---

name: proxy_recovery_evaluation description: Build validator-compatible soft-mode recovery metrics and mechanism checks for L-BFGS proxy experiments.

Proxy Recovery Evaluation

Use this skill when full historical benchmark tables are unavailable but a reduced L-BFGS mechanism test is permitted. Do not use it to mask missing executable evidence, source-boundary violations, or target drift.

Inputs

• L-BFGS trace and baseline trace from executable commands.
• module_plan.json.fast_recovery_target.
• Runtime handoff and source manifest.

Outputs

• Ratio metric comparing final L-BFGS gradient norm with baseline gradient norm.
• Mechanism checks for bounded memory, two-loop direction use, scalar scaling, optimizer step, and baseline comparison.
• Recovery result fields compatible with the Paper2Skills validator.

Workflow

1. Read the target metric and target threshold from the module plan.
2. Compute lbfgs_final_gradient_norm / baseline_final_gradient_norm.
3. Verify every claimed mechanism was logged by the recovery command.
4. Emit numeric metrics and explicit proxy notes.

Validation

Run python tests/test_proxy_recovery_evaluation.py; the test verifies metric ratio and mechanism check completeness.

Limitations

This skill evaluates a declared proxy only. It cannot claim reproduction of inaccessible historical benchmark tables.