Lean for Beginners: Cannabis · Lesson 8 of 10 · Week of November 30

Quality at the source

Most operations catch mistakes by checking the work afterwards: a count at the end, a second look at the label, a test before sale. Checking has a place, and in cannabis some of it is required by law. But a check finds a mistake after it has been made, and sometimes after it has been passed down the line. This lesson is about the other approach: designing the mistake out where it starts.

By the end of this lesson you can:

tell inspecting after the work from catching the problem at the source;
say what jidoka and an andon are, and what they are not;
explain poka-yoke, and the difference between a device that controls and one that only warns;
work out first-pass yield and rolled throughput yield, and spot the rework hiding behind a good final number;
name the rules that decide what you may do when a test fails or a label is wrong.

You need Lessons 1 to 7 first. This one builds on the required-step rule from Lesson 1 and the standard work from Lesson 6.

Build quality in, don't inspect it in

W. Edwards Deming's third point in his 14 points was to cease dependence on inspection to achieve quality. The idea is that if you rely on checking, you are paying for the mistake and then paying again to find it. He allowed 100% inspection where safety is at stake.

Shigeo Shingo described three kinds of inspection, and the differences are worth knowing:

Judgement inspection sorts good from bad after the work is done. It doesn't change what happens upstream.
Informative inspection uses the results to correct the process, for example by tracking results over time or by checking the step before.
Source inspection checks the conditions before the work, so the mistake doesn't happen. It works because it gives feedback at the moment of the error, and because a simple device can check every unit instead of a sample.

Deming and Shingo look as if they disagree, because Deming warns against mass inspection and Shingo wants 100% inspection. They fit together once you separate sorting from preventing. Shingo's 100% is at the source and done by a device. Deming's warning is about sorting after the fact.

Testing isn't waste. In Massachusetts, laboratory testing before sale is the law (935 CMR 500.160). Don't try to remove it. And a passed test is a check on a sample, not proof that every unit is good. In February 2025 the Commission issued a public health advisory about potentially contaminated products sold between September 2024 and January 2025. By the Commission's account, a flower lot that had passed testing was later found over the limits for yeast and mold. The lesson isn't that tests fail. It's that quality has to be built in upstream of any test.

Jidoka and the andon

Jidoka is a Toyota word often translated as "automation with a human touch." It means giving a machine or a person the power to notice that something is wrong and stop, so the problem isn't passed on and the cause gets fixed. It has three parts: detect the abnormality, stop, and fix the cause.

Two things it is not. It isn't "automation," and it isn't "no people." The person is the point: the machine or the operator finds it, and a person fixes the cause.

An andon is the visual signal that shows where a problem is, such as a lamp, a board or a set of lights, set off by a sensor or by the operator pulling a cord or pressing a button. At Toyota, operators can pull a cord to stop the line. A pulled andon doesn't always halt the line instantly. The line usually stops at a fixed position, so the person has time to respond.

An example, with made-up details. The potency number printed on a label is wrong halfway through a run. Under jidoka, the person who sees it stops the run, signals, contains the suspect units, and then looks for the cause. They don't keep going and sort at the end, and they don't slow the line and hope. Containing and finding the cause is the work.

For this to work, people have to be able to stop the work without being blamed for it. If stopping gets you in trouble, nobody will pull the cord.

Poka-yoke: devices that stop a mistake becoming a defect

Poka-yoke means a simple, cheap device that stops an inadvertent error from turning into a defect. Think of a connector that only fits one way. There are two kinds by what they do, and the difference matters:

A control device blocks the next step until the error is fixed. A sealer that won't run until the scale reading is in range is a control.
A warning device alerts the person, who can carry on anyway. A louder alarm for an out-of-range weight is a warning.

They also work by different means. Some check shape, size or color (contact). Some check that the right number of parts or motions happened (fixed value). Some check that steps happened in order (motion step). A device that stops a mistake at the moment it would happen is stronger than a sign telling people to be careful.

What it looks like in a cannabis operation

These are ideas to look for, not a design:

Where a person types a weight, a tag number or a batch ID by hand, ask whether a scale or scanner could send it directly. Every retyped number is a chance for a mistake.
Where labels are filled in by hand, ask whether a locked template tied to the batch record, with a scan to match, could take the typing out. Labels carry many fields from the batch record and the test result, such as batch number, net weight and the cannabinoid profile (935 CMR 500.105(5)).
Where one product can be mixed up with another, ask whether color, shape or a different tray would make the wrong one obvious.

Some rules already design errors out. For edibles, solid products must be easily and permanently scored, or sold as a single serving, servings must be separable and marked, and the THC must be spread evenly (500.105(6)(c), 500.150(3)). Those are source controls written into the law.

Evidence from other fields shows what prevention can do. In one hospital study, bar-code checking at the point of giving medication cut errors not related to timing by 41.4%, and eliminated transcription errors on the study units. That's a different industry, so treat it as a sign of what's possible and not a number for cannabis.

A device can be wrong, and people can bypass it. In April 2026 the Commission issued a bulletin about a configuration error in the tracking system's action limits, which blocked compliant concentrate batches from being marked as passing until it was corrected. A control that is set wrongly can block good product. And workers can find ways around a device that slows them down. Shingo paired devices with successive checks, where the next person checks the last step, and self-checks. Ask people which devices they work around, and why.

