Prices, fleet build-out and dispatch for each duration cohort are solved jointly to a fixed point. Each cohort’s dispatch is a linear programme, solved by dual simplex with warm starts on CEMeridian, and the outer loop is a Krasnosel’skii–Mann iteration with step 2/(k+2), because plain fixed-point iteration fails to converge once the fleet passes about 2 GW. The naive first pass, which holds prices fixed, is always reported next to the equilibrium. The method is set out in CE-WP-2026-01.
The revenue stack covers frequency response across DC, DM and DR with saturation pricing, the Capacity Market with duration de-rating and pinned clearing prices, and a Balancing Mechanism proxy measured zone by zone from actual battery cashflows. Calibration verdicts are published, including the fails.
Forward prices are anchored in three disclosed layers: a trailing anchor to realised prices, refreshed automatically each week; official government fuel-price paths for stress tests; and your own licensed forward curves if you bring them.
Each run is keyed by scenario, code version and timestamp, with a stored manifest. A solve that is not optimal stops the run. Scenario years and site appraisals have to pass a suite of validation invariants, and datapacks export the full audit trail. Every CEMeridian build is cross-checked against an independent open-source solver before CECadence uses it.
Validation
There are ten graded backtest results against realised GB prices (the prices that actually settled) from 2022 to mid-2026, each published as a pass or a fail. The model’s annual-mean price landed within ±5% of the realised outturn in three of five windows. Two windows were held out, with their rules fixed before any of their data was fetched: 2025 landed at 1.6%, and 2026H1 missed our own 5% bar and is published as it landed. The scarcity tail, the highest-priced hours, is graded and published the same way, and it missed in both windows.
Delivery
CECadence runs as multi-tenant SaaS with hard tenant isolation, or air-gapped on your own servers under an offline signed licence. Outputs include branded scenario reports, per-site appraisals with the DSCR series and binding year, head-to-head site comparisons, and CSV datapacks with the audit manifest.
