The Spec Change Is Never Where You Made It

The spec change that looked like paperwork

A Midwest protein processor and one of its packers spent a morning on a single document: the film protocol. On paper it reads like procurement housekeeping, which film, which supplier, which qualification tests, who signs off. Underneath it is a throughput decision. The supplier qualification tests on the new film run into early 2027. The protocol had to become the processor's own owned document, not the packer's, because the person who owns the spec owns the consequences when it hits the line. As the packer put it on the call, you make decisions and commitments, so we made a decision.

Here is the part that gets missed. A film change is an input change, and input changes do not stay where you make them. A different film seals at a different temperature, runs at a different speed, tolerates a different amount of line vibration, and scraps at a different rate when the loader hiccups. The procurement team scores the change on cost per thousand feet. The floor pays for it in seal failures, slower indexing, and giveaway three stations down. Both numbers are real. Almost nobody puts them in the same model.

Throughput is a system property, not a line property

The cleanest proof I have seen of this came from a packaged-meats plant running four lines. They sent a full year of operating data. I built a digital twin and calibrated it until the model ran within 2 percent of what the floor actually did, better than 98 percent accurate across the year. That is the threshold that matters: below it you are arguing about the model; above it you are arguing about the plant.

The first thing an accurate model does is expose the lies in the raw data. One line reported a rate of 1.3; the model would only produce 1.2, an 8 percent gap. The instinct in the room is to treat that 8 percent as hidden capacity and go chase it. The model said otherwise. That line had a seven-week stretch where its OEE read above 100 percent, which is physically impossible. The 1.3 was inflated by a measurement error, not real output. The honest number was 1.2. If you had chased the 8 percent with capital, you would have spent money fixing a spreadsheet.

That is the whole mechanism in one number. Throughput is not a property of a line you can read off a single sensor. It is a property of how loading, line balance, labor, and shift structure interact, and a single-step reading of any one of them lies in both directions. The plant's four lines were not four independent throughput numbers. Two were older and inefficient, two were newer and fast, and they were all manually loaded, which capped every one of them at the speed a human could feed it. The constraint was not on any single line. It was in the seam between the loaders and the lines.

Model the whole line, not the step you touched

What you do about it is the same whether the trigger is a film spec, a recipe tweak, or a new SKU. You do not score the change at the step you changed. You run it through a model of the whole system and watch where the constraint moves.

In the four-line plant, the model ran roughly a thousand scenario variations and surfaced four that were worth a human's attention. The winning move was not "speed up the fast line." It was structural: shut the two weakest lines, put automatic loaders on the survivors so throughput is no longer capped by manual feeding, and convert the freed labor into a second shift on fewer, better lines. That is a sourcing-and-staffing decision wearing a throughput costume. It reallocates labor instead of adding it, and it only became visible because the model held the whole system at once instead of optimizing one line in isolation.

The discipline that makes this repeatable is cheap to start. Lightweight wireless trackers go on the lines in a day and give you real throughput and OEE per line instead of a clipboard average. A facility scan gives you the layout the model needs. Half the plants I work with already run a line-tracking system; the other half are flying on shift-end tallies that bury exactly the variance where the constraint lives. The trackers are not the point. The model you build on top of them is the point, because the model is the only place a film change and a labor change can be scored in the same currency before you commit to either.

So when the next spec change lands, do three things this week. Name the owner of the spec, the person who carries the floor consequence, not just the purchase order. Name the downstream station you expect the change to hit, and instrument it before the change goes live so you have a before. And refuse to book a throughput number that came from the changed step alone; make it come from a model that closes within 2 percent of a year of actuals, or treat it as a guess.

What a well-run version looks like

On a floor that runs this way, no capacity number exists without a model behind it, and that model is checked against a full year of real output, not a clean quarter. OEE never reads above 100 percent, because the data is reconciled before anyone makes a decision on it. Every input change, film, formulation, supplier, has a named owner and a named downstream station it was validated against, so a spec change cannot quietly move the constraint and disappear. Disposition between procurement and operations clears in the same meeting, not in a finger-pointing session a month later when the giveaway shows up.

The film protocol that started as housekeeping was actually the cheapest version of this discipline: decide who owns the input, before the input reaches the floor. The savings on the new film were real on paper. So were the seal failures it could cause two stations down. The only mistake left to make is scoring one without the other.