Benchmark

Autodialectics Benchmark & Evolve

Use this skill to benchmark policies and drive champion/challenger evolution.

Prerequisites

autodialectics-mcp must be on PATH (pip install autodialectics).

MCP Workflow

Benchmarking

1. benchmark(suite_dir?, policy_id?) — run the benchmark suite against a policy. Returns case-by-case results with scores and decisions.

Policy Evolution

2. evolve_policy(use_gepa?) — analyze recent benchmark reports and create a challenger policy. Set use_gepa: false to skip the GEPA optimizer (simpler heuristic fallback).
3. promote_policy(policy_id) — promote a challenger to champion if comparison rules allow.
4. rollback_policy() — revert to the previous champion if the current one regresses.

CLI Fallback

autodialectics benchmark
autodialectics evolve
autodialectics promote <policy_id>
autodialectics rollback

Typical Evolution Cycle

benchmark → evolve_policy → benchmark (with challenger) → compare → promote or rollback

1. Run benchmarks with the current champion to establish a baseline.
2. Evolve a challenger from the benchmark reports.
3. Run the same benchmarks with the challenger's policy ID.
4. Compare results. Promote if the challenger wins; rollback if it regresses.

Guidance

• Never claim a policy is better without benchmark evidence from the same suite.
• When reporting benchmark results, include: total cases, pass/fail/revise counts, mean overall score, mean slop composite.
• If evolve_policy returns no_reports, run benchmarks first to generate data.
• Promotion can be denied by comparison rules — check the response status.

Arguments

If the user passes a suite directory after /autodialectics:benchmark, use it as the benchmark suite path.