Your Line Rate Is an Average, and Averages Lie

A Recipe Change That Was Actually a Capacity Decision

A Tier-1 protein processor running a national quick-service program approved capital to move a steak SKU to a pre-marinated process. The recipe logic was sound. The process logic was not free. Based on what the same plant had already lived through on a carne asada line, the team's own estimate for the new process was a 12 to 15 percent line run-rate reduction: added tumbling and adobo steps, heavier packaging, more handling, more complexity per pound.

The volume targets did not move. Seasonal demand did not get smaller to match. So a decision that looked like a recipe change on the capital sheet was, on the floor, a capacity decision nobody had priced as one. Hold volume constant, take 12 to 15 percent off the run rate, and you have manufactured a shortfall before a single piece of equipment has failed. That is the shape of the variability tax. It does not show up as a breakdown. It shows up as a line that quietly stops being able to do what the plan assumes it can.

The Line Rate Is a Fiction You Agreed To

Every capacity plan rests on a rated speed: units per hour, pounds per shift, one clean number per format. That number is an average, and the line does not run on averages. Real output is governed by interactions. The added process steps interact with conveyor timing. Changeover patterns interact with the schedule. Staffing interacts with the slowest constrained station. Demand mix interacts with all of it. Each interaction is small. Together they decide your actual throughput, and almost none of them are visible in the one number on the capacity sheet.

I saw the cost of ignoring this directly on another engagement. We built a structural simulation of a multi-plant meat processor's facility straight from its CAD files and identified two lines as the high-leverage points for optimization. The model was clean. It was also, in our own words to the client, directional and incomplete, because it lacked the operational variability data: downtime, changeover patterns, staffing reality, run-rate fluctuation. A model built on geometry and averages will confidently tell you the line is fine. The variance is exactly where the output leaks, and a model that does not carry the variance is just a prettier version of the spreadsheet that got you here. That is why the same readout flagged camera-based trim-loss and downtime tracking as the next move. You cannot optimize a number you are not measuring.

How to Find the Tax Before You Pay It

The fix is not a better average. It is capturing the variance and then refusing to trust anything the average told you.

On a packaged-meats line we were modeling, the client offered 30 days of data. We pushed for a full year. Plants are not static, and the entire value of the exercise is finding the variation: the seasonality, the bad weeks, the OEE swings that a 30-day window flatters into a smooth line. We built the digital twin to within 2 percent of the real baseline, 98.2 percent accuracy, good enough to test scenarios virtually instead of stealing plant time to try them live. But one line came in at only 88 to 89 percent. The reason was not the model. It was roughly seven weeks of OEE data with spikes that did not make sense. The variance was trying to tell us something, and the honest move was to chase the bad data rather than smooth it away to make the twin look better.

That is the discipline. Three concrete moves any plant can run this quarter:

First, lengthen your window. If your capacity case is built on a month of data, it is built on an average that has never met your worst week. Pull twelve months and look at the spread, not the mean.

Second, model before you buy. The reflex when a line misses rate is to add capital or add bodies. On the protein line above, the better lever was timing the center conveyor to deliver one item to the cutters at a consistent interval, roughly every 42 seconds, instead of overfeeding and creating the build-ups that looked like a staffing problem. Pacing the system beat adding to it. You only find that by simulating the interaction, not by staring at the rated speed.

Third, put the savings and the throughput in the same model. The recipe change above had an approved capital case. The 12 to 15 percent run-rate hit was acknowledged in a different part of the same conversation. Those two numbers belong in one model, because they are one system.

The Two Numbers Nobody Put Together

The capital case for the format change was approved on the recipe. The capacity hit was estimated, honestly, a few minutes later, and then set aside as a thing to revisit after the rollout stabilized. The savings were real. So was the lost output. They lived in the same plant, on the same line, in the same week, and nobody put them in the same model. Price the variance yourself, or the line will price it for you, one quiet shift at a time.