D255 - Thirty-five per cent of what the model proposed was already in the shipped word list
decided · 2026-08-25 · measured, and it changes what to ask for
Of the twenty-six distinct words a model proposed and swept over
thirty-six rounds, nine appear inside a name in
standard.txt - including Group and
Node, the two it repeated most. Three of the five words it
has ever banked are in there: Unset in
unsetenv, Object in
kinfo_getvmobject, Resource in
setclassresources.
Those names are tried whole and never decomposed into parts, so the words are present in the tree and unavailable to the grammar - which a parallel line of work found independently from the other direction. Once that list is decomposed, a third of what this model produces arrives for free and asking for it again buys nothing.
What no standard name contains is the vendor's own vocabulary. The
same run produced Dma, Midi,
Bios, Endpoint and Bandwidth
unprompted, and that two-thirds is the part worth paying for. The prompt
now says so in one sentence.
Two things this rules out, and both were tempting.
Morphological mutation - Group to
Groups, Grouping - is a loop over words that
already exist, and a loop does it exhaustively in milliseconds; spending
a five-to-twenty second model call on it is D231's measurement again.
Combination is what the grammar patterns already are, so more
of it is a line of TOML rather than a question for a model. The model's
job is the words a loop cannot invent, and nothing else.