There is a page on this site that is nothing but formulas. Scoring, percentiles, confidence intervals, the gradient decomposition, the type assignment, the population aggregates and the privacy rules — written out, with the reasoning, in about nine sections.
Nobody has asked for it. I think it is one of the more important pages here.
Why
Because the alternative is asking to be trusted, and this project has no standing to ask.
There is no institution behind it, no ethics board, no peer review, no funding. What I can offer instead of a stamp is that every claim is checkable: the items are published, the scoring is published, the reference data is a committed file anybody can diff, and the arithmetic that turns one into the other is written out in full.
A reader who wants to know whether the 94th percentile on their result means anything can go and read exactly how it was computed, including the part where it was pulled toward the mean and why. That is a weaker guarantee than institutional review in some ways and a much stronger one in others.
The constants are imported, not typed
A small thing that matters more than it sounds. Every number on that page — the minimum cell size, the contrast weights, the engine version, the size of the reference sample — is read from the module that owns it and interpolated into the text.
A page that restates a threshold is wrong the first time the threshold moves, and it is wrong in the most damaging way available: it looks authoritative while contradicting the code. The same rule applies to the figures on the methods page, which are recomputed by a script rather than typed in.
Saying "these numbers are generated from the source" is a stronger claim than any amount of promising to keep them updated.
The section that took longest to write
The last one, which lists the five places the arithmetic is weakest.
A formula page that only lists formulas is an advertisement. The instrument depends on an inter-ring correlation nobody has measured; three statements per level is few; regression to the mean assumes the reference group is really yours; the noise probe assumes something about shape that might be false.
None of that is comfortable and all of it changes how much weight a reader should put on their own result. It is the part of the page I would most want to read on somebody else's site, which is the only test I have for whether it belongs on mine.