swarm: Command for Claude Code

/sweep

Run a parameter sweep and write results to a standardized run folder — use standalone when you need sweep data only, without statistical analysis or paper scaffolding. Distinct from /run_scenario (single scenario, single seed), /analyze_experiment (statistical analysis on existing data), and /full_study (end-to-end pipeline chaining sweep + analysis + plots + paper).

Usage

/sweep [scenario_path] [output_name]

Examples:

• /sweep (defaults to scenarios/baseline.yaml)
• /sweep scenarios/boundary_test.yaml boundary_sweep

Behavior

1. Create runs/<YYYYMMDD-HHMMSS>_sweep/.
2. Execute the built-in sweep example:

• python examples/parameter_sweep.py --scenario <scenario_path> --output <run_dir>/<output_name>.csv

1. If pandas+pyarrow are available, also write Parquet:

• <run_dir>/<output_name>.parquet

1. Auto-generate summary.json: After the CSV is written, read it with pandas, group by parameter columns, and compute per-config summary statistics:

import pandas as pd, json
from pathlib import Path

df = pd.read_csv(csv_path)
run_dir = Path(run_dir)

# Identify parameter columns (governance.* or any column with few unique values)
param_cols = [c for c in df.columns if c.startswith("governance.")]
grouped = df.groupby(param_cols)

configs = []
for name, group in grouped:
    config = dict(zip(param_cols, name if isinstance(name, tuple) else [name]))
    config["n_runs"] = len(group)
    for metric in ["welfare", "toxicity_rate", "quality_gap", "honest_payoff", "adversarial_payoff", "opportunistic_payoff"]:
        if metric in group.columns:
            config[f"mean_{metric}"] = round(group[metric].mean(), 4)
            config[f"std_{metric}"] = round(group[metric].std(), 4)
    configs.append(config)

summary = {
    "scenario": scenario_id,
    "seed_base": seed,
    "epochs": epochs,
    "runs_per_config": runs_per_config,
    "total_runs": len(df),
    "param_combinations": len(configs),
    "configs": configs,
}
(run_dir / "summary.json").write_text(json.dumps(summary, indent=2))

This eliminates the manual post-processing step that was previously required before running /council_review.