How MyAtomy reads evidence.

A claim is a sentence. A conclusion is what a body of evidence can carry. The gap between the two is where almost every health argument goes wrong.

So the first thing MyAtomy does is restate the claim precisely: who it applies to, what the exposure is, what outcome is being asserted, against what comparison, over what period. Most popular claims turn out to specify almost none of these, and that vagueness is itself a finding — a claim that cannot be pinned down cannot be tested.

The claim is then judged as written. If the honest version is weaker than the stated version, the report says so and describes the weaker version rather than quietly grading the claim on a curve.

A mechanism explains how something could work. An outcome is what actually happened to people. These are different kinds of knowledge and MyAtomy keeps them in separate sections of every report.

Mechanistic reasoning is seductive because it feels like understanding. A compound activates a pathway, the pathway is involved in a disease, therefore the compound treats the disease. Each step can be true while the conclusion is false: the dose may be unreachable, the pathway may be redundant, the body may compensate, the effect may be real but too small to notice.

A plausible mechanism raises the prior. It does not settle the question, and a report that only has mechanism to offer will say so.

Study designs differ in how well they isolate an effect from everything else going on. Randomisation is the cleanest tool available for that, which is why trials sit high on the ladder and observational work sits lower.

But the ladder describes what a design can support in principle, not whether a specific study did. It is a prior, not a verdict.

Within any tier, quality varies enormously. MyAtomy weighs sample size, whether the endpoint measured is the outcome anyone cares about or a marker standing in for it, how participants were selected, what was controlled for, whether the analysis looks pre-planned, how much was lost to follow-up, and whether the result has been reproduced.

A badly conducted meta-analysis pooling weak trials does not outrank one large, careful, pre-registered study. Pooling poor data produces a precise estimate of a biased number.

Effect size gets read separately from statistical significance. A result can be highly significant and practically meaningless — with a large enough sample, almost anything is significant. The question a report tries to answer is whether an effect of that size would matter to a person.

Any sufficiently popular claim has supportive studies. Finding them proves very little. What is informative is the shape of the whole literature: whether findings converge as studies get larger and better designed, or whether the effect shrinks toward nothing.

Null results and failed replications get their own section in every report, deliberately, because they are the part most likely to be omitted elsewhere. A claim with strong support and strong contradiction is genuinely unresolved, and MyAtomy reports it as mixed rather than picking the more satisfying side.

Evidence is always evidence about someone, at some dose, for some duration, measured somehow. A finding in trained young men does not automatically transfer to sedentary older women. A dose achievable by injection in a rodent may be unreachable by diet in a human. A four-week change in a blood marker is not a lifetime change in risk.

When a claim generalises past the population that was studied, the report names the gap rather than treating it as a detail.

There is a real difference between studied and found not to work, and not adequately studied. Both produce the answer no, and they mean opposite things about what future research might show.

MyAtomy reports insufficient evidence as its own verdict rather than rounding it to a confident answer. Every report also states what would change the verdict — the design, endpoint and rough scale of a result that would move it — because a position that nothing could revise is not a scientific one.

Confidence describes how much the retrieved literature can settle the question. It is driven by the quality and consistency of the evidence, how directly the studies address the claim as stated, and how much of the relevant literature appears to have been found.

It is not the size of the effect. A large effect shown once in a small unreplicated trial earns low confidence. A modest effect reproduced across good trials can earn high confidence.

The evidence strength score is a separate, deliberately coarse reading aid: a 0-100 summary of how much weight the human evidence carries. It is not a probability, not a p-value and not a clinical measure. The written report is the substance; the number is shorthand.