Central schemas in data_validation/schemas/ have drifted from the datasets: decide per file (point, retire, or keep)

[emmanuelgjr]

Context

The policy ratified when merging #62: dataset-local schemas are authoritative, and the central files in data_validation/schemas/ are historical stubs unless a dataset actually validates against them. #73 applied this to agentdataflow_trace.schema.json by turning it into a $ref pointer. The same problem exists across the directory. For example, all 59 entries in datasets/exploit_dataset/ fail exploit.schema.json (reported in #73).

Current state (as of main, 2026-09-17)

Central schema	Dataset-local schema	Used by	Proposed action
agentdataflow_trace.schema.json	yes	nothing	done in #73 (pointer)
exploit.schema.json	datasets/exploit_dataset/schema.json	nothing	pointer, as in #73
riskassessment.schema.json	datasets/riskassessment_dataset/schema.json	check	pointer, as in #73
vulnerability.schema.json	datasets/vulnerability_dataset/schema.json	check	pointer, as in #73
promptinj_testcase.schema.json	none	live: promptinj_dataextraction_testcases README, tr_turkish_conversation_prompt_injection_pairs/validate.py and build_cases.py	keep as authoritative, or move into the dataset folder and leave a pointer
incident.schema.json	none	data_validation/validators/schema_validator.py usage example	verify against datasets/incident_dataset/; then either move local or keep
rag.schema.json	none	check	verify against datasets/rag_dataset/
crossframework_mapping.schema.json	none	check	the crosswalk now lives in GenAI-Security-Project/crosswalk; decide whether this dataset and schema still belong here

Rule going forward

Every dataset with entries ships its own schema.json, a validator, and a path-filtered CI workflow (the agentdataflow_toolexchange_traces/ layout is the reference). Central files are either pointers or removed.

How to help

The three pointer conversions are small and independent. Copy #73 and include a $comment listing the concrete divergences. Before converting, grep for references to confirm nothing validates against the central file. One PR per schema, please.