The State of Timed Assessment 2026

1. 57.4% of the final fifth of a speeded paper is never answered against 12.1% where the clock is generous, on tests where a wrong answer costs nothing. Report 2026-01
2. Thinking longer about a question is worth 3.1 points once the question is held constant, while the surplus time spent doing it would have covered 7.3 further questions. Report 2026-02
3. Papers do not share a difficulty shape. PI Cognitive, CCAT and Wonderlic escalate steeply, Cubiks Logiks and IBEW aptitude run flat, and Watson-Glaser opens with its hardest block. Report 2026-03
4. On a speeded paper the raw score is mostly coverage (r=0.72 against r=0.56 for accuracy), and the ordering reverses once the clock stops binding. Report 2026-04

Taken together these describe a single failure mode. A candidate meets a hard item early, spends a minute on it because that feels like the diligent thing to do, gains almost nothing on that item, and pays for it with the last ten questions of a paper that was getting harder anyway and that scores an unanswered item exactly as it scores a wrong one.

Every report in this series is computed from one corpus with one set of filters, so any two of them can be cited side by side. 1,473 attempts are on record. The filters remove attempts that cannot be scored honestly against the form as it now stands, and attempts that show no evidence of having been sat.

Candidates who failed to finish are kept, deliberately. Not finishing is the central measurement of the year. Filtering on it would be selecting on the outcome being reported. The full reasoning, including the two thresholds we tested and rejected, is in the method note.

The shortfall is not distributed across the paper. It arrives as a cliff. In the opening fifth 0.2% of items go unanswered; by the fourth fifth it is 31.9%, and in the closing fifth 57.4%. Generous-clock forms, facing the same kinds of reasoning items, end at 12.1%.

Because none of these formats deducts marks for a wrong answer, every one of those blanks is a choice that scores zero with certainty in place of a choice with positive expected value. Read Report 2026-01 in full.

The obvious explanation for an abandoned ending is that candidates get stuck. The obvious remedy, repeated in every prep guide, is that dwelling on a question destroys your accuracy on it. We tested that and it is largely untrue.

Compared without controls, accuracy appears to fall by around twenty points from the fastest band to the slowest. Compared inside the same question, where item difficulty is held exactly constant, the gap is 3.1 points, and the faster half wins on only 54% of questions.

The real cost is arithmetic. The median speeded attempt spends 102 seconds beyond a thirty-second-per-item cap, which at a median item time of 14 seconds is 7.3 further questions. 63% of attempts contain at least one question that ate a full minute. Read Report 2026-02 in full.

Candidates are told, as a general rule, that aptitude tests get harder as they go. The rule holds for half the forms measured and fails badly for the rest.

The CCAT form loses 36.4 accuracy points from its opening fifth to its closing fifth and the Wonderlic form 42.8. The IBEW form moves 0.1 points, which is to say it does not move. The Watson-Glaser form runs the other way entirely, opening at 53.1% and closing at 89.7%.

This is why mid-test self-assessment misleads. On a ramping paper the opening block is the easiest block, so feeling comfortable there carries no information. On the Watson-Glaser the reverse applies, and abandoning the paper on the strength of a poor first section is a real and avoidable error. Read Report 2026-03 in full.

A raw score decomposes exactly into how much of the paper was reached multiplied by how well it was answered. Which term dominates is decided by the clock.

On speeded forms the average attempt answers 72% of what it attempts and reaches 80% of the paper. The missing marks are overwhelmingly on items never seen. Where the clock is generous, coverage rises to 96.1% and stops discriminating between candidates at all, so the score reverts to measuring what people assume it measures. Read Report 2026-04 in full.

Plot every form by the seconds it allows per question against the share of its items that never get an answer, and the four reports collapse into one picture.

The exception is the most instructive point on the chart. McQuaig allows 18 seconds a question, exactly what the CCAT allows, and leaves 0.2% of its items blank against the CCAT’s 20.6%. Seconds per question is not the variable. Seconds per unit of work is, and a candidate cannot read that off the instructions: they have to have sat the format before.

Withdrawn, 21 September 2026. We published, and retracted within a day, the claim that time pressure degrades the answers a candidate does give. Splitting speeded attempts by how much of the clock they actually used killed it: candidates who finished with time to spare declined 27.6 accuracy points against 31.3 for the time-pressured. Almost the whole effect was present in people under no pressure at all. What was left, once that was taken seriously, became Report 2026-03, which is the better finding.

Measured and not published. Attempts that were re-sat show a gain of roughly twelve points. We are not publishing it as a finding about practice, because 61 of the 80 repeat pairs re-sat the same form, so most of the gain is item memory rather than improvement. It is the one number in this corpus that would flatter the company that collected it, which is precisely why it stays where it is until a clean comparison exists.

Not published at all. The per-question records. Joined to an attempt, timing data at this sample size is re-identifiable. Every report publishes its aggregated figures as CSV and none of them publishes the underlying rows.

• Every discard is published. The ledger in section 2 appears in identical form on every report. More than half the attempts on record are excluded, and the reason is stated rather than buried.
• No filtering on the outcome. The filters are fixed before a finding is looked at, and a filter that would select on the thing being measured is rejected even when it produces a cleaner number.
• A confound gets a control, not a caveat. Both of this year’s main findings survive a direct control: within-item for pacing, full-coverage-only for the ramp. A caveat is what we write when a control is not possible, and then the claim is weakened to match.
• Withdrawals stay visible. A retracted finding is documented with its reason on the report that replaced it, not deleted.
• Figures are files. Every chart is generated by the same script that writes the numbers, so a figure cannot drift from its table, and both the image and a vector copy are downloadable.

Speeded form

A test that allows under thirty seconds per question, so the binding constraint is the clock rather than the difficulty. On this platform: the CCAT, Wonderlic, PI Cognitive and Cubiks Logiks formats.

Generous clock

A test allowing thirty-three seconds or more per question, where nearly every candidate reaches nearly every item. On this platform: IBEW, Watson-Glaser, EIAT and Ramsay Mechanical formats.

Coverage

The share of items on the paper that received any answer at all, whether right or wrong.

Blank, or unanswered

An item that received no answer. Every format measured here scores a blank identically to a wrong answer, which is to say zero, with no deduction for the error.

Gradient

Accuracy in the opening fifth of the paper minus accuracy in the closing fifth, in percentage points. Positive means the paper gets harder as it goes.

Full-coverage attempt

An attempt that answered every single item. Used as the control population for the gradient, because no candidate is missing from any point on the curve.

Within-item comparison

A comparison drawn inside a single question, splitting its sittings at that question's own median time, so item difficulty is held exactly constant.

Re-grade filter

Every attempt is scored again against the form as it now exists on disk. Disagreements are discarded, because banks were edited after launch and an old attempt cannot be scored against a paper it was not sat on.

• Where the clock actually goes. The distribution of time across the paper rather than across items, and whether candidates who front-load their clock can be identified before they run out.
• Item-level difficulty calibration. Ranking individual questions by observed difficulty rather than by position, which would separate a genuine ramp from an ordering artefact.
• A clean repeat comparison. A second sitting on a different form of the same test, which is the only design that can separate practice from item memory. It requires data we do not yet have.

If you want a cut we have not published, or a figure checked against the source, ask and we will run it.