D253 - The model was told the wrong libraries, and earned names from ones it was never shown
decided · 2026-08-25 · found by checking where the earned names actually came from
The prompt names the libraries the identifiers belong to, and the
live run supplied four hardcoded strings: libkernel,
libSceNpManager, libSceAudioOut,
libSceVideoOut.
Every name the model has ever earned came from a library that is not in that list:
| earned | library | shown? |
|---|---|---|
sceAgcDriverRegisterResource |
libSceAgcDriver |
no |
sceNpAuthPollAsync |
libSceNpAuth |
no - NpManager is a different library |
sceAudio3dObjectReserve |
libSceAudio3d |
no - AudioOut is a different library |
It was succeeding despite its context. A model told the wrong domain is being pointed away from the answer, and the one thing this model is measurably good at - vendor domain nouns - is exactly the thing that clusters by subsystem.
Derived instead from the round's own examples. A vendor name is
sce + a module word + the rest, and the module words are a
list the grammar already carries, so the segmentation is read rather
than guessed. 80% of the example set resolves, across 47 distinct
libraries, and the three names above all resolve correctly.
The examples already rotate per round, so the libraries rotate with them: each round now names the subsystems whose examples it is looking at, and the question narrows on its own without a separate per-library mechanism.