Skip to content

Commit d8d9fa1

Browse files
committed
docs(changelog): record judge construct measurement contract
1 parent f2a8b0a commit d8d9fa1

1 file changed

Lines changed: 7 additions & 0 deletions

File tree

Lines changed: 7 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,7 @@
1+
# Judge construct measurement contract
2+
3+
## Added
4+
5+
- Add a package-level LLM-as-a-Judge measurement contract that treats each rubric criterion as one dichotomous or polytomous item, enforces zero-based category coding, and separates the three-item identification floor from the package's five-item default and seven-item recommended facet policy.
6+
- Project judged responses into deterministic persons × items matrices only when every result carries exactly the criterion identities declared by the construct specification, and preserve `spec.criterion_ids` as the authoritative output-column order so downstream item parameters cannot be mislabeled by lexical key ordering.
7+
- Keep GRM/GPCM fitting, likelihood, scoring, recovery, and all production psychometric arithmetic Rust-owned; the new surface performs validation, marshalling, and recovery evidence only.

0 commit comments

Comments
 (0)