benchmark
Seven item sources on one blueprint. Six are generation arms; the seventh is the reference set of parameterised templates (written by the build agent, not reviewed by a human educator), audited with the same battery. The table reports every dimension separately. It does not crown a winner, because an arm can win on cost and lose on correctness.
Blueprint cells60dev 24 · val 12 · test 24
Constructs / subskills16 / 35
Items180reference only (run pending)
Generation cost—measured, per-call USD
by dimension
pre-registered tests
Primary contrasts are paired by blueprint cell and Holm-corrected. The confirmatory scope is the 24-cell test split.
composition
| Construct | Cells |
|---|---|
| C01 conditional probability | 6 |
| C02 bayes rule | 4 |
| C03 independence | 4 |
| C04 probability rules | 4 |
| C05 expected value | 4 |
| C06 variance and sd | 4 |
| C07 binomial distribution | 6 |
| C08 normal distribution | 6 |
| C09 sampling distribution | 4 |
| C10 confidence interval | 4 |
| C11 hypothesis test logic | 4 |
| C12 error types and power | 2 |
| C13 correlation | 2 |
| C14 regression | 2 |
| C15 descriptive measures | 2 |
| C16 study design | 2 |