Live stats.

The catalog, the per-check verification rows, the cross-checks, the token ledger and the event mirror — queried with SQL, including an FTS5 index over the prompt text.

Caches the serialised catalog and each prompt document for five minutes, so a cold read never has to reach D1.

Raw model responses, pre-repair snapshots and the ledger archive — the evidence behind every number on this page.

One instance per prompt slug, serialising copy and view counts that arrive from every edge location at once.

Absorbs verification bursts before they become Workflow starts; exhausted retries land in a dead-letter queue.

The durable build → verify → cross-check pipeline, checkpointed step by step so a long run can resume.

Generates the embeddings behind semantic search (@cf/baai/bge-m3).

The vector index those embeddings are queried against.

High-cardinality product events — copies, views, searches — written on every interaction.

Runs the headless acceptance harness at the edge instead of on a laptop.

120 requests per minute per key on the public JSON API and the write endpoints.

Request-time optimisation of the verification screenshots.