4-osp-baseline-scout

/4-osp-baseline-scout — Adversarial Baseline Audit

Acts as an adversarial auditor identifying baselines and datasets the authors should have compared against but didn't.

Activation

Invoke the osp-baseline-scout-agent skill.

Inputs — none of these is a gate

• phases.literature.status == "completed". Without a retrieved corpus the Scout searches from scratch, which it is built to do anyway — it just costs more calls and may miss what an earlier round already found. Run it, and say in Provenance that there was no corpus to start from.
• historian is helpful and never required; the Scout works from the structured summary.
• No structured summary either? Read .brain/input/paper.md directly for the task and the baselines claimed, and record that you did.

Record any skip before you start. The orchestrator is not in the loop when the user runs this command directly, so it falls to you: for every earlier phase still pending, set phases.<name>.status = "skipped" and phases.<name>.skip_reason to their reason, or "user ran /4-osp-baseline-scout first". A phase left pending reads as "not reached yet", and the final review cannot tell the difference.

Resource notice

⚠️ Makes ~6-10 API calls to search for SOTA methods and benchmarks. Expect 1-2 minutes, or longer when it reads a paper in full — one full-text read can take up to ~75 seconds on its own.

Opening block (print before step 1)

Render the opening block exactly as .claude/defaults/phase_block_template.md defines it — that file holds the rail, the rules and the widths, and it is the only place they are written down. This is phase 5 of 7 (baseline_scout); read the rail's state from session.json. Values:

  DOING    find baselines and datasets the authors left out
  READS    .brain/raw/01_structured_summary.md
           .brain/raw/02_retrieved_literature.md
  WRITES   .brain/raw/04_missing_baselines.md
  COST     ~2 min, ~6-10 searches

Steps