Launch Control Agent

Launch Control Agent

Built from 3,000 simulated launches with the Ammonix platform code.

Synthetic telemetry from a notional heavy launcher (cryogenic core stage, two solid boosters). Not data from a real vehicle; all values are illustrative.

H0 − 00:60.0
Final countdown
Agent call
GO
Fixed limits (redlines)
Clear

Grey band: 5th-95th percentile of 120 nominal launches. Hover or drag across the traces to read values at any moment. Gaps are lost telemetry frames.

Layer 3 · Knowledge graph

Subsystems, the channels that measure them, the failure modes seen on those channels, and the call each one triggers. The blue link was found from the data alone; the red one is an exclusion rule. Hover a node to trace its connections.

Layer 2 · Rules discovered

Layer 1 · Failure-mode taxonomy

Failures caught in time, by failure mode

Share of the held-out launches with that failure where an alert came in time: before ignition for countdown failures, before booster ignition for engine-start failures, at any point for flight failures. The fixed-limit set has no limit that responds to POGO or to a TVC limit cycle, so it scores zero there by construction.

AgentFixed limits

Per failure mode

Streaming replays of the 500 held-out launches, plus the classifier snapshot at the end of the window (H0 + 150 s). Median delay counts from the moment the simulator marks a failure as visible; a negative value means the agent saw it earlier than that conservative mark.

Knowledge universe

Every star is one action state: the failure-mode probabilities (one per failure mode) the agent held for one launch at one 0.5 s check. States the agent reads alike sit close together. A launch is a chain of states: point at a star to see its launch, click to keep it.

Opening the universe…

How this universe is made