A deterministic engine, shown in full

The engine, formula by formula

E_p50 = mw_total × 8760 × p50_cf × (1 − wake) × availability
        × (1 − elec_losses) × (1 − curtailment)

E_t = E_p50 × (1 − degradation)^(t − 1)

σ_total = sqrt(Σ σ²)        # independent σ's in quadrature:
                            #   wind-data/method σ  (Path A: 0.06, Path B: 0.09)
                            #   loss uncertainties   0.03
                            #   inter-annual var.    0.06 — NOT in long-term P90;
                            #                        reported separately as 1-yr P90

P90 = P50 × (1 − 1.2816 σ_total)
P75 = P50 × (1 − 0.6745 σ_total)

Default assumptions

Mirrors packages/engine/rheingold_engine/defaults.yaml field-for-field — every value carries its source. Vintage-dependent values (capex, opex, interest) are overridden per farm from cost_vintages.csv by commissioning year.

Data vintages

Live from GET /api/health — the max date of each mart table baked into the running API.

Data lineage

The pipeline, as committed in docs/architecture.mmd (Mermaid source, rendered by GitHub in the README). Registry and market data flow left to right; the language models only ever see the frozen EvidenceStore.

flowchart LR
  A[MaStR] --> P[pipelines]
  B[SMARD] --> P
  C[Netztransparenz] --> P
  D[BNetzA auctions CSV] --> P
  E[Wind resource: ninja / GWA] --> P
  P --> M[(DuckDB mart)]
  M --> API[FastAPI]
  API --> ENG[Deterministic engine]
  ENG --> EV[(EvidenceStore)]
  EV --> GATE[Compliance gate - code]
  EV --> CR[3 critics - Claude]
  CR --> N[Narrator - Claude]
  GATE --> N
  N --> V{Citation validator}
  V -->|pass| UI[Next.js: map / dossier / memo]
  V -->|fail| CR
  UI --> H[Human veto]

Data sources & licenses

renewables.ninja profiles are CC BY-NC 4.0 — this project is strictly non-commercial. Full lineage and limitations: MODEL_CARD.md.