Writing the defect class list your inspection will be graded against

The defect class list is the contract between quality and engineering, and must exist before any images are collected. Each class needs a name, a definition, a severity, a disposition, a minimum reportable size, and a pair of images — one that is the class, one that deliberately is not. Anything less is a word, and a word cannot be scored.

The failure this prevents is specific. "Scratch" is the usual offender: it covers a light mark in the coating that ships without comment and a gouge through to the substrate that is scrapped. Two dispositions, one label. The model learns an average of the two, and the confusion matrix reports accuracy on a category that never existed.

The six fields a class has to carry

1. Name and one-line definition, written so a new inspector reaches the same verdict on day 1. "Coating void: substrate visible through the coating" leaves less room than "bad finish".
2. Severity. Critical, major or minor against function or the customer requirement — not against how obvious it is to look at.
3. Disposition. Scrap, rework, use-as-is, quarantine. This is the field the line acts on, and every class needs exactly one.
4. Minimum reportable size, in part units. "Length ≥ 5 mm on the show face" is checkable; "visible" depends on the light. Never pixels, which change with lens and standoff.
5. Where it can occur — named faces or zones. A mark that is minor on a hidden face and major on the show face is 2 rows, not a footnote.
6. Example and counter-example. The counter-example is the one everybody skips and the one that does the work: it fixes the boundary against the class next door.

A worked list for one part family

Below is the shape for a powder-coated steel bracket with one show face. The numbers are illustrative — yours come from the drawing — but the columns are the point. Any class you cannot fill in across all four is not yet a class.

Four rules that keep the list usable a year later

• Mutually exclusive by construction. If a part can honestly belong to 2 classes, write the precedence rule into the definitions rather than leaving it to whoever holds the part.
• One disposition per class, no conditions hidden in prose. A class whose disposition depends on a second judgement is 2 classes.
• An explicit unclassified bucket, reviewed on a fixed cadence. Without it every surprise is pushed into the nearest wrong class.
• Version the list. Adding a class after a model exists invalidates the earlier grading, so the change needs a date, an owner and a note on what must be re-scored.

If two experienced inspectors cannot put the same part in the same class, no amount of training data will make a model do it. It will just be confidently inconsistent instead of visibly inconsistent.

What the list decides downstream

The disposition column reaches the floor. Classes dispositioned use-as-is should never be wired to a reject actuator, however confidently detected, because each one is a good part pulled off the belt and a minute of somebody's re-check. Severity runs the other way: it names the classes an escape is unacceptable for.

The minimum-size column sorts the technology. A class defined by a number a caliper could confirm belongs with classical tools, not a learned stage — the split is argued in the measurement your caliper agrees with. What the list does not do is set sensitivity: where the cut falls stays a governed decision, described in the anomaly score is a ranking, the threshold a business decision.

The list is also what acceptance is measured against: per-class recall targets and overkill budgets can only be written once the classes exist, the argument in an exit criterion is not a good feeling. That is why the taxonomy is a scoping deliverable in MVP and product builds, alongside visual inspection and defect detection and the wider industrial vision practice.

• defect taxonomy
• quality standards
• inspection design
• severity