Stratum Atlas

Assumptions

Everything the model believes, stated before the math. If you do not buy one of these, do not buy the output.

The bridge needs real deal terms. Debt sizing, rate, tenor, target DSCR, covenants and offtake come from the customer's covered book; if they are absent the model returns UNKNOWN for those outputs instead of guessing terms. No invented DSCR. Ever.
Physical risk enters only through the Delivery (M02) and Cost (M05) distributions; the bridge adds no free parameters of its own.
Every Monte Carlo draw is a complete quarterly cash-flow path: CAPEX phasing, debt sculpting to the target DSCR, interest, amortization, covenant tests per period.
Draws are seeded (mulberry32, default 2000 draws): the same terms on the same pinned record reproduce the identical result.
Interest is capitalized during construction; debt service is sculpted as level payment sized so CFADS / service equals the target DSCR.

Derivation

The engine builds the quarterly waterfall. Revenue per quarter combines contracted offtake and merchant volume at the stated prices; CFADS is revenue less opex. Debt service is sculpted level-payment:

D S = min (\frac{C F A D S}{D S C R _{t a r g e t}}, \frac{D _{0}}{A ( r _{q} , T )}), A (r_{q}, T) = \frac{1 - ( 1 + r _{q} ) ^{- T}}{r _{q}}

where D_0 is sized debt (gearing default 65% of the CAPEX draw), r_q the quarterly rate, T the tenor in quarters. Each period tests DSCR = CFADS / DS against the covenant floor and tracks minimum DSCR, LLCR, and the amortizing balance.

Distributions come from Monte Carlo over the physical models: each draw samples a COD quarter from the Delivery mass and a CAPEX from the Cost moments (inverse-normal via Acklam's rational approximation on a seeded uniform), then computes the full path:

draw_{i} = {N P V, I R R, M O I C, min D S C R, LL C R, L_{d e b t}, L_{e q u i t y}}_{i}, i = 1 \dots 2000

Every output is reported as P10 / P50 / P90 bands across draws, plus the covenant breach probability (share of draws breaching the DSCR floor). Delay questions ('what if it slips 6, 12, 24 months') are re-timings of the cash-flow path with carrying costs and covenant consequences recomputed per period, not haircuts to the answer.

Identification strategy

Why the estimate supports the claimed meaning - and where the model refuses to claim more than the data carries.

The bridge carries no free parameters beyond the physical models feeding it - that is the identification discipline. Nothing in the money answer can be traced to an assumption invented inside the bridge; every distribution in the output has a physical or contractual parent.
Validation is reproduction: on each pilot deal the engine must reproduce the lender's own base-case model to within rounding before any scenario work is shown. A bridge that cannot reproduce the base case has no standing to opine on the stress case.
Terms are identified to a source: termsSource is 'customer_terms' or the run does not happen. There is no third state where plausible-looking reference terms stand in for the deal.

Inputs and outputs

Reads

Deal terms: Customer's covered book (spec Section 28): CAPEX, debt sizing, rate, tenor, target DSCR, covenants, offtake, opex, discount rate
Delivery distribution: Delivery Model (M02): COD mass by quarter
Cost distribution: Cost Model (M05): CAPEX moments
Merchant revenue: Merchant Revenue Model (M07) where volume is uncontracted

Writes

NPV, IRR, MOIC distributions (P10 / P50 / P90)
Minimum DSCR and LLCR distributions; covenant breach probability
Debt loss and equity loss distributions
Draw count and seed (the audit record of reproducibility)

Calibration and gate evidence

Validation design

Reproduction against the lender's own base-case model on each pilot deal, to within rounding, before any scenario output is shown.

Evidence on record

Pilot-deal reproduction results are not yet published in the sealed source. Determinism (seeded replay) is coded; the reproduction gate awaits a pilot book with full terms.

Failure modes

The ways this model declines, degrades or goes UNKNOWN - and what it prints when it does.

No deal terms on the case: every output is UNKNOWN with reason 'terms_not_provided'. The model refuses the run rather than inventing a capital structure.
A draw set with no IRR-sign-changing path: IRR is UNKNOWN, not a misleading number.
Offtake or merchant terms absent: revenue is computed from what exists only; the missing leg contributes nothing rather than an assumed price.

Sensitivity

Delay is the dominant sensitivity: a COD slip shifts the entire revenue path while capitalized interest keeps accruing, and minimum DSCR moves before NPV does.
Gearing and target DSCR jointly size debt service; a 5-point move in gearing can flip the covenant breach share on tight deals.
The CAPEX draw's spread (from the Cost Model's moments) drives the loss tails; on a wide cost prior the equity loss P90 is the honest headline.

Iron rules

The financial bridge needs real deal terms. If the customer has not provided debt sizing, rate, covenants and offtake, the model returns UNKNOWN for those outputs instead of guessing terms. No invented DSCR. Ever.

Source and audit

Written from the sealed model source. Every run is recorded - model version, frozen inputs, input hash, timestamp. There is no unrecorded model execution.