Checked Sun, 16 Aug 2026 17:16:35 GMT · probe took 1519 ms
Database
Replay protection and rate limiting
Evaluation API
HB-Eval runs as a single instance in one region, with no redundancy and no automatic failover. Service is best-effort: there is no uptime guarantee, no support commitment, and no compensation for downtime.
That is a smaller promise than most platforms make. It is also the true one, and a project arguing that reliability claims should be measured rather than asserted is a poor place to start making unmeasured ones.
Practically, this means: evaluation and monitoring can be interrupted by a deploy or an upstream incident, usually for minutes. Metrics already computed are not lost — the SDK computes them locally and the platform stores results — but a run in progress may fail to submit.
| Evaluation results | Kept until you delete them or your account |
| Monitoring sessions | Kept until you delete them or your account |
| Step-by-step snapshots | 90 days, then removed automatically |
| Alert delivery records | 90 days, then removed automatically |
| OAuth tokens | Removed on expiry |
| Observatory contributions | Anonymised at write time — no identifier is ever stored |
Last automatic cleanup: Sun, 16 Aug 2026 13:53:37 GMT. Published so the policy above can be checked rather than taken on trust.
Export — everything held about your account, as JSON, from Settings. No request, no wait.
Deletion — removes the account and everything linked to it. If you chose to donate anonymised results to the Observatory, those are copied without any identifier before deletion proceeds, and the deletion is cancelled if that copy fails. Data promised to be kept is never lost by an operation meant to remove something else.
Observatory — off by default. Contributions carry no user or agent identifier, and aggregate figures are withheld until at least five independent accounts have contributed.
Gateway protocol 2.7.0