In line quality and statistical process control

Cue 1 A run of results building into a control chart, centre line and limits appearing

AMARA 00:00 Start with the thing people get wrong on day one.

NADIA 00:04 That variation is a fault. It is not. Every process varies, and the only useful question is which of two kinds you are looking at, because the two have completely different owners.

AMARA 00:16 Name them.

NADIA 00:17 Common cause is the process being itself. Ambient temperature, material lot to material lot, the ordinary play in a mechanism. Always present, predictable within a range. Special cause is something new: a worn tool, a failed heater, a different operator, a delivery from a different supplier.

AMARA 00:36 Why does the distinction matter so much?

NADIA 00:38 Because it decides who acts. Special cause belongs to the person at the machine, now, and it can be found. Common cause belongs to management, because reducing it means changing the process itself. Asking an operator to fix common cause variation is asking them to do something the job does not permit.

Cue 2 The same chart with a run of seven points on one side highlighted

AMARA 00:59 And the chart tells them apart.

NADIA 01:01 That is its entire purpose. Results in time order, a centre line at the average, limits at three standard deviations calculated from the process while it was behaving.

AMARA 01:13 Why three?

NADIA 01:13 It is a compromise, not a law. Tighter and the chart cries wolf and people hunt causes that are not there. Wider and real signals slip past. Three has held for a century because it balances those two costs about as well as anything does.

AMARA 01:31 Most people only know one rule.

NADIA 01:34 A point outside a limit, and on its own that catches large shifts only. A process can drift a very long way with every point still inside the lines.

Cue 3 A trend developing towards a limit, with the intervention point marked early

AMARA 01:45 So what else are you looking for?

NADIA 01:48 Runs and trends. Seven consecutive points on one side of the centre line means the average has moved. Seven rising or falling means something is drifting progressively, and that is what tool wear looks like before it produces a single reject.

AMARA 02:05 That is the useful one.

NADIA 02:07 It is the whole prize. A trend caught at point five is a tool change scheduled at the end of the shift. The same trend caught when a point finally breaches the limit is a quarantined batch and a conversation with a customer.

AMARA 02:24 You also watch for points that are too neat.

NADIA 02:27 Data hugging the centre line unnaturally closely. Real processes are not that tidy. It usually means somebody is averaging, rounding, or writing down what they expect to see rather than what the gauge said.

Cue 4 Two charts side by side: in control and out of specification, then the reverse

AMARA 02:41 Control limits and specification limits. Make the distinction hard to forget.

NADIA 02:45 Control limits come from the process and describe what it does. Specification limits come from the customer and describe what is acceptable. They are unrelated quantities that happen to be drawn in the same units.

AMARA 02:59 What goes wrong when they are confused?

NADIA 03:02 A process can be beautifully in control and outside specification on every unit. Stably producing rubbish. It can also be well inside specification and completely out of control, which means it is getting away with it today and you have no idea what it does tomorrow.

AMARA 03:21 And on the chart itself?

NADIA 03:23 If the tolerance is drawn on a control chart, that chart is training people to tamper. Every time a point wanders towards a line it was never meant to be compared with, somebody reaches for an adjustment.

Cue 5 A tampering simulation: an operator correcting noise and the spread widening

AMARA 03:38 Define tampering.

NADIA 03:38 Adjusting in response to common cause variation. The part reads slightly high, the operator corrects, the next reads slightly low, they correct again. The output becomes measurably more variable than if nobody had touched the machine at all.

AMARA 03:54 That is counterintuitive enough that people will resist it.

NADIA 03:57 The arithmetic is simple. The natural swing is still there, and now your correction is added on top of it. Two sources of variation where there was one. It is not a failure of skill, it is what happens when you respond to noise.

AMARA 04:15 How do you stop it without telling people to care less?

NADIA 04:19 You never tell them to care less. You give them a rule that says when a reading is a signal and when it is not. That is precisely what the chart is for, and it is why an unread chart on a wall is worse than no chart, because it looks like control.

AMARA 04:40 Sampling. How much is enough?

NADIA 04:42 Take consecutive units so the sample reflects one moment of the process, and space the samples through the shift so the gaps between them reveal drift. A sample assembled from units picked at random across twelve hours blends the two and can hide a shift that happened at lunchtime.

AMARA 05:02 And frequency?

NADIA 05:03 Follow the rate of change. A process that drifts with tool wear needs sampling often enough to see the drift while correcting it still costs a tool change rather than a batch.

