Signal Validation Overview
Validation is the gate that decides which samples are trustworthy, cleaning noise, masking dead sensors, and flagging drift before any feature is computed.
Trust before transform
No feature is computed from a raw sample until validation has passed judgment on it. Validation runs per channel and across channels: it rejects out-of-band noise, detects and masks dead or saturated sensors, checks calibration and drift, and cross-validates redundant diagnostics against each other. A wrong sample that reaches the twin is worse than a missing one, because the twin will act on it.
The validation stages
- High-frequency noise filtering to remove pickup and out-of-band energy (see HF filtering).
- Dead-sensor masking to exclude channels that have failed, saturated, or gone silent (see masking).
- Range and rate-of-change checks against physically possible bounds for each machine.
- Cross-channel consistency: redundant magnetics, multi-chord interferometry, and neutron monitors must agree.
A validation verdict per sample
| In-band | Sensor live | Verdict |
|---|---|---|
| 1 | 1 | use |
| 1 | 0 | mask/impute |
| 0 | 1 | filter then use |
| 0 | 0 | reject |
Graceful degradation
When a channel is masked, the fabric does not stall. The sensor-topology graph model imputes the missing value from neighbors (see dead-sensor masking), and the data-quality score attached to the feature records that it was imputed. This keeps the control loop fed while telling downstream models exactly how much to trust the number.
Validation logic is common to the breeder (Hyperion) and the burner; only the physical bounds differ — 9.66 MA and 16.84 T on one machine, the 26.49 T plug and ambipolar potential on the other.