Election

Use this skill to elect a winner from existing evaluation artifacts. Treat it as the standalone selection step after candidates have already been run and graded. It does not generate candidates, perform research, synthesize evals, or re-run optimization.

When to use this skill

• A workflow already produced multiple candidate configurations and now needs a winner chosen from scored artifacts.
• Multiple candidate configurations already exist as with_skill, without_skill, old_skill, or other config directories inside a skill-eval workspace.
• Each candidate has already been run against authored evals and saved grading.json and optional timing.json artifacts.
• You need a separate election pass that picks the strongest configuration from workspace results instead of generating new candidates.
• The workflow explicitly needs comparison across multiple optimizer outputs, prompt rewrites, or skill revisions without folding that comparison into the optimization runtime.

Do not use this skill to gather datasets, synthesize evals, optimize prompts, or run missing evaluations from scratch. Those remain separate skills.

Inputs

• workspace_dir: root workspace path, a specific iteration directory, or a direct eval directory
• iteration: optional iteration selector when the workspace contains multiple iterations
• manifest_file: optional authored evals/evals.json path for expected eval coverage