benchclaw-stage4-cdm-irt-analysis

To run IRT and CDM diagnostics on item-level scores, you must dispatch this analysis as a subskill within a benchmark pipeline using the `/benchclaw-subskill` command, passing the appropriate project root, node ID, and artifact paths.

IRT difficulty and discrimination analysis evaluates item-level score matrices to estimate model ability and item performance, diagnosing quality control issues in small-batch model responses before full benchmark synthesis.

You need a sub-agent context to use this IRT diagnostic tool, as it is designed specifically as a subskill that requires explicit dispatch via the `/benchclaw-subskill` command rather than standalone execution.

You should perform cognitive diagnosis model analysis during the grey-batch validation phase to diagnose issues in small-batch model responses and ensure data quality prior to executing full benchmark synthesis.

When invoking the IRT analysis subskill command, you must pass the skill path, skill name, project root, node ID, input artifact, output artifact, and DAG dependency parameters to execute the diagnostic analysis correctly.