Penrose documentation
Penrose is an independent falsification referee for quantitative trading claims. It reconstructs a claim, tests the evidence under realistic statistical and operational constraints, and returns a calibrated verdict. It is not a trading system, signal service, or promise of profitability.
This directory is the public-facing documentation source. It is intentionally separate from
docs/, which contains research papers, engineering standards, and internal design material. The
same Markdown can be rendered as a GitBook-style website or loaded inside the Penrose dashboard.
Interpretation contract
A Penrose verdict describes how well a claim survived falsification. It never means “buy,” “sell,” “profitable,” or “safe to trade.” Promotion at P9 always requires a human.
Start here
The system in one minute
source or structured claim
↓
P1–P3 ingest, ground, and check falsifiability
↓
P4–P6 screen, bind data, and reconstruct faithfully
↓
P7 deflation, power, robustness, costs, and confirmation controls
↓
P8 calibrated verdict or honest routing state
↓
P9 human-only review and promotion
Penrose records the complete search behind a result. A strategy selected from 500 attempts must pay for all 500 attempts; it cannot present itself as a lucky singleton. This search-aware accounting is the central distinction between Penrose and an ordinary backtester.
What to trust
Penrose separates three kinds of state:
- The scientific record — append-only decisions, receipts, audit logs, trial accounting, source archives, and holdout burns.
- The active corpus — evidence eligible to inform the current learning epoch.
- Derived views — dashboard JSON, reports, graphs, connections, and queues that can be rebuilt.
The first is preserved. The second is epoch-scoped. The third is disposable and reproducible.
Documentation map
The complete navigation lives in SUMMARY.md. For application embedding, the equivalent
machine-readable navigation is in navigation.json.
Version and status
Penrose is a 0.x research prototype. Interfaces may evolve, data coverage is incomplete, and
reconstruction fidelity remains a central scientific risk. The public documentation should always be
published from the same commit as the running engine and display that engine version visibly.