Auto Paper Research

Purpose

Build research context only. Do not patch .tex, write final sentences, or add claims that are not supported by author evidence.

Required Inputs

• config.yaml
• source_index.md
• tex_inventory.json
• local materials and draft sources named by intake
• ../../../.agents/references/research-supervision-patterns.md
• ../../../.agents/references/research-supervision/paper-writing-layouts.md
• ../../../.agents/references/research-supervision/case-patterns.md
• docs/30_evidence/Experiment_Evidence_Index.{json,md} when present

Workstreams

Scene Analyst reads the current draft, results, figures, tables, notes, repo docs, and the experiment evidence index when present. It outputs author evidence, known claims, known gaps, missing files, and source provenance. It may read iteration_log.json as a weak signal for experiment intent, but must cross-check purpose and result summaries against iteration reports, configs, logs, metrics, or run artifacts before using them for paper claims.

When source PDFs or Markdown notes include figure/table suggestions, extract the proposed visual purpose, evidence source, required data, and uncertainty into research_dossier.md and figure_requirement_scan.md. Treat figure requirements as writing context even when no image asset exists yet.

Exemplar Learner reads reference papers. It outputs section ordering, paragraph moves, style profile, and useful sentence functions. It must not create author claims. Also classify source logic as technical, benchmark/evaluation, or mixed so later phases choose the right paper skeleton.

SOTA Mapper reads bibliography, related-work notes, and optional external search results when the user explicitly requested search. It outputs field map, comparison axes, candidate citations, and unsupported areas.

For review/blog work, SOTA Mapper must produce citation candidates for all named papers, methods, datasets, systems, and quantitative literature claims. If candidates come from an AI dialogue or unverified PDF notes, mark them as unverified candidates instead of omitting them.

Outputs

Write:

• research_dossier.md
• exemplar_learning_dossier.md
• style_profile.md
• sota_gap_map.md

Every fact that may support later claims must carry provenance to local source, TeX, bibliography, or the user brief.

Gate Ledger

Report a Gate ledger entry with commands run, artifacts written, unresolved source gaps, any USER_GATE or NOT_RUN reason, and the next owner before handoff.