Framework Injection (FI) was proposed as a paradigm for transferring complete domain-specific reasoning methods from human expert to AI agent, superseding both prompt engineering and context engineering. But does one FI mode fit all models? This paper reports the most significant empirical finding of the Artisanal Intelligence Program: framework mode must adapt to model capability. A controlled benchmark across three real-world cases (Ponzi fraud, MODY misdiagnosis, Brazil tax reform) and two model tiers (Haiku as smaller model, Sonnet as frontier) reveals a striking asymmetry: FI-scaffold improves smaller models by +9.3% over prompt engineering but degrades frontier models by -11.1%. Context engineering performs worse than prompt engineering in every scenario—a validated negative result consistent with Sweller's Cognitive Load Theory. We derive five empirical rules and propose a two-mode solution: scaffold mode (all five framework types, explicit, detailed) for smaller models, and metacognitive mode (evaluative 30%, ethical 30%, meta-cognitive 40%) for frontier models. We ground this asymmetry in Vygotsky's Zone of Proximal Development (scaffolding lifts), the ceiling effect (scaffolding constrains), and the compass metaphor (metacognition guides without constraining). The two-mode solution recovered +38% performance on the Brazilian tax case when switching from scaffold to metacognitive mode on the frontier model. We describe an adaptive generator that automatically selects framework mode based on assessed model capability, and articulate implications for the design of model-aware interaction systems.
Renato Aparecido Gomes (Thu,) studied this question.