InterveneLab: World-Model-Assisted Reasoning over Structural Intervention Evidence

We introduce InterveneLab to study how world-model forecasts support an agent's reasoning over structural intervention evidence. The environment contains 3,150 questions over 21 protein structure cases and 105 targets. Questions concern single-action, order, and interaction effects on local density response; budgeted queries return archived observations or terminations caused by geometry or density-map-support checks. A Qwen3.5-4B agent, adapted with supervised learning and on-policy reinforcement learning, receives observation-conditioned forecasts from a learned structural model or an analytic prior. We pair interactive query traces and replay identical acquired evidence to measure how predictive context affects final judgments. On 300 development questions at budget (K=4), learned context raises strict success from 252/300 to 267/300. At this budget, the tools-only agent already acquires complete evidence for every defined question. Only two query paths change, compared with 21 final decisions; 11 of the 14 additional correct numerical answers concern exact-zero effects. With evidence held fixed, persistent learned and analytic contexts yield 110/150 and 117/150 correct judgments, respectively, versus 104/150 without context. The learned model's standalone endpoint error is slightly higher than the analytic reference's (MAE 0.0661 versus 0.0627). In this setting, predictive context improves evidence-grounded judgments, and the benefit is concentrated in the final interpretation of acquired evidence.