ok so llama 3.3 70b q4 generates valid json schemas but hallucinates enum values past 12k context

tested llama 3.3 70b q4_k_m yesterday on openapi schema generation task. past 12k context the model generates perfect json schema structure - all required fields present, correct $ref pointers, valid type definitions - but the enum values are completely hallucinated. example: we pass in context that says status can be ["pending", "approved", "rejected"] and the model generates schema with enum ["pending", "active", "completed"]. schema validates fine but the enum values don't match what was in context. seems like long context degrades enum recall specifically while keeping schema structure intact