title	RHEINGOLD
emoji	🌬️
colorFrom	yellow
colorTo	gray
sdk	docker
app_port	7860
pinned	false
license	mit
short_description	Cited IC memos for German onshore wind farms

RHEINGOLD

Underwriting the Energiewende. Pick any real wind farm in Germany → get a full, cited investment-committee memo — valuation, debt capacity, stress tests — in about two minutes.

Why this exists

Every input is live public German data: the MaStR registry, SMARD prices, BNetzA auctions, Netztransparenz market values. The finance core is deterministic and auditable; the agent layer only narrates evidence it can cite. And the availability and O&M assumptions come from someone who spent two years building manufacturing data systems across 10+ wind plants at Suzlon — watching real turbines fail and recover.

Architecture

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]

Loading

The deterministic engine (packages/engine) is pure Python — no network, no LLM, no unseeded randomness. Agents read a frozen EvidenceStore and argue about it; a citation-integrity validator they cannot talk their way past rejects any memo whose numbers don't trace to evidence.

Screenshots

Data sources & licenses

See docs/DATA_SOURCES.md. Highlights: MaStR (DL-DE/BY-2.0), SMARD (CC BY 4.0), Netztransparenz (attribution), renewables.ninja (CC BY-NC 4.0 — this project is non-commercial).

Results

From make backtest (§12, deterministic, seed 42): the median model bid tracks the observed BNetzA award-price trajectory 2017–2024 with MAE ≈ 1.10 ct/kWh across the 22 priceable rounds (directional hit-rate 0.55). Two honest failure modes are part of the result: 2019–2020 rounds run ~1.3–1.6 ct below actual awards (undersubscribed rounds pushed clearing prices to the Höchstwert — a winner's-curse-free regime the break-even model cannot see), and the 2021–2023 rounds are unpriceable under a flat forward at crisis-year spot levels (any bid ≥ 0.5 ct cleared the hurdle) — recorded as skipped-with-reason in the output rather than hidden.

Limitations

The honest section. See docs/MODEL_CARD.md: annual (not monthly) debt model, two-pass tax (no Zinsschranke, no loss carryforward), single-int §51 modeling, merchant tail without a forward curve, Path B resource error, and every backtest caveat (winner's curse, undersubscribed rounds, binding price caps, site-selection bias).

Run locally

cp .env.example .env       # add ANTHROPIC_API_KEY for memos (optional)
make data                  # MaStR + SMARD + Marktwerte + fleet build (~30 min first run)
make dev                   # web :3000 + api :8000
make engine-test           # the golden farm must always pass

Author

Siddharth Jain — manufacturing → capital. Das neue Rheingold ist Wind.