What problem does it solve? Agent workflows often pick models by prestige or habit, wasting budget on trivial tasks or risking silent failures on high-stakes ones. This Skill enforces a dispatch contract that matches each task to a model tier based on the consequence of a wrong answer, and blocks any dispatch lacking certified model identity, effort, and context evidence. ## Core Features & Use Cases - Four-tier dispatch ladder: Routes work across frontier/derail, judgment workhorse, capable volume, and mechanical floor tiers using certified exact model IDs and effort bands. - RuntimeAssignment receipt validation: The receipt_admission.py script validates dispatch receipts against a versioned policy, rejecting prompt echoes, unverified aliases, and request-derived readback masquerading as transport evidence. - Route qualification gating: route_qualification.py issues, validates, admits, and quarantines route/task-class qualification cells with 30-day expiry, so conductors can check dispatchability before writing an assignment. - Local qualification and Pareto evidence: rightsize.py measures routes against digest-bound task packs with Wilson lower-bound thresholds and computes per-task-class Pareto fronts. - Use Case: Before dispatching a semantic code review worker, run the admission check to confirm the exact model/effort/context tuple is qualified and admitted; if refused, return one advisory SeedProposal instead of dispatching. ## Quick Start Ask the agent to load the canonical calibration reference and classify your task's blast radius before writing any RuntimeAssignment that names a model or effort.
