Demos

The solver, ported to dependency-free JavaScript and run live. Each page says what would falsify it.

All three demos call bb_core.js, a direct port of brownianbandit/mean_field.py held to the Python at stated tolerances by tests/test_js_parity.py. Nothing is pre-baked; every chart is computed when you move a control.

• The budget solve
  Survivor schedule, wavefront cutoffs, and terminal cloud at any budget and intensity. The certificate shown is the Frank–Wolfe gap plus the budget-slackness term λ(B−C); falsified if a baseline beats the returned policy by more than that bound.
• Watch a race
  Sampled paths killed live by the solved wavefront policy, replaying continuously. Falsified if the running mean of realized maxima leaves a time-uniform confidence band around the predicted expected maximum; ordinary 2-s.e. crossings are expected under continuous monitoring.
• The budget frontier
  Expected maximum against budget for the adaptive policy and both baselines, ten full solves per sweep. Falsified by a baseline exceeding the adaptive curve by more than that point's certificate; ties at both ends of the budget range are legitimate.