Model-in-the-Loop Engineering · Capstone brief

ENG-08 · Calibration drift

After a teacher-defined distribution shift (seasonal temperature, wear, or load change), can you detect that the model’s errors drift — and say whether recalibration or human takeover is the right response?

The question

After a teacher-defined distribution shift (seasonal temperature, wear, or load change), can you detect that the model’s errors drift — and say whether recalibration or human takeover is the right response?

System / materials

Logged before/after segments from an approved plant or public sensor stream. Shift must be documented, not cherry-picked.

Expected failure modes

Calling noise “drift.” Retraining on the test period.

Done looks like

Drift chart, threshold for alert, and a written runbook line: recalibrate / degrade gracefully / human only.

Five C's

CT: drift vs. noise. CR: choosing a detection statistic. CO: peer verifies shift dates. CM: runbook one screen. CZ: who acts on a false calm.

Mentor role

Quality or reliability engineer reviews shift definition at Checkpoint 1 and runbook at Checkpoint 2. School-supervised.

Rubric calibration

R1: shift documented. R2: segments reproducible. R3: pre-shift baseline. R4: false alarm discussed. R5: runbook actionable. R6: refuses silent continue.

Two ways this goes wrong

(a) Retrain on everything — looks fixed, no detection story. (b) Drift claimed from three data points.

Credit lane fit

Lane A immediately. No verified credit claim.