Research Agent Bot is an audited biomedical evidence-synthesis engine. It is not a generic drafting demo: retrieval, source admission, claim extraction, manuscript compilation, and publish gates are separate stages with sidecar artifacts for review.
topic pack
-> source retrieval and corpus seeding
-> source admission and evidence-tier classification
-> claim cards, citation registry, and tension matrix
-> v3 synthesis compiler
-> public manuscript plus supplemental evidence tables
-> deterministic surface, audit, and pre-submit gates
-> Researka submission
full_paper.md- reader-facing manuscript. It should contain prose, compact evidence snapshots, references, and no raw compiler dumps.structured_evidence_tables.md- supplemental evidence tables and full evidence-map detail.manifest.json,claim_graph.json,citation_registry.json, and audit sidecars - machine-readable trust-spine artifacts.paper_ir.json- typed PaperIR containing thesis, sections, framework axes, tables, references, and export pointers.paper_quality_score.jsonandpublic_export_manifest.json- deterministic reader/export readiness summaries.references.bib,evidence_table.csv,contradiction_map.json, andfull_paper.docx- portable reader and publisher exports.- Optional polish outputs - Typst PDF, sciwrite-lint report, and offline eval JSON when those tools are installed.
LLM PROPOSES. CODE DISPOSES.
LLMs may draft and revise prose. Deterministic code owns evidence admission, categorical labels, source-traced numerics, maturity status, and submit eligibility. Markdown is a rendering target, not the source of truth.
.venv/bin/python -m ruff check agent scripts tests
.venv/bin/python -m mypy --no-incremental --explicit-package-bases --ignore-missing-imports scripts/run_v06_synthesis.py scripts/table_renderer.py agent/journal_surface_gate.py
.venv/bin/python -m pytest -qThe live bot runs from /opt/research-agent-bot on the VPS. Deploy only from a
clean tree, fast-forward both VPS checkouts, then verify the service and HTTP
status endpoint.