quality-evaluate

Cross-Modelworkflow

Mission

Define metrics → measure → compare → report → act. Every evaluation produces a decision or action.

When to Use

Use when benchmarking AI system quality, measuring output consistency, running eval suites, comparing model versions, detecting quality regressions, grading outputs against rubrics, or generating evaluation reports.

Triggers: “benchmark this”, “run evals”, “measure quality”, “compare model outputs”, “quality gate”, “detect regression”, “grade these outputs”, “eval suite”

Skills Invoked

3-way majority vote:

• eval-output-grading (×3 vote: Haiku, GPT-5 mini, GPT-4.1)
• eval-variance — stability measurement
• eval-prompt-bench — prompt variant comparison
• eval-adversarial (Advanced) — injection/jailbreak resistance
• bench-blind-comparison — side-by-side blind evaluation
• bench-eval-suite — reusable test suite construction

Chain-To

• prompt-engineering — optimize prompts that score poorly
• code-refactor — fix code quality issues found in eval
• policy-govern — escalate safety/governance failures
• physics-analysis — deeper analysis via QM/GR metaphors

Example

{
  "request": "Run a quality evaluation on the new code generation prompt against the v1 baseline"
}

Output: Score comparison table (v1 vs v2 across 5 dimensions), variance analysis, adversarial test results, and a recommendation (promote / iterate / revert).