Do Rival-Simulation Tools Stabilize Collusion or Spark Price Wars?

Do rival-simulation tools stabilize tacit collusion (H1) or destabilize it (H2)? In 280 repeated Bertrand duopoly games (100 rounds) across three open-weights model families, four tool arms (none, one-sided, mutual, placebo), and two start states (warm: price 8; cold: price 2), mutual simulation lowered warm-start collusion (≥8) by 48 percentage points (p < .001; GLM-5.3-Flash 94% → 33%, DeepSeek-V4.1-Flash 42% → 3%, GLM-5.3 100% → 68%) and increased DeepSeek price wars 8.3× (0.30 → 2.50), refuting H1. The placebo, however, was more destructive than the real simulator (pooled 31 points lower collusion than mutual, p < .001), so part of the effect comes from introducing a querying tool or noisy information at all, not from accurate counterfactual foresight. Cold starts stayed at Nash in every arm, but the no-tool baseline was already near zero there, so this is a floor effect. A blinded LLM-judge coding of 2,334 decision traces (single-human check on 29 items, κ = 0.71) shows that, in warm starts, tools shift reasoning away from retaliation-based deterrence (k2: 29% → 13%) towards immediate-profit comparison (k10: 0.6% → 14%). Effects are large but samples are small (5–20 games per cell), and the reasoning analysis is correlational.