How to read this. These outcomes are authored, not sampled from negotiations. No trial re-runs the multi-round talks. Each trial draws random conditions, and a single model call then writes the outcome it judges plausible under that draw, in a structured form; we aggregate those written outcomes. The distributions below are therefore the model's stated beliefs about sensitivity to conditions, not an empirical distribution of simulated negotiations, and not forecasts. Wide spread = outcome is condition-sensitive; tight spread = robust. Satisfaction is the model's 0–1 rating per delegation. Exploratory and educational only — see the
methodology.