D258 - The vocabulary was never the gap; the shapes were

decided · 2026-08-25 · four unnamed imports, each one step out of reach

Every harvesting mechanism had been at zero marginal yield against the corpus for weeks, and the conclusion drawn from that was that the vocabulary was exhausted. It was the wrong conclusion, and four imports the conformance probe calls showed why.

posix_vocabulary() already existed. It already capitalised all 3,018 names harvested from FreeBSD source, and it was already injected into the grammar. The words were there the whole time. No pattern used them behind a module: prefix-posix spells sceUsleep, and the vendor writes sceKernelUsleep.

One new shape - ["prefix", "module", "posix", "tail"] - reaches sceKernelUsleep and sceKernelDlsym with nothing added to any word list. It costs 75 modules x 3,018 names against an existing space of 2.8 billion.

Two more gaps, found the same way:

• No shape allowed a second verb. sceKernelLoadStartModule loads and starts, and every name of that form was unreachable however complete the vocabulary was. prefix-module-verb-verb-object fixes it.
• posix_vocabulary joined underscore parts rather than emitting them. pmap_unset became PmapUnset, so Unset - the morpheme a vendor name actually reuses - was never offered to the generator. Emitting each part as well takes the list 2,923 to 3,741.

What this says about "exhausted"

A miss proves the name was not among those tried, and the tried set is the vocabulary crossed with the shapes. Reporting exhaustion on the strength of a full vocabulary measures one factor of a product and calls it the product.

The corpus sweep after these three changes named 38 imports it had never reached - none from module strings, none from published names tried whole, all from the generated grammar against material that has been in this repository the whole time.

And a hazard found on the way

names widens the grammar as a side effect of running, so pointing it at another project's binary imported 12,080 of that project's strings into the vocabulary with nothing said and nobody deciding. Reverted. The refusal filter also could not see the injected posix list - it read the file text - so it re-learned words the grammar already had, which is what the model in the other thread had been paying for by hand.