Your ΔE Reading Is Only as Good as the Standard Behind It

← Back to Insights

By Daniel Wang · September 2026 · 9 min read

The report said ΔE 0.27.

It was flagged as passing. On a specification with a 1.0 limit, that is not a close call — it is a number so far inside the line that nobody thinks to question it.

The part was wrong. Not marginally wrong, or wrong under a light booth if you look carefully. Four grey scale grades off, in the same room, under a D65 lamp.

I have written elsewhere about why the eye and the number disagree — how textured materials compress a reading, and why a grade should decide where a decimal point should not. That argument assumes the reading is worth listening to in the first place.

This piece is about the case where it is not. It is a more dangerous case, because everything in the output looks correct.

A ΔE is a relative number, and it knows exactly one thing

ΔE measures the distance between two coordinates in a colour space. One of those coordinates is your sample. The other is whatever the instrument has stored as its reference.

That is the whole of it. The instrument does not know which panel your customer signed off. It does not know which lot the reference came from, or how long ago it was set, or who had access to the menu since. It knows the standard in its memory, and it answers against that, faithfully.

Which means the number is only as meaningful as the reference behind it. A perfectly calibrated instrument pointed at a reference that has drifted will report a small ΔE with total confidence — and be entirely correct about the wrong question.

This is not a device failure. It is a record-keeping failure, and it is invisible in the output, because the output looks identical either way.

What our own meter’s log turned out to show

Three readings from our bench, across two days:

Look at the third column. Those are the references the instrument was holding, back-calculated from its own screen — the sample L* it reported, minus the ΔL* it reported. All three were stored under the same identifier: S00001.

That is a 44-point spread in lightness, from a near-white reference to a mid-dark one, filed under one name. Against that, ΔE 0.27 means nothing at all. It means the sample sat 0.27 away from the thing in the meter — and the thing in the meter was not the panel anyone had agreed to.

Colour systems do not usually rot through a broken instrument. They rot through a reference that nobody owns — and the output looks fine the whole way down.

A reference is either under control, or it is a number in a machine

“Under control” has a specific meaning in a quality system, and it is stricter than “we have a reference.” A controlled standard is one where you can answer, at any moment, without asking anyone:

• What is it? A physical panel, signed and dated by both parties. Not a code, not a screenshot.
• Who set it? A named person, and a named authority behind them. “The factory” is not an answer.
• When? A date on the panel, and a date in the record.
• Who can change it? If anyone with access to the meter can re-standardise it, it is not controlled.
• Does it still match the approved piece? Verified on a schedule, not on suspicion.

If any of those answers is “we would have to ask someone,” you do not have a standard. You have a number in a machine.

The single discipline that matters most: a reference that is re-set without a record is worse than no reference at all, because it produces confident answers that cannot be traced afterwards. Every re-standardisation should leave one line — date, who did it, what it was set against. On our own bench that line was not always kept, which is why the table above had to be reconstructed from screen photographs rather than pulled from a log.

SCI and SCE are two different answers, and there is usually no record of which one you got

Both meters in these photographs were reading in SCI mode. That matters more than most people allow for.

SCI — specular component included — measures the light coming off a surface including the mirror-like reflection. SCE — specular component excluded — takes that reflection out and measures only the diffuse light. On a matte reference card the two land close together. On a moulded plastic part, a brushed metal component, or any film with sheen, they do not.

And the trap runs the way you would not expect: on a glossy part, SCI is the more forgiving of the two. The specular reflection washes out precisely the differences a person sees when they tilt the part in their hand.

So a report can carry a respectable ΔE, in a stated mode, on a part whose surface finish is visibly wrong. Both statements are true, and neither is the answer.

How to separate them in one step: read the same part twice, once in each mode. If SCI comes back small and SCE comes back large, you do not have a colour problem — you have a surface-finish problem, and the next instrument to reach for is a gloss meter at 60°, not a colour meter.

The mode belongs in the report. A ΔE with no stated geometry is not a measurement. It is a screenshot.

The measuring point is part of the claim

A handheld meter reads a patch roughly 8mm across, at one angle, in one place.

On a moulded part, the places that go wrong are the edges, the gate, the witness lines and the flow marks — and those are exactly where an operator does not put the aperture, because the aperture sits more comfortably in the middle of a flat face.

A single reading in the middle of a part is a statement about the middle of that part. It is not a statement about the part. Most rejected shipments live in the gap between those two sentences.

What we do instead: a fixed set of measuring points, agreed with the factory and marked on the drawing, each one recorded with its reading. It is not more sophisticated than a single measurement. It is the same measurement, several times, in positions that were chosen before anyone knew what they would show.

Below a certain difference, the number stops carrying information

Every instrument has a repeatability — the spread you get measuring the same thing twice. For handheld colour meters in factory use, the practical figure is around 0.1 ΔE, and agreement between two units of the same model is often worse than that.

Which gives a working rule: at ΔE below roughly 0.3, the instrument is inside its own uncertainty. A reading there is not evidence that the colours match. It is evidence that the instrument could not separate them at that spot, on that day.

I want to be careful with that number. It is not a published limit, and it is not identical across devices. It is an order of magnitude — and it is why a lab that knows what it is doing will not set an accept line at ΔE 0.5 and then celebrate a result of 0.2.

A very small ΔE is not a strong result. It is frequently a weak measurement of a small question.

What has to be written down before a number becomes a record

This is the part that decides whether any of the above survives contact with a container arriving six months later and everyone looking for someone to blame.

A colour reading is a record when it carries, at minimum:

Six fields. One line of a spreadsheet per reading — and the whole difference between a measurement and an opinion with decimal places.

And then there is the question of who decides

Everything above is about making the number valid. None of it is about making it the verdict.

The order that holds up:

1. The written standard decides what is being asked. Which reference, which illuminant, which viewing condition — agreed before production, not after the argument starts.
2. The eye decides the grade. Under the agreed light, against a certified grey scale, by someone who has done it enough times to be consistent.
3. The instrument documents and screens. It tracks drift across lots, catches a supplier going out of control over months, and puts a number in the file.

When two and three disagree, two wins — that is the argument in the other piece, and it does not change here. What changes here is the third case: when three is invalid — no controlled reference, no stated mode, no named measuring point — it does not get a vote at all, however precisely it is printed.

What this costs, honestly

Controlling a reference is administrative work. It is a log, a signature, and a physical panel in a drawer that somebody is responsible for. It produces nothing a customer ever sees.

Measuring points add a few minutes per part. Dual-mode readings double the readings. Keeping a record per measurement is a habit that decays the moment a factory gets busy — ours included.

The return is narrow and it is real. When the argument comes, you can put a piece of material on a bench with a page of records beside it, and nobody has to be believed.

A ΔE that cannot say what it was measured against is not a smaller version of a useful number. It is a different kind of thing.

What to ask your supplier

1. “What reference did this reading use, and when was it set?” If the answer is not a date and a name, the reading is not evidence.
2. “SCI or SCE, and which illuminant?” A report with no stated mode cannot be compared to anything, including itself next month.
3. “Where was the aperture placed?” If the answer is “in the middle,” ask what the edges read.
4. “What is the acceptance criterion, and who signs it off?” The answer tells you whether colour has an owner at that factory, or just a meter.
5. “Can I see the standard log?” One page. It will tell you more than the last fifty colour reports.

Colour is one line in a longer spec sheet — the full component list is in The Luggage Spec Library. The grading method itself, with the viewing conditions and the accept line, is in Why ΔE Under 2 Can Still Look Wrong.

Arguing about a colour right now? Send us the details — we read them all.