Fleet keeps one authoritative record of what exists, what is running and what is reserved — a single source of truth every part reads from and writes to. A single-writer store sits behind a small service; resources, reservations, jobs, assets and events all live in one place, so there is never a second, disagreeing copy of reality. By funnelling all state through one authority, Fleet avoids split-brain: a reservation either exists or it does not, a worker is either busy or free.
Everything meaningful that happens is measured and recorded, not guessed at. Fleet emits metrics on jobs, durations, resource use, successes and failures, so it knows a GPU is actually loaded and a job actually finished — rather than assuming it. Telemetry feeds the evidence layer: a metric is tagged with how it was obtained, so a dashboard never quietly mixes a real reading with an estimate.
Every output — an image, an audio file, a document — is recorded as an asset, with where it came from and how it was made. Each is linked to the job and worker that produced it and graded for whether it was real or simulated, so outputs are tracked, attributable results rather than loose files. Because each asset records its provenance, you can always answer “where did this come from?” — which is where Fleet’s work becomes evidence.
Fleet is event-first: significant changes are recorded as an ordered stream, so there is always a replayable history of what happened and when. Reservations made, jobs started, assets produced, leases expired — each is an event; the current state is the running sum, the timeline is how it got there. An append-only history means you can audit, debug and replay — nothing important happens without a trace.