What AI can and cannot close in colour
# What AI can and cannot close in colour Two greens sat on the desk. One was the press sheet, fresh, with a bar along the tail that the console had already passed. The other was the buyer's swatch, a drawdown on a different board, slightly glazed from being handled, the green their pack had
What AI can and cannot close in colour
Two greens sat on the desk. One was the press sheet, fresh, with a bar along the tail that the console had already passed. The other was the buyer's swatch, a drawdown on a different board, slightly glazed from being handled, the green their pack had used for years. Beside them, face down, was the contract proof, which the sheet matched closely enough that the software had stopped suggesting moves. The brand manager pointed at the swatch. The minder pointed at the proof. The screen had nothing left to say, because the only comparison it had been given was sheet against proof, and that comparison was quiet. The argument in the room was about which object was the contract. No model in the building could close it.
That's the boundary. Colour software can compare a reading with a target and, on a press that will answer, ask for a move. A closed loop is that loop and nothing vaguer. A measurement of this sheet. A target somebody has agreed. A press, offset or digital, that actually changes the print when you ask it to. Miss one of the three and you don't have a loop. You have a suggestion, a hope, or a graph. Software that guesses a curve from the file, or from jobs it calls similar, hasn't measured this sheet. Guessing can be a decent way to start. It isn't a loop, whatever the brochure calls it.
A loop is three things, and you can point at each
The measurement has to be of the sheet you're judging, taken on a bar or a patch that's clean enough to stand for the picture. Powder, a thumb mark, a bar sitting on a glue flap the buyer will never see, a strip that the imposition clipped: the instrument will still produce a number. The number is about what it saw. If what it saw isn't the job, the loop will close on the wrong object and call itself finished. Somebody has to look at the bar before the reading is allowed to drive a key. That look is short. Skipping it is how a dirty tail becomes a colour correction across the whole form.
The target has to be the thing you sold. A proof on proofing stock, a drawdown on the real board, a signed sheet from the last reprint, a pack from the shelf: these are different objects, and they diverge for ordinary physical reasons. The proofing paper isn't the board. The board's shade isn't last year's. A pack that has sat in a shop has been handled, varnished, and looked at under lamps you don't own. You can store any of them as the aim, if you measure them properly and say which one is in force. You can't store "the green" and expect two people to mean the same surface. When the console and the job bag disagree, the loop will happily close on the console. The buyer will judge the bag.
The press has to answer. On offset that means the ink keys, the ducts and the water will move the sheet in the direction you asked, not merely that a motor twitched. A zone that's already wide open can't give you more density because a screen requested it. A unit running wet will go greyer when you add water to chase a number that was really a damping fault. On a digital press the equivalent is a curve or a channel that actually changes the next sheet, on this substrate, with this ink set. A front end that rewrites a curve in software and leaves the print where it was hasn't closed anything. You've edited a file. The sheet is the test. If the sheet doesn't move, the loop is a diary of wishes.
Put the three on the table when someone offers you a closed loop. Where is the reading taken. Who agreed the target, and is that object in the room. What moves on the press, and how do we see that it moved. If the answer to the first is "the file", you haven't got a measurement. If the answer to the second is "an industry aim we loaded at install", you may have a reasonable starting reference and you don't yet have this buyer's contract. If the answer to the third is "the system handles it", ask for the sheet before and the sheet after. Loops that can't show those two sheets are slogans.
Guessing a curve is a different animal
There's a family of tools that never looks at the sheet and still talks as if it were controlling colour. They take the file, the stock name, perhaps a history of jobs, and they propose a curve, a preset, or a channel adjustment before you've printed anything. Used as a first suggestion, this can shorten the walk to a printable sheet. The coverage preset on an offset press is the old, honest version: area in, key opening out, and then a person looks. The trouble starts when the proposal is described as closed, or as the colour. Nothing has closed. The sheet hasn't spoken.
A guessed curve also mixes jobs that shouldn't be averaged. Same stock code, different batch, different finish, different proofing path. A curve that made last month's brochure sing may open the shadows on this month's carton. The software hasn't failed at arithmetic. It has answered a question nobody should have treated as the final one. Keep the guess upstream of the first pull. After the first pull, only the sheet can tell you whether the guess survived contact with ink, water, substrate and the picture. If you let the guessed curve keep "correcting" without a fresh reading, you're no longer guessing once. You're repeating a bet.
