Public corpus model

A public Penrose corpus should be a verified, read-only publication layer—not the private engine exposed to the internet.

Recommended first release

Publish immutable, versioned records containing:

• normalized claim and public-safe source citation;
• verdict or routing state;
• kill reason and bounded metrics;
• claim type and normalized strategy family;
• engine commit and corpus epoch;
• search cohort and denominator;
• receipt and data fingerprints;
• reproduction class;
• source, data, and licensing disclosures.

Do not publish raw private sources, licensed datasets, holdout values, local filesystem paths, model credentials, proprietary strategy code, or unreviewed personal submissions.

Architecture

private Penrose engine
        ↓ allowlist + validation + redaction
signed publication bundle
        ↓
object storage / immutable release
        ↓
CDN + read-only corpus API
        ↓
public website and documentation

The public service should not have credentials capable of changing the private scientific record.

Record lifecycle

Records are append-only. Corrections create new versions with explicit supersession links. Public pages must distinguish active-epoch evidence, legacy records, and historical versions.

Participation model

Start with curator-published records. Add public submissions later through a quarantined intake queue with rate limits, content limits, licensing attestations, moderation, and explicit compute budgets.

Corpus Commons

Portable kill records can serve as advisory negative priors in another installation. Imported records must never create a local verdict, satisfy confirmation, or bypass P9.