AN INITIAL PUBLIC SAMPLE
Different tasks.
More than one right answer.
Explore what each rubric rewards, where it may be too narrow, and which valid approaches it should leave open.
Initial editorial review; not accepted training data.
Coverage across task families
Counts describe this sample, not the full source datasets. Empty families show where more examples are needed.
How to read this collection
These examples are a discovery sample, not a representative distribution or a set of accepted training targets. Family assignments and rubric reviews are initial editorial judgments. Dataset availability does not imply the same license for every source.
A source-provided rubric is preserved for inspection. A drafted rubric is a new candidate and requires validation. Suggested changes are kept separate from the original. A valid-alternative probe is a candidate response used to test whether a rubric is too restrictive; an invalid control tests whether it is too permissive.
Read each probe’s verification label: an editorial expectation is not an executed rubric score. A response sketch is not a complete solution. “Full response” describes the specimen’s scope, not whether it passed an automated check.