Digital presses collect this confusion because the curve lives in the same box as the raster. It feels like one system, so a change to the curve feels like colour control. It is colour control only when a measured sheet drives the next change, against an agreed target, and the following sheet shows the move. A profile built in a quiet room on a standard patch chart is a necessary piece of housekeeping. It isn't a closed loop around today's job. Today's job might be a thin film, a tinted board, a heavy solid the chart didn't contain. The profile gets you onto the press. The loop, if you have one, starts when you measure the work.
Be wary of the word automatic here. A press that measures, compares and moves while a minder does something else can be a real loop, on a stable reprint, with a bar that's worth reading and a person still glancing the picture. The glance matters because bars and pictures come apart. Close the solid and you can lose the face in the same zone. Chase the face and the logo goes thin. The loop doesn't know which of those the buyer will hate more. A person does, or a signed instruction does. Leave that choice in the software and it will pick the bar, because the bar is what it can see. You'll get a passed bar and a rejected picture, which is the most expensive way to be in tolerance.
The argument the model cannot settle
Back at the desk, the numbers were not the dispute. Sheet against proof was a small remaining difference, the kind a second pull might nibble at and might not. Sheet against swatch was a different green, enough that everyone in the room could see it without the instrument, once the two were overlapped at the edge. The brand manager said the swatch was the pack, and the pack was what shoppers knew. The estimator said the proof had been signed, and the signature was what the plant had priced. Both sentences were about a contract. The model had already done the only thing it can do with colour, which is to report a difference between two things it was shown. It hadn't been shown the swatch. If it had been shown the swatch, it would have reported that difference instead, and it still wouldn't have known which difference the business had agreed to wear.
You can make the model pick, and you can even make the pick look confident. Load the swatch as the aim and the software will chase the swatch, walking the sheet away from the signed proof, perhaps into a place this process can't hold. Load the proof and it will defend the proof against the brand. Either behaviour is obedience, not judgement. Judgement is the conversation that decides which object is in force, and what you'll do if the chosen object isn't printable. Sometimes you reprint towards the swatch and you retire the proof, and you wear the cost because the relationship is worth it. Sometimes you hold the proof, show the swatch beside a drawdown you can actually hit, and tell the brand the old green was a different ink on a different surface. Sometimes the honest answer is that this process won't get there, and the job should move, or the design should. A curve can't make that phone call.
The same unsettled argument turns up in quieter clothes. Process builds that "replace" a spot. The software offers a build, the proof of the build looks close in the booth, and the brand rejects it on the made-up pack because the hue leans, or because a metallic they used to print simply isn't there. You can hand the software the reading of the build and the reading of the spot drawdown. You can let it tell you the gap. You can't hand it the brand's right to refuse a gap that a tolerance table would forgive. Tolerances are a deal. If the deal was never made, the table is a prop. Don't let a passed number shout down a person who never agreed to be judged by it.
Substrate keeps these arguments alive even when everyone is acting in good faith. A proof approved on a bright white stock, a job that runs on a recycled board with its own colour, a varnish that shifts the green once it's dry: the wet sheet, the dry sheet and the proof can be three greens, and a loop closed on the wet bar will swear the job was right at the moment it was wrong for the bag. Measure what you ship, on the stock you ship, after the finishing that changes it, or say out loud that you're approving earlier and wearing the risk. The model will approve whatever moment you point it at. Pointing is the decision. Everything after that is a comparison.
What the floor can hand across
Hand the software a reading from a bar you've looked at, a target you've named, and the before-and-after sheet that shows the press answered. Hand it key positions or a curve version, so the next shift can see what was tried. Don't hand it a dirty bar and expect judgement. Don't hand it three possible targets and expect it to know the contract. Don't hand it a file and expect a loop. The file is the hope. The sheet is the evidence.
If you want the argument at the desk to happen less often, do the dull work upstream. At estimating, write down which object will be the aim, and get the buyer to sign that sentence as well as the proof. On the job bag, put the object, not a nickname. On the console, load that aim and no other. When a swatch turns up on press day and contradicts the file, stop and treat it as a commercial change, because that's what it is. A loop will not save you in that minute. A loop will, on the next reprint, get you back to whatever you were brave enough to name. Name the proof, or name the swatch, or name the signed sheet. Then a measurement, a target and a press that answers can close something real. Leave the name blank, and the cleverest console in the hall will close a circle around a question the room has not agreed to ask.