Waterfall Forensics.

Waterfall Forensics

We ran the contract forward for eight years. It matched to the penny.

Sequoia Mortgage Trust 2013-1: deconstructed from its prospectus into an executable encoding, then replayed against every monthly distribution the trustee ever reported — 93 periods, February 2013 through final termination.

$363M

of actual distributions reproduced

99.7%

of 2,139 comparisons within $1

$0.08

worst principal delta, ever

Every period, every class — computed from the contract vs. reported by the trustee

2015-03-25 · 16/93
ClassInterest computedInterest reportedPrincipal computedPrincipal reportedΔ
1-A1102,540.32102,540.321,253,707.391,253,707.390.01
1-AX148,555.46148,555.53——0.07
2-A1239,778.29239,778.298,356,712.888,356,712.880.01
2-AX240,360.09240,360.01——0.08
B-128,880.6928,880.6927,477.6827,477.680.01
B-219,253.8019,253.8018,318.4618,318.460.01
B-313,025.6313,025.6312,392.8512,392.850.00
B-49,626.909,626.909,159.239,159.230.00
B-512,704.7312,704.750.000.000.02

drag through 93 distribution dates — Δ is the summed absolute difference between the contract-computed and trustee-reported figures for the period

Or see all 93 periods at once: the tie-out heatmap ›

How

The governing documents are deconstructed into a provenance-linked encoding — every waterfall step, trigger, and floor provision cites the clause that defines it. A deterministic interpreter executes that encoding, period by period, against the collateral cashflows the trustee reported. State rolls forward on computed values only; an error anywhere would compound and surface. It doesn't.

The same engine then runs Monte Carlo counterfactuals — what should have happened under different prepayment, default, or servicing behavior — and exports the whole model as readable Python your own analysts run and defend. Every encoding is open to clause-by-clause verification, and every verification lands in an append-only ledger.