Diagnostic Calibration Drift and Recalibration

Sensors drift, and a drifting diagnostic silently corrupts every model that reads it, so Kronos tracks and models sensor calibration as part of the MLOps loop.

When the instrument, not the plasma, moves

Models read the machine through diagnostics, and diagnostics drift: gains change, offsets creep, detectors age. A drifting sensor is uniquely insidious because it makes a stable plasma look like it is changing, corrupting every model downstream while the plasma itself is fine. Kronos treats diagnostic calibration drift as a distinct monitored quantity, separate from plasma-regime drift.

Calibration drift is detected by cross-checking redundant and physically-related diagnostics against each other and against the physics twin. If two independent measurements of the same quantity diverge over time, or a diagnostic diverges from a well-grounded twin prediction while its neighbors do not, the diagnostic — not the plasma — is the likely culprit. A learned calibration model estimates and corrects the drift.

Detection and correction

• Cross-diagnostic consistency: redundant sensors should agree
• Twin residuals isolated to one channel, not the whole state
• Drift trend on calibration reference shots
• Learned calibration correction with uncertainty
• Flag affected historical data for relabeling via lineage

def calib_drift(channel, redundant, twin, hist):
    d_cross = disagreement(channel, redundant)
    d_twin  = residual(channel, twin.expected)
    if d_cross > TH.x and d_twin > TH.t and neighbors_ok(channel):
        corr = fit_calibration(hist[channel])   # gain/offset over time
        return corr, affected_datasets(channel)  # -> recurate + retrain
    return None, []

Confirmed calibration drift has two consequences: online, it corrects the channel before it feeds models; offline, it flags every historical dataset that read the miscalibrated channel for recuration and dependent-model retraining, discoverable through lineage. This is why lineage tracks calibration revision at read time — so a drift discovered in 2032 can trace back to exactly which 2030 data it contaminated.