Adversarial Evidence Review Methodology

Purpose and pilot status

This Adversarial Evidence Review Pilot is designed to make benefit evaluations inspectable rather than promotional. It documents what is claimed, why the claim appears plausible, what the strongest evidence supports, what it does not establish, and how the working conclusion was reached.

The pilot is not independent certification, medical certification, expert approval, or a substitute for clinical guidance. A review remains in progress until its research brief is complete and human publication approval is recorded.

The claim-first process

A review begins by inventorying the meaningful versions of a claim and defining each one precisely. Broad wellness language, biomarker effects, disease prevention, and treatment claims are not treated as equivalents.

• Inventory the claim and where it appears
• Define a testable outcome and population
• Classify and extract the evidence
• Build the strongest reasonable supporting case
• Cross-examine methods, statistics, causality, safety, and independence
• Allow the supporting side to rebut material challenges
• Adjudicate each disputed question
• Set supported, qualified, and unsupported language
• Publish only after the completion checklist and human approval

Evidence hierarchy

Evidence is classified rather than counted. Replicated, well-conducted clinical evidence carries more weight for human outcomes than observational, animal, laboratory, expert, or marketing material.

Mechanistic and animal research can establish plausibility. Observational research can identify associations. Neither independently establishes that supplementation produces a human benefit. A mushroom, extract, or multi-ingredient result is not attributed to isolated L-ergothioneine unless the study design permits it.

• Replicated systematic reviews and high-quality clinical trials
• Individual randomized controlled trials
• Other controlled human intervention studies
• Prospective observational studies
• Cross-sectional or retrospective human studies
• Case reports
• Animal studies
• Laboratory and mechanistic studies
• Expert opinion
• Marketing statements and unsourced assertions

Supporting review and cross-examination

The supporting advocate must present the strongest fair, evidence-based case rather than a weak argument designed to fail. The examiner then tests study design, power, prespecified outcomes, direct between-group comparisons, multiple comparisons, effect sizes, confidence intervals, attrition, clinical significance, causation, generalizability, replication, funding, formulation, dose, duration, registration, and selective reporting.

Rebuttal and adjudication

The supporting position receives an opportunity to answer material criticisms—for example, ceiling effects, population selection, trial duration, or measurement sensitivity. The final adjudicator addresses each disputed question separately instead of hiding disagreement inside one overall label.

An absence of evidence is reported as uncertainty. It is not proof that an effect is impossible.

Publication stages

A Preliminary Evidence Profile maps claim variants, evidence categories, verified records currently in the database, and unresolved questions. It contains no evidence rating or efficacy verdict and remains excluded from search-engine indexing while the research is incomplete.

A Draft Evidence Review follows only after primary-source extraction and the supporting case, cross-examination, rebuttal, and working adjudication are substantive. Human Review in Progress identifies the separate approval stage. Review Complete is used only after human approval is recorded.

• Preliminary Evidence Profile
• Draft Evidence Review
• Human Review in Progress
• Review Complete

Three separate rating dimensions

Completed reviews report direction, evidence stage, and certainty separately. The existence of a study does not determine confidence in the claim.

• Direction: supportive, mixed, neutral, or contradictory
• Evidence stage: mechanistic, preclinical, observational human, preliminary clinical, or replicated clinical
• Certainty: high, moderate, low, or very low

AI role and human oversight

Codex may assist with evidence discovery, structured extraction, comparison, adversarial questioning, and drafting. AI output is not a source and cannot grant publication approval. This is currently a single-system pilot; no multi-model independence or AI consensus is claimed.

Each review records the system used, date, role, supplied sources, material disagreements when present, and human changes or approval. The current human-review label is “Human review in progress — publication approval not yet granted.”

Funding, conflicts, and independence

Commercial and secondary sources may document that a claim is being made, but they are not primary proof of efficacy. Study funding, author employment, supplied materials, branded formulations, and relevant conflicts are recorded where available.

Industry funding does not automatically invalidate a result. It increases the importance of registration, transparent analysis, complete reporting, and independent replication. Website ownership and commercial relationships are disclosed separately.

Corrections, versioning, and update triggers

Each completed review displays its publication date, last evidence search, evidence cutoff, substantive update date, review version, methodology version, AI disclosure, human-approval status, and revision history.

A review may be reopened after a new trial, failed replication, material correction, changed registration record, newly identified conflict, regulatory action, or well-supported reader challenge. Changes that alter meaning are dated and explained.

Publication safeguards

Internal research briefs are excluded from the production build, navigation, sitemap, and search index. An unfinished domain may have a linked Preliminary Evidence Profile only when it provides a useful claim map, evidence-category boundaries, verified records or an explicit record gap, unresolved questions, and clear language limits. These pages remain noindex and receive no rating.

Structured data never represents a page as medically reviewed, fact-checked, certified, or independently validated unless those statements are true and documented.