content-refinement-agent

Automated peer review for LaTeX drafts runs a structured reviewer across six axes, capturing strengths, weaknesses, questions, and an overall score per iteration. It simulates the review process by applying targeted revisions and deterministic scoring to refine your manuscript iteratively.

Numeric claims in LaTeX drafts are validated against experimental_log.md through strict integrity gates. The refinement process checks citation pool integrity, prevents adding new experiments, and disallows explicit limitation claims to ensure no fabricated results enter the manuscript.

Simulated peer review accepts conference guidelines, a citation pool, draft LaTeX files, and PDF inputs. It applies reviewer-rubric-driven scoring while enforcing anti-reward-hacking rules to ensure the refined draft meets submission requirements without integrity violations.

Reverting LaTeX iterations uses score_delta logic to decide accept or revert actions. The system snapshots each iteration, compiles PDFs, re-scores drafts, and maintains an append-only worklog to identify and promote the best accepted snapshot as the final paper.

Refinement loops prevent adding limitation claims through safety gates that disallow explicit limitation statements. These anti-reward-hacking rules ensure the manuscript maintains integrity by validating claims against experimental logs and preventing unauthorized content additions.