Review Lane Guide

Default deep-review lanes:

Section lanes

• section_intro_related
  • check framing, novelty positioning, and promises made early in the paper
• section_methods
  • check definitions, assumptions, derivations, and method detail
• section_results
  • check metric computation, evidence sufficiency, and comparison fairness
• section_discussion_conclusion
  • check interpretation, limitation handling, and claim closure
• section_appendix
  • check whether appendix material supports or contradicts headline claims

Cross-cutting lanes

• claims_vs_evidence — max 8 issues
• notation_and_numeric_consistency — max 10 issues
• evaluation_fairness_and_reproducibility — max 8 issues
• self_standard_consistency — max 6 issues
• prior_art_and_novelty_grounding — max 6 issues
• pre_submission_readiness — max 12 issues (full/editor focus only; populated from high-signal PRESUBMISSION script findings)

Per-lane focus directives, DO/DON'T rules, and grouping conventions live in SUBAGENT_TEMPLATES.md. Output limits prevent LLM filler; recurring issues collapse into one entry with multiple example locations.

Output rule

Every lane must output JSON findings matching ISSUE_SCHEMA.md.

pre_submission_readiness is intentionally narrow. It can contain Critical or Major mechanical issues such as em dashes, repeated AI-tone vocabulary, abstract result gaps, or source hygiene problems, but it must not absorb methodology, theory, literature, or claim-validity reviewer work. When --focus methodology|theory|literature|logic is selected, keep these findings only in Phase 0 automated context.