content-refinement-agent

To run a simulated peer review on a LaTeX paper draft, the content-refinement-agent applies a structured reviewer rubric to generate feedback and scores across scientific depth, technical execution, logical flow, writing clarity, evidence presentation, and academic style.

Iterative paper refinement can avoid quality regression by using an accept/revert snapshot workflow that compiles each revision, scores it against the reviewer rubric, and immediately reverts changes if the score decreases.

The revision loop prevents hallucinated content by enforcing strict safety constraints during targeted revisions, ensuring no new experiments are added and requiring numeric claims to match ground truth logs exactly.

You should stop the iterative refinement loop when there is nothing actionable left to fix or when improvement plateaus, after which the workflow promotes the best accepted snapshot to the final output.

The best way to improve academic style automatically is using a targeted revision loop that evaluates writing clarity and academic style via a structured reviewer rubric, applying verbatim revision prompts while enforcing anti-hallucination constraints.

The paper refinement workflow has no external dependencies listed, relying solely on its internal scripts and references to compile each LaTeX iteration to PDF and score the results.