Context
Found by adversarial review of PR #154 (rubric-review). Not covered by the existing follow-ups (#155–#160).
Problem
PR #154 adds to both semantic agents the check: "Is the description semantically appropriate for the claimed dataset type and program of origin?" (.claude/agents/d4d-rubric20-semantic.md:58-61; same edit in d4d-rubric10-semantic.md).
But there is no program/project field in the D4D datasheet to ground this on. D4D_Core.yaml defines a 95-field exchange schema with no program/project attribute (grep for program|project in src/data_sheets_schema/schema/D4D_Core.yaml returns nothing). The only project concept lives in the batch evaluation summary metadata (D4D_Evaluation_Summary.yaml), not in the datasheet under evaluation.
Impact
With no field to read, the evaluator will infer "program of origin" from filenames, keywords, publisher, or prior knowledge — so the same datasheet can score differently depending on invocation context (filename, who's asking). This undermines the agents' stated temp=0.0 / "same file → same score" reproducibility guarantee.
Recommendation
Either (a) name the exact datasheet fields that establish program of origin and require quoted evidence (e.g. inferred only from keywords / publisher / funders), or (b) drop "program of origin" from semantic scoring and keep project strictly as batch metadata.
Evidence
.claude/agents/d4d-rubric20-semantic.md:58-61.claude/agents/d4d-rubric10-semantic.md (same "program of origin" edit)src/data_sheets_schema/schema/D4D_Core.yaml (no program/project slot)
Context
Found by adversarial review of PR #154 (
rubric-review). Not covered by the existing follow-ups (#155–#160).Problem
PR #154 adds to both semantic agents the check: "Is the description semantically appropriate for the claimed dataset type and program of origin?" (
.claude/agents/d4d-rubric20-semantic.md:58-61; same edit ind4d-rubric10-semantic.md).But there is no
program/projectfield in the D4D datasheet to ground this on.D4D_Core.yamldefines a 95-field exchange schema with no program/project attribute (grep forprogram|projectinsrc/data_sheets_schema/schema/D4D_Core.yamlreturns nothing). The onlyprojectconcept lives in the batch evaluation summary metadata (D4D_Evaluation_Summary.yaml), not in the datasheet under evaluation.Impact
With no field to read, the evaluator will infer "program of origin" from filenames, keywords, publisher, or prior knowledge — so the same datasheet can score differently depending on invocation context (filename, who's asking). This undermines the agents' stated temp=0.0 / "same file → same score" reproducibility guarantee.
Recommendation
Either (a) name the exact datasheet fields that establish program of origin and require quoted evidence (e.g. inferred only from
keywords/publisher/funders), or (b) drop "program of origin" from semantic scoring and keepprojectstrictly as batch metadata.Evidence
.claude/agents/d4d-rubric20-semantic.md:58-61.claude/agents/d4d-rubric10-semantic.md(same "program of origin" edit)src/data_sheets_schema/schema/D4D_Core.yaml(no program/project slot)