Tool Shadow

Tool Shadow

Use this skill when an agent needs repeated access to a property that is hidden inside an expensive, slow, licensed, remote, or physical tool.

A Tool Shadow is a fast, explicitly lossy approximation of an authoritative tool. It may use deterministic heuristics, a reduced solver, cached results, a learned surrogate, or a combination of these. It exists to improve perception, ranking, and design-space exploration. It does not inherit the authority of the tool it approximates.

The governing rule is:

Use the shadow to explore. Use the authoritative tool to verify.

Required workflow

1. Name the hidden property the agent needs to perceive.
2. Identify the authoritative tool or physical procedure that establishes the best available answer.
3. Decide whether an exact cheap rule already solves the problem. Do not train a model where deterministic geometry, arithmetic, lookup tables, or caching are sufficient.
4. Define the Tool Shadow profile using https://github.com/closedloop-technologies/climbhill/blob/main/resources/strategies/tool-shadows/CONTRACT.md.
5. Declare the validity envelope, expected cost and latency, uncertainty method, known failure modes, and authority level.
6. Return a versioned Functional Observation Packet for every invocation.
7. Keep predicted, verified, and observed evidence separate.
8. Compose this skill with progressive-fidelity-tool-routing before an agent spends an expensive tool budget or acts on a shadow prediction.
9. Compose it with tool-shadow-calibration before promotion to wider use.

ClimbHill mapping

Treat each candidate asset or action as an Attempt. Every shadow, reduced solver, authoritative tool, or physical result is a distinct Evaluation. A continuation analysis may recommend another candidate or a higher-fidelity evaluation, but deterministic policy or a human authority decides whether the tool may run or an attempt may be promoted.

Do not mutate an existing Evaluation when a better result arrives. Append the new result, link it to the same candidate and conditions snapshot, and record which prior evidence it supports, contradicts, or supersedes for the current decision.

Minimum outputs

• Tool Shadow profile
• immutable input and conditions hashes
• prediction and uncertainty
• validity-envelope result
• authority label
• escalation recommendation
• links to authoritative or physical evidence when available
• cost, latency, and cache metadata

Never claim

• that a shadow prediction is certification
• that average model accuracy proves safety near a decision boundary
• that the shadow can authorize its own use or promotion
• that a stale prediction still applies after the asset, conditions, tool, or model version changes
• that a model is reliable outside its declared validity envelope

Strategy resources

• Tool Shadow overview
• Contract and observation packet
• Progressive-fidelity routing
• Calibration and active learning
• GeoField bracket case study