Progressive Fidelity Tool Routing

Progressive-Fidelity Tool Routing

Use this skill when an agent can answer a question at several fidelity levels and must decide which tool call is justified.

The router is a policy layer. It is separate from every evaluator and separate from the reasoning process that recommends a next action. A Tool Shadow must not decide whether its own prediction is sufficient.

Fidelity ladder

Use the smallest level that can answer the current decision safely:

• F0 exact-cheap: arithmetic, schemas, lookup tables, deterministic geometry, cache hits, static checks, or analytic formulas.
• F1 shadow: learned or heuristic approximation for perception, ranking, and broad exploration.
• F2 reduced: coarse simulation, reduced-order model, lower resolution, small sample, or bounded dry run.
• F3 authoritative: the real solver, compiler, database query, evaluator, external service, or other source of software truth.
• F4 physical: measurement, fit test, hardware run, user observation, or production outcome.

The level labels describe evidence authority, not model sophistication.

Routing procedure

1. Hash the candidate asset and all material conditions.
2. Check for a valid cache entry produced by the same evaluator version.
3. Apply exact cheap rules before any learned approximation.
4. Use a Tool Shadow to compare many candidates only inside its validity envelope.
5. Estimate uncertainty, distance to constraints, decision reversibility, downside of a false-safe result, and the expected information gained from a higher-fidelity call.
6. Escalate deterministically when any policy gate fires.
7. Spend the authoritative tool budget on finalists, boundary cases, uncertain cases, disagreements, and irreversible actions.
8. Append the resulting Evaluation and rerun continuation analysis from a quiescent evidence snapshot.

Mandatory escalation triggers

Run a higher-fidelity evaluator when:

• the candidate may be promoted, shipped, deployed, purchased, or fabricated
• the predicted result is near a hard constraint
• uncertainty or out-of-domain score exceeds policy limits
• the shadow and an exact rule disagree
• independent shadows disagree materially
• the input, conditions, or evaluator version changed
• the action is expensive, irreversible, safety-relevant, or externally visible
• the agent appears to be exploiting a recurrent shadow error
• policy explicitly requires human or authoritative review

Cost-aware behavior

Prefer batching, memoization, incremental evaluation, early rejection, and coarse-to-fine execution. Report expected and actual cost and latency for each call. Do not optimize only for fewer calls: optimize for decision quality per unit cost.

Use the canonical routing strategy in https://github.com/closedloop-technologies/climbhill/blob/main/resources/strategies/tool-shadows/ROUTING.md and the packet contract in https://github.com/closedloop-technologies/climbhill/blob/main/resources/strategies/tool-shadows/CONTRACT.md.