Yield, rework and what a late defect costs

First-pass yield is the share of units that pass a step the first time, with no rework. Rolled throughput yield (RTY) is the product of the first-pass yields of all the steps. It is the chance that a unit passes every step first time. Sources vary on the exact terms and on how to count units that were reworked and then passed, so define it before you use it, and use the same definition every time.

RTY exposes something a good final number can hide. If the last test passes 98% of what reaches it, that looks fine. But some of those units were fixed on the way, and the fixing is done by what Armand Feigenbaum called the hidden factory, the part of an organization that exists to do the work again.

Steps multiply, so small losses stack. Ten steps at 98% each give 0.98 to the tenth power, which is 81.7%, and not 98%.

The cost of a defect found late

You will hear that a defect costs 1 to fix at the source, 10 to fix later and 100 once the customer has it. This "1-10-100 rule" is credited by secondary sources to Labovitz and Chang, around 1992, but I couldn't find any data behind the ratio, and I couldn't read the original. Real studies show something messier. In a study of requirements errors in large aerospace systems, fixing them in design cost 3 to 8 times as much as at the requirements stage, in build 7 to 16 times, in test 21 to 78 times, and in operation 29 to over 1,500 times. A study of 171 software projects found no consistent effect of delay. So costs often rise as value is added, but the multiplier varies a lot. Measure your own, and don't quote 100 as a fact.

Rules that decide what you can do

Quality work in cannabis runs inside rules. These are the ones that matter most for this lesson. Testing rules are being rewritten, so check the current text before you rely on a number.

No sale untested. Product has to be tested by an independent laboratory to the Commission's protocols (935 CMR 500.160). A test result is valid for one year (500.160(5)).
A failed contaminant test. The rules set out what you may do: reanalyze, remediate, or dispose of the batch. They limit how many times you can remediate and require a full-panel test afterward (500.160(13)). If a batch can't be remediated, the Commission has to be told within 72 hours (500.160(4)). The paths are limited, so retesting over and over until one passes is not an option. Ask your compliance lead before any retest.
Edible limits. Edibles have caps on THC per serving and per package, and the rules allow a 10% variance on single-serving potency (500.150(4), 500.160(12)).
Mislabeled product. Operators need a written policy to segregate and destroy it. An edible's use-by date may not be altered or relabeled later (500.150(3)(b)). Never fix a label error informally.
Inventory mistakes. A hand-typed weight or a wrong package tag can turn into an inventory discrepancy. A discrepancy that can't be explained may be reportable to the Commission and the police within 24 hours (500.110(9)). Ask your compliance lead where your line is.

A real example shows why source control matters. In 2026 the Commission audited labeled potency and found 13 of 63 flower products outside 75 to 125% of what the label said. The Commission didn't say why, so lab variance, sampling and label error can't be told apart from the outside. But every one of those causes can be reduced upstream, in how the sample is taken, how the label is filled in and how the product is made.

The arithmetic: rolled throughput yield

Every number below is made up to show the steps. Yours will differ, and none of these is a benchmark.

An example, with made-up numbers. A five-step infused-product line has these first-pass yields: mix 0.97, portion 0.98, pack 0.99, label 0.96, log in the tracking system 0.97.

Multiply them in order: 0.97, then × 0.98 = 0.9506, × 0.99 = 0.9411, × 0.96 = 0.9035, × 0.97 = 0.8763. RTY = 87.6%.
In units: of 1,000 started, 876 pass every step first time. The losses by step are about 30, 19, 9.5, 38 and 27, which add up to 124. Check: 1,000 − 876 = 124.
The trap: the average of the five yields is 97.4%, which looks fine. RTY shows that 12.4% of units went through the hidden factory.
Compounding: if every step is 98%, five steps give 0.98⁵ = 90.4%, ten give 81.7% and twenty give 66.8%. At 99%, ten steps give 90.4% and twenty give 81.8%.
A source fix: suppose a locked label template with a scan-match lifts the label step from 0.96 to 0.995. The new RTY is 0.97 × 0.98 × 0.99 × 0.995 × 0.97 = 0.908. That is up 3.2 points, 32 more good units per 1,000.

This assumes the steps are independent and that failed units aren't counted twice.

Notice where the biggest gain came from: the step with the lowest yield, fixed at the source.

Knowledge check

Five questions, graded for you. Sign in to take the check, save your progress and count this lesson toward your certificate. Sign-in opens on Monday, October 12.

Your task: draw a mistake map

This is a 15 to 20 minute walk, not a project. Fix nothing.

Pick one stage where information is written, copied or scanned: weighing, labeling or logging.
In three columns, write: the easiest mistake to make here; what catches it today (nothing, a person, or the system); and when it would be found (at the same step, at a later step, or after sale).
Mark each catch as sorting after, feedback or prevention.
Ask the person doing the work which mistake worries them most.

Record categories only. Leave out product names, batch numbers, potency figures and company details.

Open the printable mistake map

Your floor, your data. What you write on a worksheet stays with you. It is paper, and nothing you write on it reaches this site. If you talk about your floor in public, keep it general: leave out license numbers, batch or package IDs, supplier and customer names, and any number your employer treats as confidential.

Sources

Where the facts in this lesson come from. Facts last checked October 1, 2026. If you find one that is out of date or wrong, tell me.

07Room turns and changeoversPrevious lessonLesson 7 → 09Week of December 7Solving problemsNext lesson →