Certificates

The test suite is the library’s claim. Each certificate checks a frozen limit against QuantLib and the switching case against an independent referee; the tolerances below are what the tests assert.

Model / instrument	Frozen limit (QuantLib)	Switching referee
Vasicek, CIR bonds	discountBond 1e-10	numerical solution; Monte Carlo with exact regime paths
Hull–White, G2 bonds and bond options	discountBondOption 1e-7	expansion converges to the numerical solution; Monte Carlo (G2)
Black–Scholes, Heston, Merton, Bates, variance gamma	analytic engines 1e-8	numerical solution; averaged-parameter limit
Digitals, implied volatility, geometric Asian	analytic engines 1e-8	numerical solution; Monte Carlo (Asian)
American, barrier	FdBlackScholesVanillaEngine 5e-5, AnalyticBarrierEngine 3e-5	Brownian-bridge Monte Carlo
Swaptions, caps, coupon-bond options	JamshidianSwaptionEngine, AnalyticCapFloorEngine 2e-6	grid European equals Jamshidian with switching
Bermudan swaptions	FdHullWhiteSwaptionEngine 1e-4, FdG2SwaptionEngine	Bermudan dominates every European; grid convergence
CDS, first-to-default	MidPointCdsEngine 2e-4	frozen basket factorises; Monte Carlo joint survival
Hybrid equity–rates	AnalyticBSMHullWhiteEngine (with correlation), AnalyticHestonHullWhiteEngine 1e-6	expansion converges; parity with the switching bond
CEV, Heston vol-of-vol (first-order tier)	AnalyticCEVEngine, AnalyticHestonEngine (grid error)	coupled switching PDE; error tracks the estimate
Symbolic formulas	—	engine at the same order to 1e-12; finite differences of the numerical solution
Calibration	—	recovers a two-regime smile’s parameters exactly

Run pytest for everything (about a quarter of an hour) or pytest -m 'not slow' for all but the four slowest.