A sample that fits is not proof: gauge drift, and two complaints with one cause

Two rooms, two arguments, three weeks apart

22 April. Kilverstone's inline inspection at Rangdhanu measures 40 garments. Nineteen of them read 54.6 cm or 54.7 cm on the half chest, against a target of 53.0 and a tolerance of plus or minus 1.5. Twenty-one read 54.2 or 54.3 and pass. The report says "marginal, monitor". From the inspector's chair the results look random.

6 May. The yarn desk reports that the order has drawn 7,011 kg of lambswool against 6,813 kg booked. The over-run is 2.9%, or USD 4,198. The merchandiser starts building a case that the waste factor was set too low.

The quality team is arguing about a measurement. The costing team is arguing about a waste factor. They are in different meetings, on different agendas, with different owners. They are the same defect. The number 2.9% appears on both sides of the building, and nobody puts the two pages together.

Why the two numbers are the same number

A plain knitted fabric, once it has relaxed, is governed by one dimension: the loop length. That is the length of yarn drawn into a single stitch. Courses per centimetre and wales per centimetre both move the opposite way to it. A course is one row of loops across the fabric. A wale is one column of loops down it. This result comes from the middle of the last century, and it is the reason a knitter's control sheet carries a loop length at all.

Two consequences, and they are the whole lesson.

Width follows loop length. The knitting program fixes the number of wales in a panel. Make every loop 2.9% longer and every wale is 2.9% wider, so the panel is 2.9% wider. On a 53.0 cm half chest that is 1.54 cm. The garment lands at 54.54, four hundredths of a centimetre outside a tolerance nobody would call tight.

Weight follows loop length too. The program also fixes the number of stitches in the garment. Yarn used is stitches multiplied by loop length multiplied by the yarn's linear density, which is its weight per unit of length. So a loop 2.9% longer means 2.9% more yarn. 0.3785 kg becomes 0.3895 kg, and 18,000 garments become 7,011 kg.

One cause, two symptoms, in the same proportion, reported by two departments who have no reason to talk to each other.

What actually happened

The size set for KV-2280 was knitted on machine 14 in March, from dye lot A, and approved on fit. A size set is one garment in each size, knitted for approval. It was a good sample, and everybody was right to approve it.

When the loop lengths were finally measured across the floor in May, the picture was this. The bulk machines ran between 3.29 mm and 3.42 mm, averaging 3.355 mm. Machine 14 ran at 3.26 mm, the tightest machine in the building.

It was tight because it had been re-timed the week before the size set was knitted, and its stitch cams were reset. The stitch cams are the parts that set the loop length. The other thirty-five machines had not been checked in eleven months, and cam and needle wear only ever runs one way: looser. The size set was knitted on the one machine on the floor that was in calibration. The approval then certified a garment that no other machine was going to make.

The threshold matters as much as the average. A garment goes outside tolerance above a loop length of 3.352 mm. Nineteen of the thirty-six machines were above it. So a little over half the production was out and a little under half was in. That is exactly what an inspector sampling across cartons sees as "marginal, monitor", because the defect is not spread at random through the lot. It is spread by machine.

That is the assumption every sampling plan makes, and this one breaks it. Course 6.2 gives the plan, and it is the right plan. What it cannot know is that this lot is thirty-six lots wearing one carton mark. If you sample a knitting floor, sample by machine.

Reading the symptoms

The pairing is a diagnosis, and it is worth carrying in your head. It separates four causes with two measurements.

A yarn count sitting 2.5% coarse, inside the spinner's own tolerance, would have produced the over-run and no measurement problem at all. Both symptoms appeared, in the same proportion. That is why the answer is loop length and not the yarn.

What it would have cost to know

Measuring a loop length is not laboratory work. Take the first panel off a machine, unravel a known number of needles' worth from one course, and measure that yarn under a light standard tension. It takes about 12 minutes, including the walk.

Thirty-six machines is 7.2 machine-hours. That is USD 12 of machine time, and half a day of a technician's time, once, at set-up. Against it: USD 4,198 of yarn, an inline inspection nobody could interpret, and roughly half an order of garments that have to be conceded or re-knitted. To concede a garment is to ship it and let the buyer take it at a reduced price.

None of the remedies afterwards is good. It is worth knowing why before somebody suggests one.

• Re-blocking the garments smaller does not work. Blocking moves a dimension. It does not remove yarn. The fabric ends up slacker at the smaller size and grows back, which lesson 5 already showed from the other direction.
• Re-labelling to the next size does not work either. The half-chest grade on this style is 2.5 cm and the error is 1.54 cm. So the garments are not a size larger. They are the same size, wrong.
• What is left is correcting the cams, re-knitting the worst machines' output, and conceding the rest. All of it is priced in days as well as dollars, because the fleet is already the bottleneck.

The approval that would have held

The change is small, and it is entirely in the protocol.

1. Knit the size set across the machines the bulk will run on, not on one machine. Five or six of them, chosen at random, with the machine number recorded on each garment.
2. Knit it from the dye lot the bulk will use, or from the first bulk lot, not from a sample lot. Yarn count varies between lots inside the spinner's tolerance, and it moves the weight.
3. Approve a loop length and a garment weight alongside the fit. The fit approval is the buyer's. The loop length and the weight are the factory's, and they are the two numbers that make the fit repeatable.
4. Record the first-off loop length from every machine at every program change. Stop the machine that is out, rather than measuring the garment three weeks later.

An approved fit sample proves that this garment can be made. It does not prove that it will be. On a floor of thirty-six independent machines, the distance between those two statements is a whole order.

Act as a senior knitwear technologist. Something is wrong with a knitted bulk and I need a diagnosis before I answer the buyer. Facts: style [STYLE], gauge [GAUGE], yarn [SPEC], quantity [QTY], the measurement point with its target and tolerance [SPEC], and what the readings actually are [NUMBERS]. Yarn booked [KG] and yarn drawn so far [KG]. Number of knitting machines running the order [NUMBER]. The size set was knitted on [WHICH MACHINES] from [WHICH DYE LOT] and approved on [DATE]. Loop length recorded at set-up, if any [MM PER MACHINE]. Blocking board dimension [CM]. The state at which the factory measures [OFF MACHINE, RELAXED, BLOCKED, AFTER WASH]. Do the following. First, tell me at which state my measurements were taken, and what that state can and cannot prove. Second, put my two symptoms side by side: did the measurement move, did the weight move, and in what proportion. Give me the single cause that explains both, or say clearly that there are two causes. Use this key. Weight over with the measurement on spec points at the yarn count. Both over in the same proportion points at loop length. The measurement over with the weight on spec points at stretching rather than knitting. Both over with courses and wales per centimetre unchanged points at the program. Third, tell me what to measure next, in order, and how long each measurement takes. Fourth, work out whether the defect is spread through the lot or concentrated on particular machines, and say what that does to my inspection sampling. Fifth, price the options: concession, re-knit, re-block and re-label. Say which of them physically cannot work, and why. Sixth, write the control that would have caught it, as a one-page instruction for the knitting floor.