Two kinds of blank, kept apart. A stealth model hides its maker: nobody says whose it is, and measurement is the only way to find out. An unlabelled model has a maker on the tin but no tokenizer family on its catalogue label, and measurement fills that blank in.
The fingerprint is shared only inside the maker's own model line: an in-house vocabulary rather than a borrowed one.
A clean fingerprint that no other measured model shares so far. Where something comes close, the nearest model is shown.
Their providers report token counts too disturbed (by caching or shifting overheads) to trust. Publishing a corrupted fingerprint would be worse than none.
Queued for the next measurement run; results land here as each one completes.
Because the known models are the reference library. A stealth model can only be placed by comparing its fingerprint against models whose families are established, so every known model measured makes the next unknown easier to identify. That is exactly how Ox Alpha was placed within minutes of measurement. Method and limits: the methodology page.
