Route unresolved work to the capability that can actually resolve it.
A bigger model does not fix a missing document. Route application work across governed queries, calculations, rules, specialized capabilities, models, providers, deployments, and human judgment.
Escalate the uncertainty that is actually blocking the work
Missing evidence
Retrieve or assemble required governed information before asking a model to infer from absence.
Deterministic conflict
Verify, reconcile, or query authoritative state when facts disagree.
Model uncertainty
Escalate to a specialized or stronger eligible model only when evidence is sufficient but interpretation remains unresolved.
Policy uncertainty
Evaluate policy rather than asking another model whether the operation is allowed.
Authority gap
Route to an eligible reviewer when understanding exists but authority does not.
Effect ambiguity
Reconcile downstream state instead of using model escalation as recovery.
Treat confidence as evidence, not authority
Classifier probability, model score, sample agreement, verifier result, calibrated probability, and Context Sufficiency do not mean the same thing.
Allow the application to abstain
If no eligible capability can satisfy the contract, the application can collect evidence, wait, route to human judgment, open an exception, or stop without a consequential effect.
Establish which routes may participate before choosing one
The first routing question is not which available model scores highest. It is which routes are permitted for this Execution Instance.
Use exact capabilities when they can satisfy the task
Exact query
Retrieve an authoritative current value.
Calculation
Compute a known result or feature.
Rule
Apply explicit business logic.
State machine
Resolve a defined transition.
A successful invocation is not automatically a sufficient business result
A Capability Sufficiency Gate can decide whether the current result satisfies the application contract or needs a specific next action.
Make the routing policy inspectable
Expose hard constraints, routing objectives, sufficiency conditions, fallback, escalation, abstention, and human escalation in the Execution Graph.
Record why the route was selected
Preserve candidate capabilities, eligible routes, rejected routes and reason codes, selected deployment, evaluation profile, cost, latency, validation, and stop reason.
Map how one AI workflow should route unresolved work.
Use an architecture review to define eligibility, sufficiency, deterministic-first paths, escalation budgets, fallback, abstention, and route evidence.