Trust & policies

Every imported fact carries a provenance row: source, dataset, version, retrieval date, evidence type, license and — where relevant — PMID or DOI. Raw payloads are archived unchanged. Hover any source badge to see the chain; administrators can trace any value to its raw record.

Derived values reference a versioned formula from the public metric catalog. Snapshots store their inputs hash; identical inputs yield identical ranks.

Observed Published Curated Regulatory Guideline Computed AI-generated

Labels are never merged. Regulatory status is always tied to a jurisdiction and an indication. Population survival is never presented as an individual prognosis.

Phase 1 contains no AI-generated text. When AI synthesis is introduced, it will be database-grounded only (answers must cite records present in the index), cached with the model name, prompt version and record snapshot, and always labelled "AI-generated". Language models are never used to map source labels to entities; at most they propose candidates that a curator reviews.

CancerIndex holds no patient-level data: all statistics are aggregate and published by their sources. The site sets no tracking cookies; the only cookie is the administrator session. Server logs retain IP addresses for a short period for abuse prevention.

Read-only public access; administrative actions require a token and are written to an audit log. Connectors run with rate limits, idempotent writes and an anomaly guard that refuses destructive updates when a source shrinks unexpectedly. Report vulnerabilities to contact@spboucher.ai.

A connector goes live only after its license is reviewed; sources under review are listed with no public records. Attribution requirements are honoured on every page and in exports. See the source registry.

Found an error, a mis-mapped label or a licensing concern? Write to contact@spboucher.ai with the page URL and the identifier (CI-…). Corrections are recorded as change events on the affected entity.