bench/README.md

Genesis AI Code Bench

Developed by: Within Us AI
Generated: 2026-01-01

A lightweight evaluation harness for Genesis-style datasets that focuses on the signals developers care about in practice:

• Structure validity (JSON parsing, required fields, schema consistency)
• Tool-trace validity (JSON array of tool calls with tool + args)
• Diff validity (patch_diff blocks contain recognizable unified-diff markers)
• Self-grade validity (score bounds, confidence bounds, presence of notes)
• Governance presence (audit/tests flags when expected)
• Economics presence (cost budgets + latency targets)

This bench is intentionally fast and offline-friendly. It does not execute repo tests; it scores dataset quality and readiness for downstream training workflows.

Quick start

python bench.py --jsonl path/to/train.jsonl --max_rows 5000

Metrics produced

• format_valid_rate
• required_fields_rate
• tool_trace_valid_rate
• patch_diff_valid_rate
• self_grade_valid_rate
• governance_present_rate
• economics_present_rate
• uniqueness_rate (hash-based)

Recommended use

• Run before upload to ensure Viewer-ready consistency
• Run after merges to confirm schema stability
• Compare v1.0 vs v1.1 addon impact