Add website-aligned title and summary columns to all 91 task cards, immediately after task_id. Every original cell is preserved and verified. Keep the default/v1 Viewer configuration and existing repository access.
Configure native SLE leaderboard metrics and freeze published results
Declare Harbor as the evaluation framework and define five unambiguous 0–100 metrics for the 90-task capability scope. Add frozen eight-model values, exact Hub mappings for three API models, metric provenance and links from the dataset card. End-to-end and Hypothesis Discovery remain outside the composite; repository access is unchanged.
Document SLE scope, scoring, published results and dataset access
Expand the dataset card with the 90+1 task scope, the published eight-model snapshot, distinct scoring definitions, repository roles, task-card fields, loading instructions, versioning and research contacts. Preserve existing dataset metadata, licenses, split configuration and contamination canary.
5 days ago
Separate GitHub grading code from gated reference assets
