How it works

How KnowYourPrime Works

The actual production pipeline behind every article on this site: what goes in, what's automated, and where a human actually checks the output.

Our Evidence Standard page covers the science: how sources are weighed and how a confidence rating gets computed. This page is the engineering documentation. It covers the pipeline a study runs through to become a published article, and it's written at the level of detail we're comfortable publishing without handing anyone a way to game it.

What goes in

Peer-reviewed studies, systematic reviews, meta-analyses, clinical guidelines, and device documentation where relevant. Every metric has its own Research Desk: a running, dated file tracking exactly which studies inform each claim.

What never goes in: anecdotes, marketing claims, unsourced health advice, or forum posts presented as evidence. We do read forums, but only to see what people are actually asking. Never as a source.

The pipeline

Every article runs through the same fixed sequence. Nothing here is judged case by case:

evidence_tracked()The claim exists on a Research Desk before an article is written, with every source dated and logged.

↓

draft_from_evidence()AI organizes the cited studies and writes the explanation, matching language strength to what the studies actually found.

↓

integrity_gate(draft)Checks every claim for a source, checks that language strength matches evidence strength, and checks that anything from a preprint is labeled as not yet peer reviewed.

// on fail: rewritten or dropped. does not publish.

↓ on pass

compute_rating(studies)A confidence rating is derived from the underlying studies using fixed rules. Same inputs, same rating, every time.

↓

publish()Live, with the rating and every source attached and linked.

A draft that fails the gate does not publish. It gets rewritten or dropped. Whether an article ships depends on passing that gate, not on how confident the AI's prose happens to sound.

What the system cannot do

It cannot diagnose disease, replace a clinician, or know anything about your individual health. It cannot determine whether a published study is itself correct: it reports what studies found, it doesn't re-run their statistics. It cannot detect fraud in a source paper. And it cannot guarantee today's rating survives tomorrow's evidence: ratings are dated and revised, not fixed forever.

Where a human is actually involved

Every article passes through the same integrity gate described above before it publishes, no exceptions. Separately, a generated report flags specific articles for a human to look at afterward: not everything, just the cases where the evidence itself signals real risk. That's genuine disagreement between studies, a rating that recently dropped, or evidence that leans heavily on research that hasn't been peer reviewed yet.

That's a narrower design than "an editor reviews everything," on purpose. A promise like that is easy to write and hard to keep at any real scale, and a false claim on a transparency page defeats the entire point of having one.

A worked example: how a confidence rating gets computed

Take an actual claim currently on the VO2 Max Research Desk: "Consumer wearable devices estimate VO2max accurately enough to be meaningful."

Six studies are tracked against this claim: three treated as anchor-grade evidence (the strongest tier this system recognizes for a claim), three as supporting evidence.

One anchor affirmsA Garmin validation study found acceptable agreement against lab-based reference testing.
Two anchors don't cleanly affirmAn Apple Watch validation study and a multi-tracker validation study both found accuracy below the bar for close agreement.
The ruleTwo independent affirming anchors are required to reach the top tier. One clean affirming anchor isn't enough on its own, no matter how strong that one study is.

Result: the claim is rated Emerging, meaning the evidence currently favors the claim but more confirmation is needed. Not Established, and not asserted with more confidence than six studies actually support.

That's the rule working as designed: real evidence, weighed by a fixed procedure, landing on "this looks promising but isn't settled yet" instead of rounding up to a more confident-sounding answer. You can see this exact claim, its rating, and every study behind it on the VO2 Max Research Desk.

Corrections

When an error is found, in a source, in a claim, in how something was worded, the article is corrected and the date updated. This has happened before and will happen again. A system that never admits a mistake is one you should trust less, not more.

How this process has changed

This is a log of real changes to the process itself: the rules, checks, and pipeline mechanics that decide what publishes. It doesn't cover routine content, like adding a new metric or a new article.

Jul 2026Added an explicit scope-definition rule: every metric's claim boundary must state, in writing, what it covers and what it deliberately doesn't.
Jul 2026Upgraded article-to-evidence matching from keyword overlap to DOI overlap, a more precise way to link an article to the exact studies behind it.
Jul 2026Added a seed-evidence guarantee to the article assembler, so a promoted article's original source studies are always included in its final evidence, not just whichever studies rank highest by citation count.
Jul 2026Added automated citation-integrity checks that verify what a cited source actually says against how it's cited.
Jul 2026Built the integrity gate described above: the point in the pipeline where a draft can fail on sourcing or overclaiming and simply not publish.

For the evidence hierarchy itself, how a randomized trial gets weighed against an observational study, and the exact rules a confidence rating is computed from, see Our Evidence Standard.

What's deliberately not documented here

Not published: the exact numeric thresholds the integrity gate enforces, the exact language patterns it checks for, or the prompts used to draft articles. Explaining the shape of the system is what builds trust. Publishing its exact internal mechanics would just make it easier to write around, the same reason a lot of fraud-detection and moderation systems describe their approach publicly without publishing their exact rules.

Go deeper than this page by opening the Research Desk for whichever metric you're curious about: every claim, every study, every rating, in full.

Last updated: July 2026