Operations JSON workflow repair
Turn repetitive operations tasks into reliable small-model routes with strict output control, prompt scaffolds, and sparse repair data.
Local optimization required no training or owned GPUs; the open route matched Sonnet performance at 5.2x lower latency and 6.0x lower cost.
The benchmark separates output-control gains from training gains, so the case study can show what changed before the final served route.
Workflow
The task is routine operations transformation: produce a bounded JSON output, repair malformed calls, fill structured arguments, or normalize workflow state.
Quality contract
The parser, schema, retry policy, and held-out examples matter as much as the model. A smaller route is useful only when it returns the contract reliably.
Optimization path
Understudy first tightens the harness: prompt, schema, parser, token cap, and retry behavior. Training comes later when the eval shows interface repair is not enough.
Buyer fit
This is a fit for back-office or product operations workflows where bad structure breaks downstream systems and frontier reasoning is often more capability than the job needs.
The detailed evidence lives on the operations benchmark.