Your scrap number is wrong. Your rework number is worse.

Walk any plant and ask two questions. What was your scrap rate last month? What was your rework rate last month? You will get a confident answer to the first and a shrug to the second. That gap is the whole problem.

Scrap gets reported because somebody has to throw the part in a bin and the bin has to get weighed. Rework gets absorbed. An operator fixes a short shot on the next cycle, a welder grinds down a bead and runs it again, a coil gets re-tempered. None of that shows up unless somebody writes it down, and writing it down is not the job.

Where the data actually lives

In most of the plants we walk into, quality data lives in four places: a paper traveller that follows the job, a shift logbook, an Excel file somebody emails on Friday, and the head of the operator who has run that press for eleven years. The MES, if there is one, usually captures a defect code and a quantity. It does not capture the barrel temperature at the time, the regrind percentage in the hopper, or the fact that the mold was pulled for cleaning two hours earlier.

So when you go to attribute first-pass yield loss to a specific process parameter, the join does not exist. You have a defect. You have a timestamp. You have a historian full of setpoints. Nobody has ever tied them together at the shot or coil or heat level.

What it takes to make it analyzable

iFactory's scrap and rework analytics checklist is a decent starting inventory: defect codes, operator and shift attribution, machine ID, process parameters at the time of the event, and a cost tag per defect. If you cannot answer all five for a given scrap event, you cannot do attribution. You can only do accounting.

The practical work is less glamorous than it sounds. It is standardizing a defect code list so "flash" and "flashing" and "excess material" stop being three different things. It is getting the historian tag names to match what the quality team actually calls the parameter. It is deciding whether a reworked part counts against first-pass yield (it should) and getting the shift supervisors to agree. EFESO's scrap reduction case describes the same pattern from the consulting side: most of the project is reconciling scattered machine and process logs before any model gets built.

The boring win

Once the join exists, the analysis is often unsurprising. One shift runs hotter. One mold produces 60% of the sink marks. One alloy heat has a rework rate three times the others and nobody noticed because the parts shipped. The value is not in the algorithm. It is in finally being able to ask the question and get the same answer twice.

Start there. The model can wait.