Investigate the Model Gap Before Choosing a Fix

When a model and a production record disagree, neither wins by default. The mismatch is a finding that needs an explanation.

A model can be well calibrated in one part of the plant and wrong in another. A production record can also be accurate within a boundary different from the one the model uses.

Name the mismatch precisely

State the line, product, period, metric and size of the difference. Compare units, time windows and definitions before changing any parameters.

An aggregate output match can hide an overprediction on one line and an underprediction on another. Break the comparison down where the intended decision depends on the result.

Keep the original input and model version so the investigation remains traceable.

Test three kinds of explanation

First, check the record. Examine counts, conversion factors, timestamps, equipment coverage and the reference rates used in calculated metrics. A suspicious percentage needs reconciliation, not automatic deletion.

Second, check omitted operating behavior. Shared crews, finite buffers, maintenance, release timing or changeovers may explain why an isolated line cannot reproduce a standalone rate.

Third, check the model itself. Routing, distributions, resource logic and parameter choices may be wrong. A good fit elsewhere does not eliminate that possibility.

Use independent evidence to discriminate among the explanations. Do not tune several parameters until the total matches and call the cause solved.

Validate the decision, not only the baseline

After resolving a discrepancy, test appropriate conditions that were not used to fit the model. Compare the outputs relevant to the proposed investment or staffing change.

Set an acceptable uncertainty based on whether it could change the decision. A universal accuracy percentage can be misleading when the metric, operating range and cost of error differ.

If a critical mismatch remains unexplained, show which conclusions depend on it. The team may still perform useful preparation, but the uncertain scenario should remain conditional.

A model earns influence by making assumptions and errors inspectable. It never earns the right to declare the floor wrong merely because it has matched the floor before.