Public methodology

Evidence policy

Every physiological claim in this publication carries a grade and an identifier — a PMID or a DOI. This page defines what the grades mean, which sources qualify, what happens when sources disagree, and what happens when something turns out to be wrong. It is outside the paywall on purpose: a grading system that cannot be audited is decoration.

The number that sets the standard

In the 2026 umbrella review of dietary patterns and inflammatory markers, the 60 included meta-analyses were assessed for certainty under GRADE. High certainty: 2%. Moderate: 52%. Low: 30%. Very low: 15%. Across its 61 narrative syntheses, high certainty was zero (PMID 40657695, p. 1180). Methodological quality of the 30 reviews, under AMSTAR-2: high 30%, moderate 33%, low 27%, critically low 10%.

A grading scale in a field with that distribution is not a badge. It is the only way to tell which sentences are load-bearing.

The scale

Grades describe the strength of the evidence behind a claim, not the size of the effect and not how widely the claim is repeated. A claim can be popular, mechanistically plausible, and graded D.

Where a source reports GRADE certainty directly, that assessment is used rather than a re-derived one, and the source is named. Where a claim sits between two grades, it is recorded at the lower one. Where a grade cannot be justified at all, none is assigned and the gap is listed — several such gaps are named in the bibliography chapter.

Absence of evidence and evidence of absence

These are kept visibly apart everywhere in the text, because merging them is the single most common failure in this category:

• Tested, no effect. Quality trials looked and found nothing. This is a positive finding, it is the most valuable class of statement in the publication, and it carries the mark shown in the legend above.
• Untested. No trials exist in either direction. Nothing is known.
• Overstated. Something real is present, much smaller or much narrower than advertised.
• Category error. The claim measures something other than what it says — a derived ratio treated as an input, a processing classification treated as a nutritional property, a marker treated as an outcome.

Which sources qualify

• Umbrella reviews and systematic reviews with meta-analysis are preferred over individual trials, and trials over observational studies.
• Mechanistic and animal work can explain a finding but cannot by itself raise a claim above grade C.
• Review quality is assessed with AMSTAR-2 where the source reports it; a rating of low or critically low is stated alongside the claim rather than quietly dropped.
• Authors’ own caveats travel with their numbers. Where a meta-analysis says its estimates must be read with extreme caution, that sentence appears next to the estimate, not in a footnote.
• Every claim records the date it was checked. Nutrition evidence moves, and an uncited assertion is not carried forward on reputation.

Every figure is read in the primary source

Not in an abstract, not in a summary, and not in another publication’s retelling. This rule exists because it has already caught two failures during assembly, and both are documented in the text rather than quietly fixed:

• An internal retelling of the umbrella review recorded four dietary patterns at the wrong certainty level. One of them had been tested and found to have no effect at moderate certainty, and was written down as “insufficient data” — a claim about the world converted into a claim about our ignorance.
• An automated page summariser produced a plausible, well-formatted certainty table containing ratings that do not appear anywhere in the paper. Full-text extraction showed the real wording. Automated summaries are not accepted as an evidence source here.

Contradictions are published, not resolved by preference

Where two credible syntheses disagree, both appear with their designs and their sample sizes, and the disagreement is described. Four live contradictions are carried in the text at present, including one where a homogeneous randomised null sits against a heterogeneous randomised positive for the same dietary pattern. Picking the convenient one and citing it alone is how this field produces its confident sentences.

Scope of application

The strongest current evidence for dietary effects on inflammatory markers was measured in adults who already have at least one chronic condition. Extrapolating a result to a healthy population is treated as a separate, ungraded assertion. And the precise statement is that such effects have not been tested in that population, which is not the same as having been disproved.

Language

This publication never states that a food, pattern or supplement “reduces inflammation”. The EU list of permitted health claims, Reg. (EU) 432/2012, contains the word “inflammation” zero times — checked literally against the full official text. There is no authorised way to make that claim about a food in the EU, and it would be imprecise in any case, since trials measure named markers that move independently of one another.

So the sentence pattern throughout is descriptive: what was measured, in whom, how much it moved, and how certain the synthesis was. “Anti-inflammatory” appears only as the name of somebody else’s claim.

What is deliberately absent

• No claims that food treats, prevents or reduces the risk of disease (EC Regulation 1924/2006).
• No before-and-after imagery, weight outcomes, or time-bound result guarantees.
• No elimination protocol. No systematic review in this field has tested one, and the documented psychological risk sits on the other side of that absence.
• No symptom questionnaire and no health question at checkout. Nothing in this product collects special-category health data, and the tracking artefacts stay on the reader’s own machine.
• No supplement affiliate links anywhere in the graded text. A commercial interest in a product changes how its evidence gets written about, and no disclosure line fixes that.
• No named practitioner persona. This publication has no clinical author, and it does not invent one.

Conflicts of interest, including other people’s

The publication earns money in exactly one way: readers pay once for access. There is no sponsorship, no affiliate revenue, and no advertising. If that ever changes, it will be disclosed on this page before it appears anywhere else, with the date it changed.

Cited authors’ interests are disclosed too. Where a meta-analysis that supports a graded claim has a co-author affiliated with an industry that benefits from it, that is stated beside the claim. It does not invalidate the result; it is a fact a reader is entitled to weigh.

Corrections

Errors are corrected in place and logged. A correction entry records what the text said, what it says now, what prompted the change, and the date. Corrections that change a grade or a safety statement are emailed to everyone who has access.

The log is public at /errata, together with the reporting address. It is empty until the first correction, and it says so rather than pretending otherwise.

Open gaps

The bibliography chapter ends with a list of what could not be verified, including two safety questions that are still open and are named rather than filled in with plausible text. That list is part of the method. A publication in this field with no unverified items has not checked carefully enough to find any.

Translations

Machine translation is used for a first pass. No language is published until a native speaker has reviewed, at minimum, the evidence grades, the contraindications, and the legal notices. A mistranslated grade is a safety problem, not a copy problem.

Use of AI

Drafting and literature triage use language models. Every citation is verified against the source record before publication — see the primary-source rule above, which exists precisely because a model fabricated a certainty table during assembly. No synthetic voice or likeness is presented as a real person.

See the grading applied: six of the twenty-six claims, free