When an Instrument Drifts and When It Just Breaks

Why this matters

An out-of-tolerance calibration result tells you the instrument is wrong. It does not tell you since when, and "since when" is where the whole cost lives. One shop gets a result like that and reviews four jobs. Another shop gets an identical result on an identical instrument and has to look at everything back to the previous calibration, because it has no way to place the failure in time.

The difference between those two mornings is whether the failure was a drift or a break, and whether the shop kept the kind of record that can tell them apart. A drift can be interpolated, which means past readings can often be corrected rather than discarded. A break has no rate, so it cannot be interpolated, and every reading back to the last independent confirmation is suspect at an unknown magnitude.

This card is about the handheld instrument in a technician's hand. A sibling card owns the same fork for a sensor wired into a control system, where the consequences run through the controller rather than through your paperwork.

Why this is a records question before it is an instrument question

A single as-found result is one data point, and one point has no shape. It cannot distinguish an instrument that walked steadily away over a year from one that was fine until it was dropped in a truck bed six weeks ago. Both produce the same line on the certificate.

What distinguishes them is a series: repeated readings of a stable artifact you own, taken at a fixed cadence with the same instrument and the same coupling. The sibling card on traceability owns what a check standard is and why it does not need to be accurate. What belongs here is the payoff, and it is not about accuracy at all: the series is what converts an unbounded review of past work into a bounded one.

What each one looks like in a series

Drift shows as a level that moves in one direction, reading after reading, at a roughly steady rate. It is proportional to elapsed time or accumulated use rather than to any event. Common generic mechanisms are component aging, repeated thermal cycling, mechanical relaxation in an elastic sensing element, and slow degradation of a sensing surface. Drift is the polite failure: it is small for a long time before it is large, and it gives you a rate you can project forward.

A break shows as a step. Readings cluster in a narrow band, then jump, then cluster again somewhere else. It is caused by an event rather than by time: an impact, a drop, moisture ingress, a damaged lead or connector, an overrange excursion, thermal shock. The step size carries no information about when between the two readings it happened.

Two directions matter as much as the shapes. If the instrument reads high, then on a decision of the form "fail if above the limit" you have been condemning equipment that was acceptable, and on a decision of the form "fail if below the limit" you have been passing equipment that was not. If it reads low, both of those invert. The direction on the certificate is what tells you whether the review set is a goodwill problem or a safety one.

The case: two instruments, two logs, two very different mornings

A shop reads a stable check-standard artifact once a month with each of its field instruments, at a working point near 50 units. Instrument tolerance is plus or minus 1.0 units. The values below are illustrative, and the shapes are what transfer.

Instrument A, twelve monthly readings: 50.05, 50.10, 50.15, 50.20, 50.30, 50.35, 50.45, 50.50, 50.60, 50.65, 50.75, 50.85.

The level rose at every single reading, with no month flat and none reversing. Total movement is 0.80 units across the year, which is about 0.07 units a month. At the twelfth reading the error is 0.85 units against a 1.0 unit tolerance, so the instrument is still in tolerance and roughly two months from leaving it at that rate. That projection is the first thing the series buys: the instrument gets booked for calibration now, on a schedule, rather than failing in the field.

Instrument B, ten monthly readings: 50.05, 50.05, 50.10, 50.05, 50.10, 50.05, 50.10, 51.90, 51.95, 51.85.

The first seven readings sit inside a 0.05 unit band with no trend at all. The eighth jumps by 1.80 units off the top of that band, and the ninth and tenth hold at the new level. That is a step, not a slope, and the error at the new level is about 1.85 units against a 1.0 unit tolerance, so the instrument is out of tolerance and comes off the truck the day the eighth reading is taken.

Instrument A: correcting past work instead of discarding it

Because the movement is monotonic and its rate is known, the error at any past date is readable straight off the log. A reading taken in month 9 was taken by an instrument sitting 0.60 units high, and a reading taken in month 6 by one sitting 0.35 units high.

That converts the review into a filter with a stated test: a past call is at risk only where the margin between the reading and the threshold it was compared against was smaller than the instrument's error on that date, in the direction that would flip the outcome.

Work one. A pass-or-fail call made in month 9, where the reading sat 0.4 units on the passing side of the limit. The instrument was 0.60 units high that month, which is larger than the 0.4 unit margin, so the true value was on the other side of the limit and the call flips. That job goes in the review set. A month 9 call with 1.5 units of margin does not flip, because 0.60 cannot cross 1.5, and it stays out.

