Guard the fleet.
The response layer: what to do about it, from the smallest dodge that answers a threat to where your sensors look next, and the same watch pointed at your own telemetry. No breach, no burn.
Three models, and a recommendation you can refuse.
Every response carries its cost and the miss it buys, so the call stays with a human. Open any model for the mechanism behind it.
-
/0.6 Defensive maneuver response 0.15 m/smedian dodge Δv
Recommends the smallest burn that answers a warning: it keeps a threatened asset outside its keep-out volume, within the propellant it has left, and makes no recommendation at all for a pass that already clears. Each carries its direction, timing, fuel cost, and the miss it buys. Scored on three labelled-synthetic campaigns and two benign controls.
-
/0.7 Sensor tasking trained scheduler
beats myopic + randomTurns the warning queue into a schedule: which object each sensor looks at next, so a network spends its scarce minutes where the uncertainty and the threat are, not on whatever is overhead. Trained on held-out scenarios with real visibility geometry, it breaks from the greedy pick to catch a closing pass that would otherwise be stranded. This is the kill-chain gap the government keeps naming.
-
/0.8 Spacecraft health monitoring OPS-SAT benchmark
anomaly detectionThe same discipline pointed inward, at a satellite's own telemetry: flag the segment where a channel breaks its nominal pattern, a spike, a stuck sensor, a slow drift, before an operator sees it. The detector is fit only on healthy history, so it needs no catalogue of past failures to recognize a new one, and it is scored in the shape of the public OPS-SAT benchmark.
What each model is graded against.
A response layer is only worth what its benchmark is worth, so here is the benchmark. These are the corpora and the controls, named, before any number is quoted.
- Defensive maneuver Three labelled-synthetic threat campaigns and two benign controls. The controls matter as much as the campaigns: a model that recommends a burn for a pass that already clears has failed, and the benign cases are what catch that.
- Sensor tasking Held-out scenarios with real visibility geometry, scored against both a myopic greedy baseline and a random floor. The trained scheduler stays a benchmark claim: the Command Center schedules the watch queue, it does not command sensors.
- Spacecraft health A corpus in the shape of the public OPS-SAT benchmark, with the detector fit only on healthy history so it needs no catalogue of past failures. Real telemetry drops into the same interface when an operator brings it.
No chart on this page. The figure exporter emits the one-sidedness discriminator and nothing else yet, and a chart that does not exist is not going to be drawn to fill a space. The corpora above are what GUARDIAN is graded against, named before any number is quoted.
See how a call is separated from a benign case →Bring us a fleet. We will show you the dodge.
Book a technical briefing and we will run GUARDIAN against your protected assets, with the propellant cost and the miss it buys on every recommendation.