Learned control for multi-model inference.
A Mistral 14B LoRA adapter trained to choose one, two, or three model calls under explicit quality, cost, latency, and verification constraints.
What changed after training
The frozen test split contains 1,200 cases from eight task families absent from training and validation. Deterministic Loom is the operational baseline.
adapter + guard90.8%
deterministic Loom+11.5 pp
95% CI9.4 to 13.7
under-escalation0.0%
caught by guard13 / 13
The raw base checkpoint is a capacity control, not a comparison with an instruction-tuned assistant. Full lane results, raw predictions, and scoring receipts are linked below.
Inspect constrained routing
The scorer enumerates every valid writer, reviewer, and finalizer assignment, rejects over-budget plans, and ranks the remaining candidates using the disclosed utility function.
Reason codes
Machine-readable plan
| Policy | Writer | Reviewer | Finalizer | Credits | Latency | Quality | Pass rate | Utility |
|---|
Weights, data, raw predictions, receipts, checksums, and limitations are published together.
The study measures routing-plan generation. It does not establish universal model rankings or final code correctness.