The Cloze Paradox
Mei, 19, is preparing for the DET with a target of 120. Her tutor's log from three weeks of daily practice shows a stable pattern. On Read and Select - spotting real words among invented ones such as 'brantle' and 'plim' - she averages 94%. On Read and Complete, where six blanks in a 100-word passage each show a few letters, she averages 41%; 8 of her 11 errors fall on function words ('nevertheless', 'whereas') and verb–preposition pairs ('depend ___', 'result ___'). On full passages she answers 80% of detail questions and 45% of inference questions correctly. Her vocabulary is estimated at 5,800 word families, close to what a 120 profile typically needs. In debrief she says: 'I know all these words. I just run out of time and guess at the end.'
1. Which sub-skill is the binding constraint? Name the two rival explanations the data weakens and say exactly which numbers weaken them.
Read the worked answer
Binding constraint: choosing the right word in context - local cohesion and collocation - not word knowledge.
- Rival 1: vocabulary breadth. Weakened twice. 94% on Read and Select means she can separate real forms from invented ones near ceiling, and 5,800 word families is already in range for her target. More decisively, 'nevertheless' and 'whereas' are high-frequency items: a breadth deficit cannot explain why 8 of 11 errors land on the most common words in the passage.
- Rival 2: time pressure alone. A pure speed problem scatters errors across the whole passage. Hers cluster by category (function words, verb + preposition). That is a signature of usage knowledge, not of a clock. Her own remark - 'I guess at the end' - is a hypothesis to test, not evidence; self-report about time is notoriously unreliable.
The explanation that survives both facts: she recognises words in isolation but has not automatised the grammatical frame they sit in ('result in', 'depend on', contrastive 'whereas').
2. Allocate her next two weeks. State what you cut, what you add, and one measurable exit criterion.
Read the worked answer
Cut: Read and Select drops to 5 minutes a day as maintenance - she is at ceiling and further gains there cannot move her score. Inference work is deferred, not abandoned; it is a second-order problem while cloze sits at 41%.
Add (20–25 min/day):
- Ten cloze items per day from unseen passages, coded into three bins: (a) function word/cohesion, (b) verb + preposition, (c) word form.
- A post-mortem on the blank's sentence plus one clause either side only - this trains local cohesion rather than rewarding a slow global reread.
- Collocation retrieval aloud: read the verb, produce the preposition and a full clause.
Exit criterion: 65–70% on unseen six-blank passages by day 12, with the error distribution shifting away from function words. If bin (a) does not shrink, the diagnosis was wrong and the plan should change - the criterion is a test of the hypothesis, not just a score target.
3. What single observation would most change your diagnosis?
Read the worked answer
A timed-versus-untimed comparison with error-position data.
- If untimed accuracy jumps above roughly 75% and errors concentrate in the final third of passages, the constraint is automaticity under time, not knowledge. The plan then flips to speed work: repeated timed cloze, chunked reading, deliberate practice at 60–70% of her comfortable pace.
- If she picks the correct word when the blank's sentence is shown in isolation but fails inside the passage, the problem is cross-sentence cohesion and working memory, not collocation - different remediation (signalling words, paragraph mapping).
A second, cheaper check: ask her to justify a wrong answer. If she can explain the grammar rule after the fact, the knowledge exists but is not retrieved under load - again a fluency, not a knowledge, problem.