Penrose v0.1.0

Penrose is an independent, power-aware falsification referee for quantitative trading claims. This is the first public release, a research prototype rather than a finished product.

What's in v0.1.0

• The full falsification pipeline: ingestion, grounded claim extraction, sandboxed reconstruction, a robustness and power gate battery, and a calibrated verdict (kill / underpowered / watch / research-supported).
• The statistical core: a Deflated Sharpe Ratio scoped to the size of the search seen, three-fold sign stability, a regime kill-lens, a bootstrap edge interval, a permutation test, walk-forward consistency, cost and capacity modeling, and a single-use per-claim locked holdout gated on significance.
• A self-calibrated detector: placebo, injected-edge, native-breadth, dead-state, and persistence-matched controls plus a multi-null battery (eval 82/82; clone-and-go reproduces it with no key and no network).
• Five working data adapters out of the box: Coinbase, Kraken, and Deribit (keyless live), plus FRED and Alpha Vantage (free key), over a bring-your-own data contract.
• Pennie, the corpus-grounded chat assistant, and the corpus of invalidations.
• Plain-language gate documentation (docs/GATES.md), agent onboarding (AGENTS.md), and two companion papers (systems + evidence standard).

Pre-1.0 expectations

This is a 0.x release: interfaces may change, costs and capacity are modeled rather than measured, and independent replication remains future work. The headline guarantees are described honestly, deflation scales with the search seen, and the holdout is single-use and conservative. Please try to break it and open an issue.