BLUM 2.2.0 — verified research and Cloudflare observatory

BLUM 2.2 establishes GitHub as the canonical platform for source, review, CI, documentation and releases. Verified main revisions publish one-way to the existing Hugging Face Engine runtime and the Cloudflare observatory.

The complete product includes Brain, Training Ground, Paper Trading, Alpha and the new Evolution page. Public APIs are read-only at both the gateway and origin; administrative research requires a private runtime key. Backend and gateway report the deployed source revision, and release verification checks that they match.

ASCENT runs bounded calibration experiments with restricted arithmetic candidates, chronological selection, independent holdout gates, persistent lineage and explicit resource limits. It retains research artifacts without automatically replacing a production champion, merging code, starting paid GPU jobs or placing broker orders.

Historical data decisions are reproducible: 75,381 valid equity bars are admitted, five malformed bars excluded, eight missing-volume observations remain missing. Forex conflicts and unsupported weekend observations are excluded, leaving one causal example—insufficient for validation. The public reasoning data and all 95 archived snapshot versions are quarantined for new training/promotion because provenance and temporal-validity gates fail. Existing database and recovery backups remain preserved; their entire contents have not been certified.

Validation includes 838 backend tests, 9 frontend tests, 12 gateway tests, static product build, TypeScript checks, exact source-hash audit and live deployment verification. A historical diagnostic generated 549 genuine baseline forecasts; validation rejected the candidate for insufficient slice sample, no confirmed calibration gain and robustness regression. This diagnostic is not forward-performance or production-promotion evidence.

Product: https://blum.signallayerlabs.workers.dev

Operations and recovery: Deployment. Data decisions: Data quality. Research boundaries: ASCENT.

Recovery verification retained the existing ASCENT generation and restored the financial database from its physical backup. The post-restart count comparison shows 20 fewer recent historical predictions and 60 fewer prediction outcomes than the pre-restart observation. Embedded PostgreSQL uses periodic backups and does not provide zero-loss container restarts; strict durability requires external PostgreSQL. This comparison is a count audit, not certification of every row.