Pointing the model at the right question
Three changes, each from a measurement rather than a hunch.
It was being told the wrong libraries. The prompt
names the libraries an identifier set belongs to, and the live run
hardcoded four. Every name the model has ever earned came from a library
not among them - libSceAgcDriver,
libSceNpAuth, libSceAudio3d. Now derived from
the round's own examples by reading the module word out of each name,
which the grammar already lists: 80% of examples resolve across 47
libraries, and all three earned names resolve correctly. Because the
examples rotate per round, the libraries rotate with them, so each round
asks about the subsystems it is actually looking at (D253).
It had no memory of its own failures. A round grows
a local clone of the grammar, sweeps, and discards it; only successes
reach the bank. So a word that failed was proposed, accepted and swept
again every round - Group twelve times, at thirty-five
million candidates each. tried_before now holds every word
swept this run, deliberately separate from the bank, which stays
evidence-only (D254).
A third of what it proposes is already in the shipped
standard list. Nine of twenty-six distinct words appear inside
a standard.txt name, including three of the five it has
ever banked. Those arrive free once that list is decomposed into parts,
so the prompt now says the standard vocabulary is covered and asks for
the platform's own domain nouns - which is the two-thirds no standard
name contains (D255).
And two things deliberately not built. Mutation and combination both looked like obvious wins and are both jobs for a loop, not a model: a loop enumerates every variant of five hundred words in milliseconds, and combination is what the grammar patterns already do.