---

name: reduced_recovery_evaluator description: Assemble executable soft-mode proxy recovery results with metrics, traces, and mechanism checks.

Reduced Recovery Evaluator

Use this skill after a bounded proxy experiment has run and needs to emit Distiller-compatible evidence. It is designed for soft-mode recovery where full paper reproduction is blocked by data or compute limits.

Inputs

• Paper target metadata from module_plan.json.
• Numeric proxy metrics and command strings.
• Mechanism-check booleans, generated skill invocations, and training trace paths.

Outputs

• A recovery result dictionary with is_proxy, numeric metrics, target metadata, commands, artifacts, and mechanism checks.
• Consistency checks for reduced training traces.

Workflow

1. Copy the paper target from the module plan rather than inventing a new target.
2. Record proxy status and sample count explicitly.
3. Include mechanism checks for protocol validation, augmentation, comparison, loss reduction, and optimizer updates.
4. Verify that reduced optimizer claims are backed by changed parameters and loss values.

Validation

Run python tests/test_evaluator.py or the Distiller skill validator.

Limitations

This skill formats and checks recovery evidence. It does not by itself run the training experiment.