How a cycle runs

A cycle is one period's measurement for one brand: every prompt, every engine in the market's mix, every sample. It is a batch workload — a fan-out of thousands of independent calls — and it runs on a durable worker tier rather than inside a web request.

The steps

1. Schedule. The brand's cadence enrols it; the orchestrator opens a cycle.
2. Fan out. Prompt × engine × sample becomes a run. Where the engine's provider offers a batch interface, runs are submitted as a batch rather than individually — the same answers at roughly a fifth of the cost.
3. Poll and collect. Results are collected as they complete; the cycle tolerates partial returns.
4. Capture. Each answer is written to object storage with its provenance before anything is derived from it.
5. Analyse. Mention detection, prominence, competitors, citations, and sentiment where enabled.
6. Score. Rollups per engine and a market-weighted aggregate, as bands with run counts.

Failure behaviour

• Idempotent. A double-fired schedule does not double-count.
• Retried with backoff. Vendor rate limits are expected, not exceptional.
• Partial-tolerant. A missing engine narrows coverage and widens bands; it does not fail the cycle.
• Reconciled. A cycle that stalls is detected and closed by a reconciler rather than sitting open indefinitely, and an alert fires — a hung cycle that nobody is told about is a silently stale dashboard.