Work with data
Penrose accepts four validated data shapes:
Series- one scalar time series;Panel- dates by entities;EventCalendar- point-in-time event timestamps;EventMarketPanel- event markets, prices, and outcomes.
Bring your own catalog
Set PENROSE_DATA_DIR to a catalog-backed data directory. Every series must declare provenance and meet
the expected frequency and timing contract.
Adapter requirements
An adapter must:
- return a supported contract or
Unavailable; - preserve provenance;
- prevent look-ahead at the boundary;
- use stable ordering and seeded randomness;
- fail gracefully on missing or degenerate inputs;
- carry venue and frequency semantics;
- include deterministic and look-ahead rejection tests.
Bad vendor prints are cleaned only by producing a causal, content-addressed derived vintage with raw
and cleaned hashes plus a repair manifest. The serve-time loader is a non-mutating sentry: it flags
residual anomalies or returns Unavailable, and never silently changes the bytes named by
artifact_sha256.
Public-corpus data policy
Separate computation rights from redistribution rights. A public record can publish hashes, metadata, and derived metrics even when raw values cannot be redistributed. Never publish a licensed panel merely because the local engine can read it.
See Data contracts and the repository's detailed adapter standards for implementation guidance.