Across the full life of a contract, Zena's blended diffusion has the lowest mean absolute error of any model tested — beating both the pure information-bridge and the Jacobi / Black–Scholes baseline a standard desk quotes from.
Mean absolute error by regime (cents per $1 of notional)
| Model | Interior | Transition | Pinning |
|---|
| Bridge | 5.39 | 5.06 | 0.63 |
| Jacobi | 1.56 | 2.74 | 6.88 |
| Zena hybrid | 1.21 | 1.78 | 0.93 |
A diffusion model misprices ATM calls in the pinning regime by ≈6.9¢ per dollar of notional — it never sees resolution coming, so it misses the binary-payoff limit p(1−K) and the implied vol that diverges like Σ*/√(τ−t). Zena's blended diffusion inherits the bridge's pinning where it matters and the Jacobi dynamics in the interior, and is provably no worse than the bridge on any market.
Ground truth: a conditioned-Jacobi data-generating process favoring neither model, cross-checked against the 50-market Polymarket/Kalshi tick panel. Fitted information rate β̂₀ ∈ [0.31, 2.14], mean 1.19 — order unity across three orders of magnitude in volume. The structural prediction σ_imp → Σ*/√(τ−t) is confirmed at z ≈ 2.3 (p ≈ 0.02). The engine in this demo is the one benchmarked here.