I — The bench
A gage study is planned before it is measured
Ten parts, three appraisers, three trials, randomized: that is decided before a single number exists, and the numbers come back hours or weeks later. The SVEND Gage R&R bench holds that gap: it issues the run sheet, receives the readings, keeps the order it issued, and judges the completed study.
A Workbench analysis operates on an existing dataset. A gage study cannot, because the method specifies what to collect before any data exists. It is therefore a separate bench rather than a row in the analysis catalog, and it holds state.
declare parts and appraisers → issue a randomized run sheet (seed recorded) → capture readings, by hand or off the instrument → judge the completed study by ANOVA → a record whose readings are hash-chained
II — The declaration
| Field | Behavior |
|---|---|
| parts | Labels. A label such as A1 or SN-4471 identifies an item directly; a positional reference such as “part 7” requires the operator to establish which item was measured. Up to 40. |
| appraisers | Named, for the same reason. Up to 10. |
| trials | Repeat measurements per appraiser per part. Up to 6. |
| characteristic, units | What is being measured, and in what. A study that names neither reads back as a column of numbers. |
| gage | The instrument label as typed. It may also resolve to a registered device; the label and the referent are kept apart deliberately. |
| USL / LSL | Optional. Left empty when the study is judged on study variation alone. An empty field records the absence of a specification limit; a zero would record a limit of zero. |
III — Presentation order is part of the method
An appraiser who measures the parts in the same order every trial is measuring their own memory alongside the gage. So every appraiser gets a fresh permutation for every trial, and the run sheet is issued in that order rather than in part order.
The order is generated from a seed that is recorded with the study. That matters on a real floor: a sheet gets printed, half filled in, and printed again, and the reprint has to be the same sheet. Re-planning a study that already has readings recorded against it is refused: the numbers on the floor belong to the plan that produced them, and a fresh permutation would silently re-file them against different rows.
IV — Capture
Readings can be typed, or sent by any digital gage or receiver that presents itself as a keyboard. The run sheet is built for the second case: the first unmeasured row takes focus when the sheet opens, Enter or Tab commits the value and moves to the next unmeasured run rather than the next field on the page, focus survives a save, and the active row is marked so an operator watching the gage rather than the screen can see where they are.
Text that is not a reading holds the row instead of advancing past it, and a row left blank is treated as a deliberate skip and collected on the way round rather than silently lost. See the digital gage input contract for device setup.
V — The record
A study is edited while it is being set up, so chaining the study row itself would prove nothing. What is chained is the act: each submission of readings is recorded as an event carrying who submitted what, against which rows, and when, and those events are hash-chained.
| Act | What the record does |
|---|---|
| A correction | Is a later event. The earlier reading stays; the history shows the change rather than replacing it. |
| An erasure | Is an act. The record carries an event whose value is null. |
| Re-attributing a reading | Breaks the chain, which is the alarm working. |
| Fixing a typo in the plan | Does not break anything. The declaration is deliberately outside the chain, so correcting a part label cannot invalidate the measurements. |
VI — The verdict
A completed study is judged by crossed Gage R&R through the same engine the Analysis Workbench uses: ANOVA variance decomposition reporting repeatability, reproducibility, the operator-by-part interaction, %GRR by study variation, and the number of distinct categories. A part-worked sheet is refused, naming what is missing, because a crossed study is judged whole or not at all.
Worked against a full AIAG-shaped study (10 parts, 3 appraisers, 3 trials, 90 readings) the bench returns %GRR 6.7% and 20 distinct categories, with a seeded appraiser bias landing in reproducibility where it belongs.
VII — Boundaries
- Crossed and nested. A nested study deals each appraiser its own parts, for destructive testing where no part can be measured twice.
- Gage stability is not on this bench. It has a surface of its own (one master of known value, measured for as long as the instrument is in service) and it belongs to the Device Trust System on the Team tier, because checking a master against its known value over months is calibration work rather than a study run on parts and appraisers.
- Linearity and bias run in the Workbench. They require reference values across the measurement range rather than a run sheet.
- Device identity is taken from the declaration. Structured device capture is specified but not shipped.