The golden sample is a decision record, not just a good part

A golden sample is a physical part, formally approved, that fixes where accept ends and reject begins for one named attribute. The part is the easy half. The half that does the work is the record attached to it: the drawing revision, the attribute it demonstrates, who approved it and in what role, the date, where it lives, and what triggers a re-check.

Quality teams mostly know this. It stops being obvious the moment a camera is involved, because the sample changes job: from something an inspector glances at to the origin of every threshold, every training image, and every argument 8 months on.

Four things in the same drawer, doing different jobs

The pair that decides an inspection is golden and limit. A threshold has to sit between them, so where only one exists the boundary gets chosen by whoever configures the cell rather than whoever owns the specification — a vision project setting quality policy by accident.

The record the part has to carry

1. Part number and drawing revision. A sample approved against rev C stops being evidence the day rev D changes the radius it shows. Most often omitted; invalidates everything else.
2. The attribute, named singly. "Surface finish on the show face, as moulded" is a reference. "Good part" is a mood. One part carries several attributes only if each has its own approval.
3. Which side of the boundary it sits on — reference condition, or worst acceptable. Put it on the label; in 12 months nobody remembers which drawer held which.
4. Approver, role and date. Quality owns the boundary, engineering the mechanism. A record signed only by the integrator has outsourced the specification to someone who never sees the complaint.
5. Storage location and condition. Named cabinet, named plant, stated environment. A sample nobody finds inside 10 minutes is not a sample at 02:00 on a night shift.
6. Revalidation trigger: a calendar interval plus an event list — drawing revision, supplier change, tooling refurbishment, any lighting change.

Ageing, handling, and the re-check nobody schedules

Reference parts decay invisibly day to day and obviously across a year. Painted surfaces yellow under continuous light. Unprotected steel flash-rusts in a humid plant. Elastomers take a compression set in a tight bag. Polished aluminium collects handling marks from the hands that keep proving it is fine. The sample drifts toward reject while the drawing has not moved.

• Store it sealed and dark. Opaque box with desiccant for anything coated or printed; its own bag for a critical edge.
• Handle it with gloves, and log it. Coming out of the box for a camera set-up is an event, not a favour.
• Photograph it on approval day — the only evidence of its condition before 2 years of handling marks.
• Re-check on both triggers. The 12-month clock catches drift; the event list catches step changes.

A reference part handled a hundred times is no longer the reference. It is a hundred marks with a label on it.

What an automated cell inherits from it

An inspection system does not soften a vague reference; it industrialises it. If the sample leaves a question open, someone answers it anyway — whoever picked the training images, or nudged a setting on a Tuesday — and that answer holds for 3 years. The record moves the decision back to its owners, where it becomes the defect class list your inspection is graded against.

The reference also decides which approach is available: a controlled set of conforming samples is the entry requirement for training on good parts only, the trade set out in when you have forty defect images, and when you have four thousand. A sample that has drifted toward reject pushes conforming parts across the line, which the floor experiences as good parts pulled off the belt every shift, not as a paperwork problem.

Where the record lives matters too. It needs revision control, an approver identity and an audit trail, so it belongs in the quality system or the plant execution layer — one reason a site weighs a custom MES against an off-the-shelf one. The surrounding imaging decisions sit under visual inspection and defect detection; the practice itself under industrial vision.

A reference only one person can read is an unattended system waiting to fail quietly — the pattern in what breaks when AI agents run in production, and why automation work begins with written decisions.

• quality reference
• machine vision
• inspection setup
• sample control