Cue 1 A distribution moving inside a fixed tolerance, with Cp constant and Cpk falling

AMARA 05:16 A process is in control. Is it a good process?

NADIA 05:20 Unknown, and the two questions are genuinely separate. In control means stable and predictable. It says nothing at all about whether what comes out is acceptable. A process can be beautifully stable and consistently outside specification.

AMARA 05:34 So capability is the second question.

NADIA 05:36 And it only means anything once the first is answered. A capability figure calculated on an unstable process describes a fortnight that will not happen again.

AMARA 05:47 Cp and Cpk. Separate them.

NADIA 05:49 Cp asks whether the process would fit inside the tolerance if it were perfectly centred. Specification width over process spread. It takes no account of where the process actually sits. Cpk asks whether it fits where it really is: distance from the average to the nearer limit, over three standard deviations.

Cue 2 Four capability cases side by side against the diagnosis table

AMARA 06:09 Why carry both?

NADIA 06:10 Because the gap between them is a diagnosis. If both are low the process is too variable and centring will not rescue it. If Cp is healthy and Cpk is much lower, the spread is fine and the process is simply off centre.

AMARA 06:28 Which is the good news.

NADIA 06:30 Much the better news. Off centre is frequently a setting and sometimes fixed the same morning. Too variable is a project. Two numbers tell you which conversation you are about to have, and that is why reporting only one of them wastes the method.

AMARA 06:47 What target should people work to?

NADIA 06:50 The convention is 1.33 for an established process and 1.67 where safety or a regulator is involved. Confirm what your customer actually specifies. Assuming the convention when the contract says something else is an expensive way to find out.

Cue 3 A line diagram with gates marked and a defect escaping past each in turn

AMARA 07:05 First pass yield. Why insist on it when scrap is already counted?

NADIA 07:10 Because rework is not scrap and it is not free. It uses existing people, existing space, existing machines, and it produces a unit that eventually sells. So it appears in no report while consuming real capacity.

AMARA 07:24 The hidden factory.

NADIA 07:26 A second plant running inside the first. You can walk into it. There are people in it. It is on no organisation chart, and the first time a plant measures first pass yield the number is usually uncomfortable enough that somebody questions the measurement rather than the result.

AMARA 07:45 Quality gates. What makes one real?

NADIA 07:47 Whether it can stop something. A gate that records a result and lets the product continue is a measurement point with an intimidating name. The authority to halt is the entire content of the idea.

Cue 4 A held batch moving through release, rework and scrap paths

AMARA 08:01 Where do you put them?

NADIA 08:03 Where the cost of continuing jumps. Before the constraint, because your slowest station must never spend its time on something already scrap. Before anything irreversible: sealing, curing, potting. Before packing, and before despatch.

AMARA 08:16 And the moment the lorry is waiting.

NADIA 08:19 That is the moment the gate either exists or does not. Which is why the authority to override cannot sit with the person who carries the despatch number. Not because they are dishonest, but because you should not build a control that depends on somebody resisting their own objective.

AMARA 08:39 A batch is held. Walk me through the outcomes.

NADIA 08:42 Three. Release, when investigation shows it conforms and you record the reasoning. Rework, by a defined and approved method, then reinspection against the original criteria. Scrap, when it cannot be brought into conformity.

Cue 5 A traceability tree from a finished unit back to material lots, and forward again

AMARA 08:56 Against the original criteria. That will be argued.

NADIA 08:59 Constantly, and it is the line that must not move. Reinspecting reworked product against a relaxed standard is how a specification quietly becomes a suggestion. If a reworked batch fails again and is held again, that is the system working, not the system being awkward.

AMARA 09:17 What about use as is?

NADIA 09:19 It exists, and it is not a decision a plant may take on its own. It requires the customer or the design authority. A plant that grants itself concessions has stopped having a specification and has not told anybody.

AMARA 09:34 Who signs the scrap?

NADIA 09:36 Not the person carrying the output number. A common arrangement is that quality may release or rework any quantity, and scrap above a defined value needs a second, more senior authorisation. The threshold is arguable. The separation is not.

AMARA 09:52 Last one. Traceability.

NADIA 09:53 It sets the size of your worst day. With batch traceability a bad material lot means recalling the units that contain it. Without it, you recall everything made between two dates you are guessing at.

AMARA 10:07 How do you know yours works?

NADIA 10:09 Test it before you need it. Take a finished unit at random and time how long it takes to name every input. Then take an input lot and time how long it takes to list every unit containing it. Most plants have never run the second test, and it is the one a recall actually asks.