v4

Mirrored from benchmarks/graph/v4/README.md. Edit the source document in the repository, not this generated page.

Status: FROZEN snapshot of round 4, frozen 2026-04-25. Do not modify files in this directory except for explicit benchmark-result errata.

v4 is a diagnostic round, not a keep/cull round. v3 settled retention of the agent-facing orbit_graph_* MCP surface; v4 maps where the surface helps, where it hurts, and how it fails — so future tool-shaping work has measured targets.

• Method + pre-registered report shape: METHOD.md
• Report: RESULTS.md
• Fixtures: ./tasks/ (12 NEW fixtures; no v1/v2/v3 carries)
• Synthetic-fixture code island: ./_fixture_code/ (re-included in .orbitignore via narrow negation; see METHOD)
• Frozen run data: ./runs/
• Shared scripts: ../scripts/
• Shared harness overview: ../README.md
• Prior frozen rounds: ../v3/, ../v2/, ../v1/
• Follow-up validation round: ../v5/
• Closing entry on the v1–v3 evidence series: ../../../docs/design/knowledge-graph/5_null_result.md

Re-running a single cell against frozen v4:

GRAPH_VERSION=v4 python3 benchmarks/graph/scripts/run.py \
  --provider codex --arm hybrid --task callers-2hop-graphbenchpolicy --seed 1

See ../../CONVENTIONS.md for version-freeze rules.