Your Line Doesn't Have a Rate. It Has a Curve.

The First Phase Is the One That Decides Everything

A precooked sausage producer reached out recently to scope a digital twin pilot. One line, multiple formulations, persistent variance between planned and actual cases per shift. The first conversation didn't get into changes or recommendations. It got into data. Specifically: CAD layout, an NDA, and Ignition extracts at the bagger and the case-pack stations.

Before any model gets built, before any improvement gets recommended, the work is to replicate the variability that already exists. That phase comes first in a four-phase approach: replicate current OEE and variability, run high-volume experiments against the model, integrate live data, then move to prescriptive line adjustments. Each phase is non-trivial. The first one is where most engagements quietly fail, because the question it answers is the one most plants have never seriously asked: how much of our throughput variance is structural, and how much is noise?

What system interaction actually means on the floor

The honest answer in prepared foods is almost always: more structural than the schedule assumes. Cook time changes with fat content. Bagger fill rate changes with link length and casing diameter. Pallet pattern changes with case dimensions. Labor minutes per thousand units changes with how much hand-arrangement the SKU requires before the bagger ever sees it. Each station has its own rate curve over the formulation space. The line's actual rate at any moment is the minimum of those curves at whatever recipe is running.

When that gets averaged into a single number, "the line runs 2,400 cases per shift," you've thrown away the structure. You've turned a vector into a scalar. The schedule is then built against the scalar, and ten weeks later operations is explaining why we keep falling short on the days that tilt toward the slower SKUs.

This failure mode shows up in any prepared-foods plant with a thermal bottleneck. Retorts, pasteurizers, ovens, fryers. The bottleneck sets the pace. The formulation sets the bottleneck's pace. The schedule pretends the bottleneck has a single pace. Three layers of abstraction stacked on top of each other, and only the bottom one is actually controlling the floor.

The same dynamic shows up on labor lines. A sliced-meat producer with strong data infrastructure gets crisp labor optimization models out of a clean source extract. Implementation isn't the bottleneck for them; the model is already saying the right thing. The bottleneck is the planning rhythm that still treats each line as a constant. Until the schedule respects what the model is showing, the gain stays on paper.

Capturing the curve, not the average

In a recent modeling effort at a precooked meat plant, the upgrade that mattered most was capturing variable operator schedules and process-time distributions across SKUs instead of a single mean cycle time. Strict input meant logging conveyor speeds, validating physical container specs early so packaging volume errors didn't cascade into cost and artwork, and running OptQuest against ranges rather than midpoints. The simulation got tighter. More importantly, it got argued with less.

Once the model carried the ranges, the conversation with operations shifted. They stopped pushing back on the simulation. They started pushing back on the schedule that the simulation revealed was unrealistic. That's the move you want. The model isn't there to tell people they're wrong. It's there to make the constraint visible enough that planning can respond to it.

This is also why digital twin pilots that skip the replicate-first phase tend to die. If the model can't recreate last quarter's actual throughput at the SKU level, no operator on the floor will trust its predictions about a redesigned cell. Replication isn't a checkpoint; it's the credibility currency of the entire engagement.

The one-week diagnostic any plant can run

If you suspect formulation is governing your throughput more than your schedule admits, do this in the next five working days.

Pick the line that misses rate most often. Pull six weeks of run data broken out by SKU. For each SKU family, compute median throughput, the 10th percentile, and the 90th percentile. Do not average them. Plot them against the schedule's standard rate.

The gap between the 10th percentile of your slow-SKU mixes and the planned rate is your committed-but-undeliverable volume. It is not a forecast. It is the ceiling that the formulation imposes on certain days regardless of how the crew runs. The gap between the 90th percentile and the planned rate is the volume you are leaving on the table on the days the mix runs in your favor; that is the margin you can promise into when you sequence well.

Bring both numbers to S&OP. Show the curves. The conversation that follows is the one most plants need and almost none have. It is not "how do we run faster on Wednesday." It is "should we be promising Wednesday's volume in the first place, or sequencing the mix differently to flatten the floor."

What this is not

It is not an argument for better forecasting tools. The plants that beat their plan are not the plants with the smartest planning software. They are the plants that schedule against the curve.

It is also not an argument for reducing SKU count. Sometimes that helps; usually the formulations are already there for revenue reasons that operations cannot unilaterally undo. The more useful move is to stop pretending that one line plus four recipes equals one rate. They equal four rates, sometimes five, and the schedule that respects that fact is the one that gets believed.

The capacity is real. So is the variance. Nobody puts the two in the same model. That is the work.