Where these numbers come from
Every hard number on this site carries a small mark next to it, like this oneMeasured. example only, no real figure attached. Hover it, or tap it on a touch device, and it names exactly where the figure came from. This page explains what each mark means.
countedCounted. this tag, as used throughout the site
Counted from the code.
A commit count from git log, a route count from the router modules, a model count from the schema. Anyone with access to the code could recount it and get the same number.
measuredMeasured. this tag, as used throughout the site
Measured by a script that's committed and can be rerun.
Dermaid's accuracy figures, the test-set audit, Kofi's retrieval accuracy. The script that produced the number exists and could be rerun.
ReportedReported. this tag, as used throughout the site
Reported by the client. I have not verified it myself.
Reported by a client or employer, not independently verified by me. Used sparingly, and marked as such rather than folded into a counted or measured figure it doesn't belong next to.
estimatedEstimated. this tag, as used throughout the site
My own estimate, not a measurement.
Used only where a real count or measurement doesn't exist, and never dressed up as more precise than it is.
Why this exists
Commit counts on this site were read from git, not remembered. Route and model counts were counted from the code, not estimated from memory. Where an old README claimed more than the code supported, the claim was corrected rather than carried forward. That’s a habit, not a feature, and it should be checkable rather than taken on faith, which is the whole reason for the marks.
The time a published number was wrong
Dermaid, a skin condition classifier, originally reported 95.2 percent accuracy. That figure came from a model-selection validation split, not a genuinely held-out test set. Checking the shipped test set found that 24 percent of it was perceptually identical to a training image. Re-evaluated on the deduplicated remainder, the honest figure is 92.7 percent against a 34.0 percent majority-class baseline. A second idea, that a difference in where the clear-skin images came from explained the model’s strongest result, was tested directly and did not hold up, and was retracted rather than published as a finding.
Owning that in public is worth more to a reader here than any single accuracy figure would be on its own.