Optional bounded decision routing

Use deterministic routing for clear rules or a small known task. A decision model is useful when repeated bounded classifications have enough volume and measured benefit. It proposes one of the eligible routes; the coordinator still owns interpretation, authorization, verification and next actions. No model or score can approve installs, data disclosure, permission changes, release readiness or an external write.

Read decision-model contract before selecting a provider. Research actual Jev/Laya publisher docs, exact versions/model artifacts, licenses and compatibility through an available capability researcher. Resolve local/cloud data and budget choices before adoption. This skill does not bundle weights, credentials or a running decision service. Do not simulate a successful API call from vendor examples.

Derive eligible routes from the current task and actual host capabilities. Evaluate typed outputs on held-out project examples, including ambiguous/unanswerable cases, large option sets, changed tools and adversarial content. Scores are model estimates; calibrate and choose escalation thresholds from measured error consequences.

The optional Rust orchestrator route-proposal CLI checks a caller-supplied allowlist, finite scores and explicitly supplied minimum score/margin. It recommends one declared route or escalation and never executes tools. Follow its decision-model contract; a recommendation verifies neither capability availability nor consent. All candidates supplied by the provider must be included. If scores are incomplete, badly calibrated or stale, escalate and use ordinary planning/verified rules.