| name | benchmark-overview |
| description | Fast benchmark-faithful orientation for scenario splits, baselines, metrics, artifacts, and reproducibility constraints in robot_sf_ll7. |
| category | benchmark-evidence |
| kind | atomic |
| phase | context |
| requires_write | false |
| requires_slurm | false |
| requires_benchmark_artifacts | true |
| delegates_to | [] |
| output_schema | skill_run_summary.v1 |
Benchmark Overview
When to use
Use this skill when you need a fast, benchmark-faithful read of scenario definitions, seeds, baselines,
metrics, and artifact layout before changing code or producing claims.
Read First
docs/benchmark_spec.md
docs/benchmark.md
docs/benchmark_planner_family_coverage.md
docs/benchmark_camera_ready.md when the task is publication-facing
Workflow
- Identify the active benchmark contract for the request.
- Check scenario/config entrypoints and baseline family boundaries.
- Map required reproducibility artifacts and seeds.
- Return allowed conclusions and known caveats.
Proof and Guardrails
- Verify artifact provenance and contract mapping before interpretation.
- Preserve fail-closed policy where fallback/degraded behavior exists.
- Avoid implying success from runs with incomplete seeds, mismatched episodes, or missing metrics.
- Use issue/context notes to capture non-obvious assumptions.
Output
- benchmark contract and why it applies,
- what evidence supports the interpretation,
- known caveats and non-covered cases.
Guardrails
- Stay within the skill scope declared in
.agents/skills/skills.yaml.
- Prefer repository scripts and canonical docs before ad-hoc commands.
- Record blockers and validation gaps instead of overstating completion.