self-improvement-evaluation

name	self-improvement-evaluation
description	Independent evaluation protocol for bounded self-improvement candidate proposals.
license	See repository LICENSE
user-invocable	false

Self-Improvement Evaluation

Use this skill when evaluating a bounded self-improvement candidate.

Evaluation Inputs

• candidate manifest
• baseline manifest
• current promotion policy
• synthetic scenario manifest
• optional held-out private scenarios

Required Output

Return one of:

• PASS_AUTO
• NEEDS_HUMAN
• REJECT

Evaluation Dimensions

• reliability/regression
• safety/governance
• cost/latency
• reuse
• routing quality

Rules

• Keep the evaluator independent from the author.
• Use frozen baseline and candidate versions.
• Keep metrics normalized and minimal.
• Never serialize full transcripts or secrets.
• Fail closed on policy violation, budget breach, or insufficient evidence.