ShellBench Research Runbook

Use this skill for a real benchmark campaign, not a one-off local score.

Read references/runbook.md before provisioning machines. It is the normative campaign contract and contains the commands, gates, artifact schema, and recovery rules.

Non-negotiable gates

1. Use remote Crabbox AWS beasts for benchmark execution. Never run scored trials on the operator laptop.
2. Pin one public-task commit, runner commit or patch hash, provider model ID, harness version, reasoning level, and judge route for the whole campaign.
3. Run one r0 qualification for every distinct harness and model-family route, using exactly ten pinned representative tasks. Do not start full-suite jobs until model identity, real traces, tools, usage, judge routing, and artifact export pass.
4. Retain and audit every r0, but force it out of leaderboard scoring. Qualify with independent full-suite repetitions r1 through r3. After a clean audit, add r4 through r6; the research result is six total repetitions.
5. Run every provider-supported non-maximum reasoning level. Never label a reasoning level as tested unless the route applies it and the trace or proxy evidence proves it. Record unsupported levels instead of fabricating them.
6. Use gpt-5.6-sol at high as the default judge. Keep the judge alias, credentials, logs, and identity audit separate from the agent route.
7. Start checkpointing after the first completed trial and continue at least every ten minutes or ten new results. Verify each local archive before it counts.
8. Upload every verified checkpoint and final archive to the private S3 prefix from SHELLBENCH_TRACE_S3_URI. Never put bucket names or credentials in git, PR text, public logs, or generated reports.
9. A run is not research-clean when traces are missing, observed model identity differs from the request, reasoning is unproven, coverage is incomplete, or infrastructure failures dominate.

Required commands

Generate reasoning-specific plans with unique labels:

python -m scripts.native_eval.plan \
  --tasks-root "$TASKS_ROOT" \
  --output "$CAMPAIGN/manifests/run-index-high.json" \
  --public-tasks-commit "$PUBLIC_TASKS_COMMIT" \
  --run-date "$RUN_DATE" \
  --reasoning-effort high \
  --judge-model-id gpt-5.6-sol \
  --judge-reasoning-effort high \
  --repetitions 3

Use repeatable --harness and --model filters for r0 or phased plans. Set --repetitions 6 only after the first three repetitions pass qualification.

Generate each family/harness r0 separately:

python -m scripts.native_eval.plan \
  --tasks-root "$TASKS_ROOT" \
  --output "$CAMPAIGN/manifests/r0-openclaw-gpt56.json" \
  --public-tasks-commit "$PUBLIC_TASKS_COMMIT" \
  --run-date "$RUN_DATE" \
  --phase r0 \
  --qualification-family gpt-5.6 \
  --harness openclaw \
  --model gpt56-sol \
  --reasoning-effort high \
  --judge-model-id gpt-5.6-sol \
  --judge-reasoning-effort high \
  --task "<task-01>" \
  --task "<task-02>" \
  --task "<task-03>" \
  --task "<task-04>" \
  --task "<task-05>" \
  --task "<task-06>" \
  --task "<task-07>" \
  --task "<task-08>" \
  --task "<task-09>" \
  --task "<task-10>"

The planner marks r0 leaderboard_eligible=false with exclusion reason r0_non_scoring_qualification.

After extraction, produce the research tables and strict identity report:

python -m scripts.native_eval.research_audit \
  --run-index "$CAMPAIGN/manifests/run-index.json" \
  --extracted-root "$CAMPAIGN/extracted" \
  --output-dir "$CAMPAIGN/summaries/research"

Treat any false row in model_identity_audit.csv as a blocker. The exporter does not invent prices: missing exact spend remains explicitly unavailable.

Stop conditions

• Stop a route after any r0 contains the wrong or multiple model IDs.
• Stop when the judge's observed provider model cannot be proven.
• Stop when a harness version differs from its campaign pin.
• Preserve and exclude infra-dominated runs, then rerun the same repetition with a suffix and lower concurrency.
• Never delete, overwrite, or silently replace an artifact or repetition.