Rule counts depend on the representation the modeler chooses, so different modelers may produce different counts

What it is

CCT's predictor is a rule count, but the boundary of "one rule" is the modeler's choice: "select a file, then press delete" can be written as one rule or split into "select," "press delete," and "confirm deletion," and both the difference set and the predicted learning time change accordingly.

Why it happens

A production model is a hypothetical decomposition of task knowledge; the same behavior admits different condition-action segmentations. Granularity, how conditions are merged, and whether error handling is represented explicitly all change the count. Relative rankings are fairly stable under one representational convention, while absolute values drift with it.

Studying it

The property itself is testable: have several modelers independently analyze the same pair of interfaces and compare divergence in difference counts and predictions; repeat under a shared coding manual and measure how much divergence shrinks. The methodological conclusion is to report the convention and run sensitivity analyses rather than treat a count as an objective quantity.

Where it stops holding

When candidate designs differ by less than the modeling divergence, counts cannot distinguish them; comparisons across teams or time are not comparable unless the convention is fixed. This limits use to relative comparison under one convention; it does not invalidate the theory's direction.

Applying it

• Agree on a rule-writing convention first—granularity, condition-merging criteria, whether error handling is separate—and put it in the review template.
• Compute difference sets for candidates with one modeler under one convention, then check granularity sensitivity: re-split legally and see whether the ranking flips.
• Report estimates with the convention and a count range, never a single bare number.

Related

• Same group: B4.14.1 Learning time can be estimated from the number of rules the user must newly acquire · B4.14.2 Mastered rules transfer; the more two interfaces share, the cheaper the second is to learn · B4.14.3 The theory yields quantitative predictions of transfer gains between old and new systems
• Nearby: B4.11 GOMS models · Q4 Measurement and reliability
• Search terms: granularity of analysis · modeling assumptions · production system