Yield variance is usually a measurement problem before it is a process problem

A yield project usually starts in a meeting where somebody puts a chart on the screen, a line that has drifted downward over four months, and everyone in the room begins explaining it. The operator thinks it is the new supplier's material. Maintenance thinks it is the seal on the second dryer. Planning thinks it is the shorter runs the sales team keeps asking for. All three explanations are plausible. All three arrive with a project attached and a number on the end of it.

What almost nobody asks in that meeting is whether the line on the screen is measuring the same thing in April that it measured in January. In process manufacturing it very often is not, and the drift on the chart is a drift in recording practice wearing the costume of a process loss.

Yield is not observed. Nothing in the plant emits it. It is computed from an input quantity captured somewhere by somebody, an output quantity captured somewhere else by somebody else, and a standard that a person set at some point for some grade. Each of those three has a habit of moving quietly, without anyone filing it as a change, because none of them feels like a change at the time.

The number is a ratio of three decisions

It helps to write out what is actually going into the calculation, because once you do, the fragility is obvious.

• Input quantity What was charged. Which depends on where you weighed it, whether the tare was the drum or the drum plus liner, whether the mid-run top-up got booked to this batch, and whether returned rework counts as an input or arrives for free.
• Output quantity What came out. Which depends on when you weighed it, how much was still in the line, whether the QC sample was deducted, and how the material held back for reprocessing was classified.
• The standard What it should have been. Which depends on who set it, for which grade, on which basis, and how long ago.

None of these is an instrument reading. Each is a convention. Conventions survive on shared memory, and shared memory has staff turnover.

Yield is assembled out of three decisions, and any of them can change without anyone recording that a change occurred.

The pattern, stated plainly

What it usually turns out to be

The weighing point moved

A plant weighs the charge at issue from the store, then someone rationalises the store and weighing moves to the vessel platform. Between those two points sit spillage, residue in the transfer container, and the small quantity that never leaves the bag. The process is unchanged. The recorded input is not. The same thing happens on the output side when the weighing moves from post-packing to pre-packing, or when a bulk bag starts being weighed with its pallet.

Rework is counted twice, or not at all

Recovered material, sweepings, and off-spec product blended back into a later batch are the most common single source of yield that looks better than it is. If the recovered material is charged without being recorded as an input, the receiving batch shows a yield it did not earn, and the batch that produced the recovery already booked the loss. Two batches, one quantity of material, one of them flattered. Over a month this averages out. Over a grade or a shift it does not, which is exactly the cut people look at when they go hunting.

The moisture assumption nobody has revisited

Plenty of process yields are quoted on a dry basis but weighed wet, with the conversion sitting in a standard factor that was set once and never touched. Ambient moisture is seasonal. Supplier moisture varies by lot. The factor does not.

The standard was written for a different product

Standards tend to live in a laminated sheet or a supervisor's head. The grade mix shifts, the plant starts running a finer cut or a tighter specification, and the standard stays where it was. Variance then appears against a benchmark that no longer describes what is being made. The plant is not losing more. It is being compared to something it stopped producing.

Make the measurement boring before improving the process

This is most of the reason Uloborus keeps batch records, material issues, rework and QC results on one set of records rather than three loosely reconciled ones. Not because a single system is tidier to build, but because a yield figure assembled across three systems has three separate opportunities to be built out of parts that do not describe the same batch. When the numbers disagree, the argument becomes about which system is right, and that argument has never once ended in a process improvement.

Where this argument is wrong

If your dryer seal has failed, you do not need a measurement programme. You need the seal. There are losses you can see, smell and sweep up, and treating them as a data question is a way of looking busy while material goes out of the door. The measurement argument applies to variance that moves without a visible cause, not to variance somebody can point at.

It also costs operator time, and that time is real. A plant running one grade in long campaigns with a high value product and a small number of batches per month can reconcile by hand and should. Per batch discipline earns its keep when you have many grades, short runs, rework flowing between batches, and enough volume that a point of yield is worth more than the effort of proving where it went.

There is a cheap test before funding anything. Take one batch from last month at random and reconstruct its yield by hand from the paperwork, without asking the person who made it. If two people doing that independently get the same answer, the process problem is real and worth chasing. If they do not, that is the project.