Your Throughput Ceiling Is a Configuration Setting

A report that says one percent and a floor that says otherwise

A Midwest protein processor sent over its data return ahead of a capacity review. The downtime report for the plant came back at roughly 1 percent. Two people from the project team had walked that same floor five weeks earlier and watched the trim line and the sealer run somewhere between 60 and 79 percent down across two days. Same plant, same month, two numbers that cannot both be true.

The tell was in the issues column. Every row carried the same text. Not similar text, the same text, copied down the sheet like a fill-handle drag. A downtime code that is populated identically on every event is not a record of what happened; it is a record of what the system does when nobody selects a reason.

The labor line had the same problem in a quieter form. Leadership had been running its numbers against a fully burdened rate somewhere north of $30, call it $32. The actual rate in the return was $26.02. Non-work hours came back at roughly 25 percent of paid hours, which nobody in the room could immediately reconcile against the headcount. None of these are exotic data problems. They are the ordinary result of a measurement layer that was configured once and never audited against the floor.

The ceiling is an output, not a property

Plants talk about a throughput ceiling as if it were a physical fact of the building, a property of the equipment the way tensile strength is a property of steel. It is not. The number you plan against is the output of a system: sensors, where they sit, what they are wired into, how the downtime codes are structured, what the timeout threshold is set to, which labor rate the model imports, and which of two conflicting sources the report happens to read.

At the same company, one roll stock line was reporting 100 percent quality. Everyone knew that was wrong. The sensors on that line were new, still being tuned, and the reason they read clean was that they had not been taught what a defect looked like yet. A maintenance tech re-pointed the measurement at the line's PLC instead of the add-on loggers and a real OEE appeared, lower and considerably more useful. Nothing on the line changed. No belt was adjusted, no die swapped, no operator retrained. The ceiling moved because the measurement moved.

This is the general form: system interaction governs throughput, and the measurement layer is part of the system. When a threshold is set so that a stop under a few minutes never registers, the plant's micro-stops become invisible and the line looks healthy at a rate it never sustains. When quality is inferred from a sensor that has not been calibrated to the actual failure mode, first-pass yield reads high while the rework cage fills up. When labor is modeled at an assumed rate rather than the paid rate, every savings case is off by the spread, in this plant's case by about $5 an hour on every hour in the model.

The failure mode is not that the data is wrong. Data is always somewhat wrong. The failure mode is that the wrong data is confident and specific, and confident specific numbers beat contradictory eyewitness accounts in almost every meeting.

Audit the instrument before you argue about the number

The move is to treat the measurement layer as a piece of equipment with its own qualification requirements, and to run that qualification before any expand-or-optimize conversation.

Start with the contradiction hunt. Take one week of reported data per line and put it next to something independent: a PLC counter, a shipped-cases figure, a payroll export, or a supervisor's own shift notes. You are not looking for agreement, you are looking for the size and direction of the gap. In the plant above, the useful discovery was not that downtime was misreported, it was that the reported figure and the observed figure differed by a factor large enough that no amount of analysis on the reported figure could produce a decision worth making.

Then read the codes, not the totals. A downtime report is only as good as the operator's ability and willingness to pick a reason in the moment. Count the distinct reason codes used per line per week. One or two means the crew is picking whatever is first in the list, or the system is defaulting for them. Fix the pick list before you fix the number: fewer choices, floor language instead of engineering language, and the top five actual causes at the top.

Next, settle the source of record per line. If both a PLC tag and a bolt-on logger can produce a rate, one of them wins and it gets written down. Plants running two sources end up in a standing argument that gets resolved by whoever presents first, which is a governance failure dressed as a data failure.

Finally, refresh the assumptions the model imports. Labor rate, non-work hours, and crew size should be pulled from the last closed quarter, not from the number someone quoted in a kickoff meeting. In this case that single correction changed the shape of every labor case on the table before a single improvement was implemented, and it changed it in a direction that made the plant's real opportunity easier to see, not harder.

Only after those four steps is the expand-or-optimize question answerable. A plant that believes it runs at 99 percent uptime will conclude it needs another line. The same plant, measured honestly at 60 percent, has a year of recoverable capacity sitting inside the building and no capital case at all.

What a well-instrumented floor looks like

Reported OEE and a PLC-sourced count agree within a few points, and when they diverge, someone owns closing the gap that week. No line reports 100 percent quality, because no line runs at 100 percent quality. Each line shows at least five distinct downtime reasons in a week, and the top three are named in the language operators actually use. Every model in circulation imports its labor rate from last quarter's payroll, and the number is visible on the sheet rather than buried in a cell. When a capital request arrives, it carries thirty days of measurement that has been reconciled against an independent source, and the request is not scored until it does.

The plant did not have a downtime problem or a labor problem. It had two instruments and no agreement about which one was telling the truth, and it was about to spend capital resolving that argument the expensive way.