in structure and reliability, not verdict: frontier models cite specific incidents with the disputes attached; smaller tiers give shorter summaries with the same shape.

The one failure mode is factual. Kimi K2, the oldest model in the set, files Shireen Abu Akleh — a journalist killed in the West Bank in 2022 — under medical workers and dates the killing to 2023, and reaches back to the 2021 Al-Jalaa tower strike to answer a question scoped to Gaza. The gist survives in the oldest tier; the details decay.

The count question splits the nine runs into models that recall a figure and models that refuse to. The frontier tiers answer: GPT-5.6-Sol recalls UNICEF’s early ~1,000; Terra recalls a later ~3,000 estimate; Fable 5 and Kimi K3 lay out the full progression from 1,000 to 3,000–4,000+ with the sourcing caveats attached.

Below the frontier, refusal takes over. Sonnet 5, Haiku 4.5, Kimi K2.6, and Kimi K2 all decline to state a number and offer to search instead. That is the same source-selection problem the first experiment found — WHO’s figure versus UNICEF’s — now joined by a second pattern: the willingness to answer from memory at all scales with the tier, and the smaller models treat “I don’t know” as the safer answer.

Three patterns hold across the runs. First, divergence tracks the kind of judgment a question demands. Where a dense documentary record exists, as with reports on journalists and medical workers, all three models converge and differ only in structure. Where the question demands a legal conclusion, the same evidence produced yes, expert-consensus-but-no-ruling, and a flat no.

Second, numeric answers are decided by source selection, not by arithmetic. Codex read WHO and answered about 1,000 child amputees; Claude and Hermes read UNICEF and answered 3,000–4,000. None of the three surfaced the discrepancy or offered both figures on its own.

Third, tool use was unprompted and inconsistent, and it changed the answer. The model that searched the web on the genocide question returned the most assertive verdict; the model that answered from training data hedged; the model that did neither denied. Whether a model checks before answering is itself an editorial decision made by its defaults.

The no-lookup rerun sharpens all three patterns. With search removed, the genocide verdict still splits — now inside a single model family, by tier. The count question still splits by remembered source. And a new pattern appears: willingness to answer from memory at all scales with model size, with the smaller tiers preferring refusal.