| name | analyze-camera-ready-benchmark |
| description | Analyze a camera-ready benchmark campaign for consistency, runtime hotspots, fallback/degraded planners, and reproducibility metadata. |
| 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 |
Analyze Camera-Ready Benchmark
When to use
Use this skill when reviewing *_policy_analysis* or camera-ready campaign outputs and you need
consistency, reproducibility, and fallback-mode visibility before claiming benchmark conclusions.
Read First
docs/code_review.md
docs/benchmark_spec.md
docs/context/issue_691_benchmark_fallback_policy.md
docs/benchmark_camera_ready.md
Workflow
-
Resolve target campaign
- Use the campaign root provided by the user.
- If not provided, ask for a specific campaign id/path.
-
Run analyzer
- Command:
uv run python scripts/tools/analyze_camera_ready_campaign.py --campaign-root <campaign_root>
- This writes:
reports/campaign_analysis.json
reports/campaign_analysis.md
-
Validate core consistency
- Check
campaign_analysis.json findings for:
- episode count mismatches
- summary-vs-episodes metric mismatches
- adapter impact metadata mismatches
- fallback preflight planners
-
Summarize runtime + quality
- Start the report with claim boundary, evidence status, fallback/degraded exclusions, and major
caveats before planner rankings or success interpretation when evidence is mixed or limited.
- Highlight slowest planners by
runtime_sec.
- Highlight weakest planners by success/collision/SNQI from derived episode means.
- Note experimental planners running with fallback mode and make uncertainty visible before any
ranking language.
-
Recommend actions
- If inconsistencies are found, propose concrete follow-ups in issue-facing wording.
- If behavior is expected, explicitly document rationale and residual risk.
Typical Commands
uv run python scripts/tools/analyze_camera_ready_campaign.py \
--campaign-root output/benchmarks/camera_ready/<campaign_id>
Optional custom outputs:
uv run python scripts/tools/analyze_camera_ready_campaign.py \
--campaign-root output/benchmarks/camera_ready/<campaign_id> \
--output-json output/benchmarks/camera_ready/<campaign_id>/reports/custom_analysis.json \
--output-md output/benchmarks/camera_ready/<campaign_id>/reports/custom_analysis.md
Proof and Guardrails
- Classify each run mode as
native, adapter, fallback, or degraded.
- Fail-closed rule: do not treat fallback/degraded as successful benchmark evidence.
- Fallback/degraded rows must appear as caveats or exclusions before comparative rankings,
aggregate success language, or recommendations.
- Mixed/limited evidence outputs must lead with claim boundary, evidence status, and caveats; only
then provide comparative interpretation.
- Require the produced JSON + markdown reports as proof artifacts before summarizing results.
- Only report planner ranking where metrics and episode counts are internally consistent.
Output
Provide a concise judgment with:
- campaign id/path,
- evidence list (
.json + .md),
- consistency status,
- planner ranking by observed-mode,
- concrete remediation steps or explicit pass conditions.
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.