When your compute budget is a rack in a ministry basement rather than an elastic cloud, model selection stops being a matter of taste. We learned to compose small specialists — and then discovered the pattern wins even where compute is unlimited.
Decomposition beats scale
A contract-review task is not one problem; it is extraction, classification, comparison and drafting. Each step has a model size that saturates its difficulty. Route each step to the smallest model that saturates it and the pipeline is faster, cheaper and easier to debug than one large model prompted to do everything.
Orchestration is where quality lives
The gains come from the seams: validation between steps, retries with sharpened instructions when confidence drops, escalation to a larger model only when a specialist flags uncertainty. Quality is a property of the pipeline, not of any single model inside it.
Measure work, not benchmarks
Benchmarks reward models; operations reward systems. We track cost per completed task, revision rate, and time to acceptance — numbers a client can feel. On those numbers, a well-orchestrated set of small specialists is the strongest configuration we ship, online or off.