[mshafir]

Five faults reported from one playthrough of a generated world, and the thread running through all of them is that nothing said anything was wrong.

The opening card never appeared. It waited on the director reporting the world's lore, and a scenario arrives with its lore already written, so the director never asks and never reports — meaning every world generated with improvisation turned on started with no premise, no protagonist and no idea which way the first beat lay. It now goes up as soon as the lore is known.

The people the story hangs on were improvising, so a beat opened, an errand landed in the journal, and the model wrote a pleasant line about the weather instead of the scene the beat existed for. They now answer from an authored tree or from the deterministic menu, never from a model, and they are drawn in their own colour so a bearing on the map points at somebody findable. The reverse fault sat beside it: one authored tree anywhere in a scenario silenced improvisation for the whole cast, because the scripted path answered for people it had nothing written for.

Model calls are priced off the catalogue rather than two hard-coded rows, so a world written on any of the other twelve models is no longer costed at exactly $0.00. The spend shows while the world is being written and again when it is done. Asking to keep the working records every prompt and answer and makes them readable in the app — on the writing screen, and on a page in the menu while playing — rather than only in a file nobody opens.

A tiny size, for trying the pipeline without paying for a world worth keeping: two beats, a handful of places, and conversations written only for the people the story turns on. The boundary grows when a seed offers nowhere for a story to happen, and never past the size it was chosen to be cheaper than.

Underneath all of it, two measured causes. Offline authoring ran on the live director's twenty-second timeout, and 39 of 40 calls in a real run took longer than that — which is why the reported world had fallback lore, procedural names and no conversations at all. And a schema mismatch was treated as permanent on the grounds that it would repeat; the same run had one model answer the dialogue schema 8 times in 26, so it is retried now, with the last attempt spent on a dearer model where the catalogue names one.