ADR-002 — Model orchestration over a single frontier model
Context
The best model for one step is rarely the best model for every step. Extraction, classification, routing and validation are cheap, well-bounded problems. Requirement analysis and gap detection are not. A single frontier model applied everywhere is expensive, slow, and creates dependence on one provider's pricing, availability and policy decisions.
Decision
Route each task to the model that fits it. Smaller specialised models handle extraction, classification, routing and validation. Frontier models are reserved for genuinely difficult reasoning. Model selection is an architectural concern encoded in the system, not a default inherited from whichever provider was chosen first.
Consequences
- Positive: lower cost and latency for routine steps; provider independence; the architecture survives model generations because models are components, not foundations.
- Negative: routing logic must be designed, tested and maintained, and evaluation must cover every model in the chain rather than one.