Pass A and the Coverage Grid
20 September 2026. Two things: whether the 36 writing hits are real, and the family-by-dimension grid the plan requires before the prompt is written. Corpus at b6319a4.
Are the writing hits real? Mostly yes, and the largest cause is a stem that cannot discriminate
Pass A hit 46 of 439 items. The distribution is not even:
| Section | Hits | Rate |
|---|---|---|
| writing | 36 / 131 | 27% |
| math | 8 / 130 | 6% |
| reading | 4 / 178 | 2% |
28 of the 36 writing hits carry the stem "Which choice completes the text so that it conforms to the conventions of Standard English?" Of those, 12 test concision, register or redundancy rather than a rule.
The clearest case, l-writing-padder-1
PASSAGE The final section of the manual is ______ and can be understood by anyone.
STEM Which choice completes the text so that it conforms to the conventions
of Standard English?
* A. clear
B. clear and easy to understand
C. clear, plain, and understandable
D. clear in a way that is understandable
A is the keyed answer and it is the most concise. All four are grammatical. The item tests redundancy against a passage that already says "can be understood by anyone" — a real and good test — but the stem asks a question every option passes. exactly_one came back 0.26.
The split, and what each half needs
- 12 stem/construct mismatches —
padder(6 with twins), plusbreather,understudyand mixed items. A real, repairable defect with a precise repair: the stem, not the options. These are theW-REDUNDANT,W-REGISTERandW-SOUNDS-FORMALtraps. - 1 marginal false positive —
w-gen-dangle-4atsecond_live0.51, a hair over the line. Its distractors are all genuinely wrong; the margin is noise. - The remaining ~23 are dangling-modifier and boundary items where the distractors are wrong by degree rather than flatly wrong. Worth an author's eye, and they are not the same job as the 12.
The coverage grid
An empty cell is deliberate and visible. This is the artifact the plan says goes to you before the prompt is written, because it shows the majority family carrying three live dimensions and a fourth held on its owner's own failure prediction.
| Dimension | Status | Wager 258 | Lineup 181 | Prime Suspect 70 threads | flavor |
|---|---|---|---|---|---|
| defensibility | ships | 258 | 181 | 70 | — |
| connection to theme | ships | 238 | 157 | 9 briefs | — |
| open-question test | ships | 258 | 181 | 70 | — |
| earned wrongness | ships | 45 | 94 | 70 | — |
| trap recognition | ships | via cons | via cons | via cons | 30 glosses |
| contested-ness | held at gate | 258 | 48 | — | — |
| fairness | held | 45 | 94 | — | — |
| con recurrence | dead, run one | out | out | — | — |
| resonance of voice | CUT | — | — | — | 150 lines |
- The Wager's 45 is the same 45 on every axis. Earned wrongness and fairness both reach a sixth of the family. The 213 get carried beside them every run rather than averaged in.
- Contested-ness reaches 48 of 181 Lineup items. The other 133 carry no trap label of any kind, which is exactly where its owner predicted it fails.
- Con recurrence is out. Only 8 of 47 con/trap pairs clear the five-appearance floor.
- Four dimensions ship on the Wager — defensibility, theme, the open-question test and trap recognition — with earned wrongness reaching a sixth of it.
Resonance is cut, and the rule that said otherwise is left standing
The pre-registered register run: 150 lines, 12 declared registers, each line scored twice with the id never sent.
| Measure | Result |
|---|---|
| Self-agreement across two runs | 139/150 — 93% |
| Exact match against the declared bucket | 57/150 — 38% |
| Pairs meeting the merge rule | 0 |
| Mechanical verdict | ships on 12 bands |
Jev reads something reliably and it is not the declared register. needling is declared on 40 lines and predicted on 85. hesitation and withdrawal are never predicted at all. savage → needling happens 18 times while the reverse barely happens.