Measurement of what candidates do under an assessment clock

Annual review

Publication register

1. Report 2026-01Edition 32026-09-23

   How much of a timed aptitude test is never answered

   Abandonment by position on speeded and generous-clock forms.

   57.4% of the final fifth left unanswered

2. Report 2026-02Edition 12026-09-23

   What an extra thirty seconds on a question actually buys

   The return on deliberation, measured inside the same question.

   3.1 points gained, 7.3 questions lost

3. Report 2026-03Edition 12026-09-23

   Which aptitude tests are built as a difficulty ramp

   Accuracy by position, with selection controlled out.

   3 of 6 forms escalate, one runs backwards

4. Report 2026-04Edition 12026-09-23

   What candidates actually score on aptitude practice tests

   Percentile bands per form, and what a raw score is made of.

   r=0.72 with coverage against r=0.56 with accuracy

How this unit works

• One corpus, one set of filters. Every report is computed from the same population by the same code, so any two of them can be cited side by side. Method and standards
• Every discard is published. More than half the attempts on record are excluded from analysis. The ledger showing why appears in identical form on every report.
• Confounds get controls, not caveats. Where a control is impossible, the claim is weakened to match rather than dressed with a disclaimer.
• Withdrawals stay visible. One finding was retracted this year. It is documented, with its reason, on the report that replaced it.
• Figures are files. Every chart is written by the script that computes the numbers, so a figure cannot drift from its table, and every report ships its figures as CSV.

In the pipeline

• Where the clock actually goes. The distribution of time across the paper rather than across items.
• Item-level difficulty calibration. Ranking questions by observed difficulty rather than by position.
• Does a second sitting help? Measured, and not being published. Of eighty repeat pairs, sixty-one re-sat the same form, so most of the apparent gain is item memory. It is the one number here that would flatter us, which is why it stays unpublished until a clean comparison exists.