Incoherence Scaling: Initial Experiment Artifact

This note describes the initial scaling artifact generated by:

examples/run_incoherence_scaling.py

Scope

• Runs replay sweeps over three scenario tiers:
• short|low branching
• medium|medium branching
• long|high branching
• Uses ReplayRunner with configurable --replay-k.
• Exports:
• incoherence_scaling_replays.csv
• incoherence_scaling_summary.csv

Current Incoherence Signal

The current script uses a temporary proxy from replay toxicity variation to produce: - incoherence_index - error_rate - disagreement_rate

This is a bridge until full step-level decision replay integration lands in the scaling runner.

Intended Next Upgrade

• Replace proxy signal with decision-level replay metrics from swarm/metrics/incoherence.py:
• benchmark-based E
• replay disagreement D
• I = D / (E + eps)