BranchCast: Planning over Named Tool Outcomes Before Latent Generation

A timeout and a silent-wrong completion may be neighbors in a learned latent space, yet a planner should retry one and reject the other. BranchCast preserves this decision-relevant distinction by predicting a named tool outcome—success, invalid call, retry/timeout, or silent error—with recoverability and severe-failure probability before generating within-outcome latent variation. We test the complete interface with a crossed comparison that fixes the encoder, proposal and scoring networks, beam search, horizon, four generator evaluations, and a 20–22 ms active-transition latency band. Across five paired seeds and 2,250 held-out WebArena, non-weather ToolBench, and code-execution decisions, replacing implicit-branch diffusion with BranchCast raises success by +6.5 points (95% CI [+5.4,+7.7]) and lowers severe failure by -6.2 points (95% CI [-7.5,-5.0]). The success gain remains +5.0 points when severity reranking is disabled and grows from +2.6 to +10.8 points across training-defined branch-load quartiles. The interface also changes model selection: rectified flow leads diffusion by 4.0 points behind an implicit interface, whereas diffusion leads same-scaffold rectified flow by +0.9 points once both expose named outcomes. Outcome structure is therefore part of the world-model contract with search, not merely an auxiliary prediction target.