Drift: how instruments change between calibrations

Every reference standard and working instrument drifts. A platinum resistance thermometer's ice-point resistance creeps; a thermocouple's Seebeck coefficient changes with thermal history; a humidity sensor's response shifts with contamination. Drift is why calibration intervals exist.

Putting it in the budget

• Compare the last two or more certificates for the same point; the change over the interval is the observed drift.
• For an instrument used across its interval, a rectangular component of half the expected change is a common treatment.
• If drift is steady and well characterised, a time-dependent correction may be applied with its own uncertainty.

A reference whose drift is not tracked silently moves every calibration it supports; a drift record is part of the traceability evidence.

Getting a number out of the certificates

Drift is estimated from history, which means the estimate is only as good as the file. Line up the reported error at the same calibration point from successive certificates and the pattern is usually visible immediately.

That is about 0.0073 °C per year, monotonic and consistent, which is the well-behaved case. Two options follow. Treat it as an uncertainty: a rectangular component of half-width equal to the expected change over the interval, so 0.0073/√3 ≈ 0.0042 °C for a one-year interval. Or correct for it: fit the trend, apply the predicted correction at the date of use, and carry the uncertainty of the fit instead — smaller, but only defensible while the trend continues to hold.

When the history is not monotonic

Errors that jump around rather than trending are not drift; they are a repeatability or handling problem, and correcting for them would be nonsense. Take the spread of the observed values as the uncertainty rather than the trend, and look for the cause: an instrument that moved 0.02 °C in one direction and back again has probably been dropped, thermally shocked, or calibrated at a different immersion depth.

Instruments with no history

• Use the manufacturer's long-term stability specification as a conservative rectangular limit until you have two certificates of your own.
• Replace it with measured history as soon as you have it. Manufacturer stability figures are typically conservative, and a real record usually shrinks the term.
• Thermocouples are the exception to expecting slow, smooth drift. Their thermal history matters more than elapsed time: a base-metal couple held at high temperature can shift substantially in weeks and then be stable for a year.
• Interim checks — an ice point, or a check standard measured monthly and trended — turn a one-line-per-year record into something that can actually locate when a change began. That matters enormously on the day an out-of-tolerance finding triggers a look-back.

Drift history is also the evidence for the calibration interval. A reference moving 0.007 °C a year against a 0.05 °C tolerance is not being sent away often enough to be interesting; one moving 0.03 °C a year needs a shorter interval, and the certificates are how you argue either case.