Two things this does that a blanket review cannot. It shrinks the set to calls that were genuinely close, which in most shops is a small fraction of the work. And where the customer relationship or the contract requires it, the earlier reading can be restated rather than retracted: the reading was 0.60 units high on that date, here is the corrected value, here is the log that establishes the rate. That is a much better conversation than "our instrument was wrong, we do not know by how much."

Two limits on correcting, and both matter. The correction is only as good as the assumption that the drift was steady between readings, which a monthly series supports and a yearly one does not. And a correction applied to a safety decision is not acceptable at all, because the decision has already been made and acted on by a person.

Instrument B: bounding it, because correcting is not available

There is no rate to interpolate. The step happened somewhere between the seventh reading and the eighth, and nothing in the data narrows it further. Every consequential call made in that window was taken on an instrument that may have been 1.85 units off, or may have been fine.

So the review set is defined by time rather than by margin: all consequential calls between the last clean check and the discovery. The check standard is what made that window one month instead of one calibration interval. Without the monthly series, the last independent confirmation would have been the previous calibration, and the shop would be reviewing a year of close calls with no way to prioritize them.

The instrument also does not go back into service after a battery change and a hopeful re-read. A step of that size is an event, the event is usually physical, and the instrument goes for calibration with a note describing the step so the lab knows what it is looking for.

The condition that behaves like a break but does not look like one

A third pattern shows up occasionally and it is worse than either: the level stays put while the spread widens. The mean of the monthly readings barely moves, so nothing looks wrong on a chart of levels, but the scatter around it grows from a few hundredths to a few tenths.

That is an intermittent fault, usually a connector, a lead, or a marginal contact. For review purposes it behaves like a break, because you cannot say which individual past readings were affected. It is caught only by recording every check-standard reading rather than recording whether the check passed, which is one more argument for keeping values rather than verdicts.

Where a drifting instrument is tolerable, and where it is not

A drifting instrument is usable for trending on your own equipment for as long as its error is small compared to the changes you are watching, because a slowly varying offset largely cancels out of a difference. That is a genuine allowance, not a compromise, and it is the reason a shop can plan a calibration a month out rather than parking an instrument the day it starts moving.

It does not extend to safety decisions, and there is no version of this where it does. An instrument used to establish an absence of voltage comes out of service at the first unexplained movement in its check-standard series, not at the projected crossing. The control that catches a break mid-interval on that class of instrument is not the schedule and not the artifact: it is proving the instrument on a known live source before the test and again after it, the live-dead-live sequence at 120.5 of NFPA 70E-2021, in whichever edition your employer's electrical safety program has adopted, inside the electrical lockout duty at 29 CFR 1910.333(b)(2) in general industry or 29 CFR 1926.417 in construction. 29 CFR 1910.334(c)(2) requires the instrument, leads, cables, probes and connectors to be inspected for external defects before each use, which is the check that catches the damaged lead behind most intermittent faults. Where a check-standard reading has to be taken on operating equipment rather than on the bench, 29 CFR 1910.333(a)(1) permits an energized reading only where de-energizing is infeasible, and instrument and probe ratings must meet the circuit under IEC 61010-1, binding through the listing mark.

How to verify you got this right

Look at what your shop records after a check. If the record says "pass" or "verified," you have thrown away the one thing that separates a bounded review from an unbounded one, and the fix costs nothing: write the value.

Then take your longest check-standard series and plot it by eye. A slope means book the calibration before the projected crossing and expect to be correcting past work rather than discarding it. A step means find the event, and treat everything back to the last clean reading as suspect. A widening band with no movement in the level means an intermittent fault, and that instrument is a replacement candidate rather than a calibration candidate, because a fault that comes and goes will pass the lab's bench test on the day it is tested.

References

• 29 CFR 1910.334(c)(2), pre-use visual inspection of test instruments, leads, cables, probes and connectors; 29 CFR 1910.333(a)(1) energized-work gate; 29 CFR 1910.333(b)(2) and 29 CFR 1926.417 electrical lockout
• NFPA 70E-2021, 120.5, absence-of-voltage verification, adopted through an employer electrical safety program or contract; IEC 61010-1 measurement categories, binding through the instrument's listing
• Manufacturer documentation for expected stability, storage conditions and events that void an instrument's calibration
• See related: Why a Drifting Sensor Is Worse Than a Dead One, which owns the control-system sensor case; The Calibration Schedule Worth Keeping; What Traceability Means and Why It Matters to You