Understanding Confidence Scores

Introduction

Every dimension, bore, chamfer, thread, radius, roughness and GD&T frame the Werk24 API returns carries a confidence score between 0.0 and 1.0. It answers one question: did the API find a concrete reason to doubt this reading? This article explains what the score means, what lowers it, and how to use it.

What Is a Confidence Score?

A reading starts at 1.0 and loses confidence only for a reason the API can name: an unclear label, a value that does not fit the drawing, a tolerance that is on the drawing but could not be read. A single serious reason takes a reading below 0.8; so do two smaller ones. A minor reason on its own keeps it above.

So the score has one line that matters:

• 0.8 or above: a reasonable reading. Nothing serious was found to doubt it.
• Below 0.8: at least one concrete reason to doubt the reading was found. Check it.

Within each side, a higher score means fewer or smaller doubts. A clean reading scores 1.0.

The score is not a probability. A reading at 0.9 is not "right 90% of the time"; it is a reading with a small doubt about it.

What Lowers a Score?

Whatever the feature, how the label was read counts:

• Unclear text. The characters of the label could not be read with certainty, or part of the label could not be read at all.
• A label only partly understood. Some of the text on the label could not be interpreted.

Each kind of feature also has its own checks:

• Dimensions, bores, chamfers and threads
  • A tolerance that is on the drawing but missing from the reading, for example a stacked +0.1/-0.2 or a fit such as H7 that could not be read. Whether a dimension carries its own tolerance or falls under the general tolerance does not change its score; only a tolerance that was there and was lost does.
  • A tolerance no drawing would carry: an upper deviation below the lower one, or a tolerance band far tighter or far wider than the ISO 286 grades allow for that size (usually a misplaced decimal point).
  • A label whose value does not match the length drawn at the drawing's scale. Dimensions drawn deliberately out of scale are not held to this.
• Surface roughness
  • A value that is not on the series designers use. Ra 3.2 and Ra 1.6 are standard values; Ra 3.5 is not, and is most likely a misread 3.2. Rz values are held to the R10 preferred numbers (Rz 4, Rz 10, Rz 16, ...), microinch values on imperial symbols to the usual averages (16, 32, 63, 125, ...), and N grades to N1 to N12.
• GD&T frames
  • A frame without its tolerance value.
  • Datums that do not fit the characteristic: a datum on a form tolerance such as flatness, which takes none, or a perpendicularity, parallelism, angularity, runout, concentricity or symmetry tolerance without one.

Where Are Confidence Scores Used?

The scale described here applies to the confidence field of dimensions, bores, chamfers, threads, radii, roughness and GD&T frames. See Confidence in the API reference.

Other objects carry a confidence field too, and do not follow this scale yet: on notes it says how clearly the note was detected, and fasteners and standard features (undercuts, key slots, center holes and the like) report 1.0.

How to Interpret Confidence Scores

Best Practices for Using Confidence Scores

1. Use 0.8 as the threshold for automation. Accept readings at 0.8 or above automatically and route the ones below it to review. That is the line the score is built around.
2. Raise it for critical work. Where every wrong value is costly, review everything below 1.0, or review the features that matter most whatever their score.
3. Review the lowest scores first. When review capacity is limited, the readings furthest below 0.8 are the ones with the most reasons to doubt them.
4. Filter on your side. The API returns every feature it read, whatever its score; it does not drop low-confidence readings for you. The confidence_min parameter of the V1 VARIANT_MEASURES ask is no longer applied.