Data Integrity

Reputation and risk data is stored write-on-change with full history: deployer reputation snapshots and wallet-flag histories record what we knew and when we knew it. Supported endpoints accept ?as_of=, answering "was this wallet flagged at the time?" — the question a listing review or backtest actually needs — without retroactive relabeling. Nightly signed dataset manifests pin what each dataset contained on each date (GET /api/v1/manifests, keyed).

Every v1 API response carries an X-Request-Id, logged in an audit trail — 90 days standard, 730 days on Business keys. When a number ends up in your product, your report, or in front of a regulator, you can point at the exact request that produced it.

Every closed month of trade tape and candles is exported to Parquet with per-file SHA-256 checksums in a manifest ledger, stored with two independent providers under S3 Object Lock — write-once retention that we cannot modify or delete during the lock window, by construction. Read-back drills restore and re-verify checksums from cold storage on a schedule.

Correctness is a monitored production system, not a launch-day claim: 136 checks every 15 minutes, a 207-case suite twice daily, and 425 cases nightly compare served values against on-chain ground truth and independent references. Results are public, live, and include the failures.

Latency claims come with sample sizes, vantages, and error bars — including a continuously self-audited detection head-start metric reconciled against every confirmed launch. Where a comparison was one-off, the page says so and dates it.

Three dedicated EU servers across two providers, a warm standby replicating with ~1 minute of data loss tolerance, and a rehearsed full restore (roughly 3–3.5 hours, drilled monthly from off-site backups — measured, not assumed). Deploys are blue-green and zero-downtime; the paid WebSocket streams run on separate processes that deploys never touch.