Is the platform real-time or batch? A generation just took the better part of a minute — is that a limitation or a choice?

The full executive answer

It is a hybrid by design, and the latency profile matches the decision cadence — which is the right engineering answer, not an excuse. Reading is instant: dashboards, stored insights and any analysis whose inputs have not changed return from cache immediately, because the cache is keyed on your actual data state. Deep generation — a board-grade scorecard synthesis or a phased roadmap — is deliberately a considered operation: tens of seconds to produce, because in that window the system may run a second provider for consensus, validate every number’s provenance, check guardrails, and possibly execute a corrective retry. A board decision brief that took forty-five seconds to verify is not slow; a board decision brief produced in two seconds without verification is reckless.

The scheduled layer handles what should not wait for a human: recurring jobs monitor generation quality and alert on degradation, score due forecasts against new reality, aggregate cost attribution, and refresh external intelligence. And for resilience there is an optional queue mode — generations persisted as jobs that a worker claims and completes with bounded retries and backoff, so long generations survive infrastructure cold starts and provider spikes. What we deliberately do not offer is streaming token-by-token prose for board artifacts: an output is either fully validated or it does not exist, and partial unvalidated text on a boardroom screen would break that contract. Boards decide in meetings, not milliseconds — the architecture is honest about that.