Publication

Read the selected-result record, evidence freeze, formal run manifest, current rule config, and available human adoption facts before drafting. Stop with a science revision when the run identity, metric value/definition/unit, sample set, aggregation, adopted claim, or rule binding conflicts. Display-only changes such as final size, typography, color, marker, linestyle, hatch, legend, and caption may proceed without recomputing a Metric.

Create a short paper_plan.md that states what each section answers, the frozen evidence it uses, the mathematical or visual role, and the conclusion boundary. Write one coherent TeX source rather than concatenating agent fragments. Repeated result values must come from generated/results.tex, not manual transcription.

Use execution-owned data and scripts for data figures. Match final insertion width, provide a non-color encoding, and draw intervals only when the run defines them. When only an original real image exists without data or a script, disclose that it cannot be redrawn. Never present a conceptual illustration as measured data.

Build the electronic target with XeLaTeX, then check citations, placeholders, anonymity, critical numbers, page/size rules, metadata, embedded fonts, and the text layer. Rasterize every page and inspect all of them. Compilation alone leaves visual_review_pending; review_ready requires explicit page-complete visual notes and is not human approval.

Use mathmodel_agent.publication.build_publication(case_root, request) for the deterministic pipeline. See docs/modules/publication.md for the request fields, rule configuration, CLI hook, and exact commands.