Reading order:capability→ responsibility→ implementation technology
Research architecture- Rust handles reusable causal computation; Python remains the research/ML control plane.
- Arrow/Parquet provide the typed cross-language evidence boundary; DuckDB/Polars query it cheaply.
- NautilusTrader owns execution simulation semantics rather than notebook-specific fill logic.
Production architecture- FastAPI/Python own signal orchestration; LightGBM is a fail-closed quality filter, not signal authority.
- PostgreSQL/SQLite persist lifecycle and reconciliation so state is not inferred from presentation.
- Exchange adapters sit behind idempotent intent and hard-risk boundaries.
Product boundary- Next.js/TypeScript, content pipelines and access services consume approved projections only.
- FastAPI API/CLI/MCP surfaces are read-only and cannot mutate private trading controls.
- Technology choice follows authority ownership rather than repository topology.
Curren cross-system architectureTechnology annotations show where each implementation choice belongs — not a logo wall