Auto Paper Harden

Purpose

Audit manuscript readiness. Do not do large rewrites in harden; route failures to the owning phase.

Required Inputs

• R/A/C/L/P phase artifacts
• patched TeX files
• bibliography files
• guard or compile reports when present
• ../../../.agents/references/research-supervision-patterns.md
• ../../../.agents/references/research-supervision/pre-submission-review.md
• ../../../.agents/references/research-supervision/scientific-plotting.md
• docs/30_evidence/Experiment_Evidence_Index.{json,md} when paper claims use experiment evidence

Audits

Artifact completeness checks required files, non-empty content, identifiers, and cross-references.

Logic transfer compares original_logic_map.md, writing_rationale_matrix.md, and revised TeX to detect lost logic or shallow substitution.

Claim support checks every claim against author evidence or citation_support_bank.md, including over-strong verbs and missing boundaries.

For blogs, reviews, surveys, and tutorials, citation audit must fail or route back when named literature, clinical, economic, benchmark, dataset, or method claims appear without citation support rows. A disclosure such as "source PDF was not verified" is acceptable only as a boundary, not as a replacement for a verified-reference plan.

Figure audit checks figure_requirement_scan.md, figure_contract.md, caption_claim_map.md, and existing figure references. If the source material asks for figures/tables but no figure contract exists, route to REWORK_LAYOUT or $auto-paper-figure instead of declaring COMPLETE.

LaTeX and compile audit runs the static guard and configured compile command when available.

Reviewer-risk audit covers motivation, novelty, evidence, baseline fairness, method clarity, result interpretation, limitations, reproducibility, ethics/data statements, and template compliance. Also run the five-dimension pre-submission lens: macro logic, writing detail, grammar/wording, format, and figure quality.

If a claim is blocked by missing or weak experiment evidence, write a concrete request to run_request_register.{json,md} and return RUN_REQUEST. The request must identify the blocking claim, needed evidence, experiment type, minimum run artifacts, suggested $run prompt, and acceptance check. Do not silently weaken or remove a central claim unless the operator approves that tradeoff.

Use a strict analysis bar: if seed counts, raw metrics, statistical comparisons, figure provenance, or comparison units are missing, mark the claim as insufficiently supported and request the missing evidence instead of polishing it into manuscript prose.

Use deterministic scripts where possible:

• .agents/skills/auto-paper/scripts/artifact_check.py
• .agents/skills/auto-paper/scripts/integrity_audit.py
• .agents/skills/auto-paper/scripts/citation_bank_check.py
• .agents/skills/auto-paper/scripts/claim_register_check.py
• .agents/skills/auto-paper/scripts/figure_requirement_scan.py
• .agents/skills/auto-paper/scripts/revision_audit.py
• .agents/skills/auto-paper/scripts/style_metrics.py
• .agents/skills/auto-paper/scripts/latex_guard.py

Outputs

Write:

• audit_report.md
• compile_report.md
• citation_audit_report.md
• revision_audit_report.md
• logic_transfer_audit.md
• final_gate_ledger.md
• run_request_register.{json,md} when $run needs to supply missing evidence

Final decision must be one of COMPLETE, RUN_REQUEST, USER_GATE, REWORK_RESEARCH, REWORK_ARGUMENT, REWORK_CITATION, REWORK_LAYOUT, REWORK_PATCH, or ABORT.