Seven Weeks of Impossible Numbers

A Year of Data, and Seven Weeks of It Were Impossible

A sliced-meats processor handed over 30 days of production data first. We asked for a year. The reason was specific: a plant is not static, and the value is in the variation, the seasonality, the swing between the weeks that run clean and the weeks that fall apart. A month cannot show you that. A year can.

The year was good data, some of the best we had seen. But as we built the digital twin, one line told an impossible story. Roughly seven weeks of OEE reading above 100%. There is no such thing. It was a structural data error, an intermittent issue the plant had actually struggled with for weeks without resolving. We had a choice: normalize it, quietly bend the model to match it, or call it out. We modeled to about 98% accuracy and showed the gap on that line rather than chase a number that could not be real.

The Tax You Cannot See Yet

Every plant pays a variability tax. It is the throughput you lose to the gap between your good weeks and your bad weeks: the staffing you carry for a peak that rarely comes, the promises you miss when the line underruns, the schedule built for an average day that never actually happens. Run to the average and you are planning to a plant that does not exist.

But here is the part that gets skipped. You cannot even measure the tax until your data can be trusted, and operational data lies constantly. Sensors double-count. Manual entries inflate. A line reads above nameplate. If you build a plan on data you never validated, you are not just paying the variability tax. You are paying it blind, because the weeks where the tax is actually levied are buried inside numbers that are physically impossible and were never flagged.

Validate Before You Average

Before any optimization, ask for a year, not a month, so the variance is visible at all. Then validate before you model. Any week that reports a physically impossible number, OEE over 100%, run rates above nameplate, a yield over theoretical, gets quarantined, not averaged in. Model to the clean baseline and show a gap where the data is bad, instead of bending the model to match a lie. A model that is 98% accurate and honest about the other 2% is worth more than one that is 100% accurate because it matched seven weeks of broken sensors.

Then size the plan to the variance, not the mean. The bad weeks are where the tax is paid, so the plan that reduces it has to be built against the low-decile week, not the comfortable average. That is the difference between a schedule that survives a bad week and one that only works on paper.

What a Well-Run Floor Looks Like

A well-run floor validates before it averages. No week with a physically impossible number survives in the baseline. Plans are sized to the low-decile week, not the mean. The spread between the best and worst week is tracked as a single number and shrinks quarter over quarter, and when the data is bad the report shows a gap rather than a fabricated match.

The model was 98% accurate because the team refused to make it 100%. The seven impossible weeks were the most honest data in the whole set; they told us exactly where the reporting could not be trusted.