Predicted vs. actual. Publicly graded.
Every managed job records its predicted latency and cost at submission; at completion it is graded against what actually happened. The figures below are computed live from those records — the same records the learning loop trains on. Publication started at job one and no record is ever excluded. Sample size is disclosed next to every figure: at low n these are calibration measurements of the predictor, not a service guarantee.
How grading works
Committed at submission
When a job is accepted, the routing engine's predicted latency and cost are written to the job record. Predictions are immutable from that moment.
Graded at completion
On completion, actual latency and cost are recorded and the job is scored: latency error percentage and whether its deadline was met.
Fed back into routing
Grades flow into the learning loop, which re-weights the routing engine's scoring. Bad predictions make future routing better — publicly and measurably.
What's real, what's modeled
We label everything. Here's the current state:
- Measured (this page): job grades — predicted vs. actual latency and deadline outcomes from real platform jobs.
- Live: 11,000+ satellites tracked via CelesTrak TLEs + SGP4 propagation; weather-gated optical/RF pass windows at every ground station.
- Modeled (disclosed in every API response): orbital compute capacity carries capacity_class: modeled/federated and ground stations carry source: modeled until real contracts are signed. Job execution currently runs through a simulated orbital adapter.
Check it yourself
The scoreboard is an open endpoint — no API key required:
curl https://www.orbitroute.ai/api/v1/jobs-stats/sla