SciGame is a controlled benchmark that converts scientific literature into structured scientific cards. Each card provides a pair of original and dimensionally consistent counterfactual environments with the same task interface. The benchmark contains a historical collection of canonical laws and a temporal collection curated from papers published after the model's knowledge cutoff. Evaluation is split into candidate‑guided identification and open‑ended discovery settings, each run under prior‑only, fixed, and adaptive protocols. Experiments with five frontier models yield three main findings: familiar priors hurt performance when the governing relation changes; adaptive experimentation boosts the strongest models under limited budgets but harms weaker ones; good explanations can hide quantitative errors. Even when agents interpret evidence correctly, recovering the exact symbolic form remains the main bottleneck. Current agents can actively investigate systems, yet inducing exact laws in unfamiliar settings remains unsolved. Review