From Diagnostics to Decision

One trace up the stack: how a raw diagnostic sample becomes a validated feature, a twin state, a prediction, and a bounded action.

Following one signal

The clearest way to understand the architecture is to follow a single measurement from the machine to an action and back. This trace ties the layers, colors, and line semantics together in one path.

The upward trace

• A Thomson-scattering sample crosses a cryo-rated feedthrough into L2 (blue).
• L2 validates, normalizes, and feature-engineers it, recording lineage.
• GNNs fuse it with the constellation and impute any dropped neighbors.
• PINNs update the twin's equilibrium (breeder) or potential (burner) estimate.
• The twin projects the state 50-100 ms ahead; anomaly ensembles score it.

The decision

If the forecast approaches a limit — a disruption boundary on the breeder, a plug-density floor on the burner — the MPC agent (yellow) plans a corrective trajectory over the twin. The trajectory is a proposal, not a command. It carries the twin state and lineage that justify it.

The downward trace

• L4's rules engine checks the trajectory against safety bounds (purple/gray).
• If accepted, L1 executes it deterministically within the command boundary (orange path).
• If the model is uncertain or the bound is violated, the plant falls back to conservative control.
• The outcome is archived; any surprise becomes a pink signal for retraining.

This is the whole architecture in miniature: telemetry rising, authority descending, intelligence proposing, determinism disposing, and the loop closing through feedback. It combines the layer connections with the line semantics and the color grammar.