Update to Maven Bio

{
  "description": "Turn already-discovered facts into evidence-backed claims with evidence ratings, verbatim quotes, and canonical source references. Use as the final evidence pass before presenting findings, or when a claim must be tied back to a specific filing. This is a building block the workflow skills compose.",
  "included_files": [
    {
      "relative_path": "LICENSE",
      "size_in_bytes": 802
    },
    {
      "relative_path": "agents/openai.yaml",
      "size_in_bytes": 335
    }
  ],
  "name": "synthesize-evidence",
  "skill_md_contents": "---\nname: synthesize-evidence\ndescription: \"Turn already-discovered facts into evidence-backed claims with evidence ratings, verbatim quotes, and canonical source references. Use as the final evidence pass before presenting findings, or when a claim must be tied back to a specific filing. This is a building block the workflow skills compose.\"\n---\n\n# Synthesize Evidence\n\n## Using this skill\n\nUse the connected Maven Bio MCP server at `https://mcp.mavenbio.com/`. Follow the user's explicit scope, depth, and output preferences; the workflow and output structure below are defaults. Report coverage limits instead of silently narrowing an explicitly requested set.\n\nHyphenated primitive names refer to other skills in this Maven Bio bundle. Consult the relevant skill when composing its workflow. Use the available MCP tool schemas for arguments; pass document identifiers to `read_document` through `ids`, and include a claim-specific `query` when using `format=\"citations\"`.\n\nThis primitive is the bridge from research findings to defensible claims.\n\n## Use When\n\n- a claim needs stronger support\n- a workflow has facts but not enough provenance\n- you need to downgrade or tighten a statement based on what the documents actually say\n- a narrative depends on quantitative claims such as incidence, prevalence, utilization, outcomes, cost, treatment windows, response rates, or safety rates\n\n## Core Tools\n\n- `search_documents`\n- `read_document`\n- `get_recent_events`\n- `search_documents` with `source_types=[\"fda_filings\"]` (when claim provenance is a regulator label)\n\n## Output Contract\n\nReturn structured claim objects:\n\n- `claim`\n- `evidence_rating`\n- `citations`\n- `notes`\n- `gaps`\n\nExample citation object:\n\n```json\n{\n  \"source_ref\": {\n    \"id\": \"doc_abc123\",\n    \"kind\": \"document\",\n    \"tool\": \"read_document\",\n    \"url\": \"https://example.com/doc\"\n  },\n  \"quote\": \"verbatim supporting text\"\n}\n```\n\n## Numeric Claim Audit Mode\n\nUse this mode when the output needs quantified disease burden, treatment utilization, standard-of-care rates, clinical outcomes, safety rates, pricing, market size, or other numeric claims.\n\nBefore writing prose:\n\n1. Extract a checklist of required numeric claims from the user request.\n2. For each claim, run targeted `search_documents` queries rather than one broad topic search.\n3. Read the strongest source candidates with `read_document`, usually `format=\"sections\"` or `format=\"citations\"`.\n4. Return a working claim table with:\n   - `claim_needed`\n   - `value_found`\n   - `source_ref`\n   - `quote`\n   - `evidence_rating`\n   - `status`: `found`, `adjacent`, `conflicting`, or `missing`\n   - `notes`\n5. Only synthesize numbers marked `found` or clearly `adjacent`; preserve `conflicting` and `missing` items as gaps.\n\nDo not substitute a nearby number silently. For example, if the user asks for AIS share of all strokes and the source only supports global ischemic-stroke share, record the mismatch rather than treating it as the requested number.\n\n## Quality Bar\n\n- use verbatim supporting quotes\n- prefer primary or authoritative documents when available\n- use the actual document ID, NCT ID, or URL returned by the tool result; do not invent plugin-local aliases\n- for regulatory claims (approval status, label content, exclusivity, designations), cite the document ID (`doc_...`) plus its date. Direct quotes from the label text are preferred over paraphrase: locate the filing with `search_documents(source_types=[\"fda_filings\"])`, use `read_document` with `format=\"sections\"` to find the relevant section, then pull `format=\"text\"` narrowly for the passage you will quote\n- if host web fetch is blocked, treat `search_documents` plus `read_document` as the standard fallback path\n- downgrade evidence strength when the support is indirect\n- if the support is missing, record the gap instead of forcing a claim\n- do not write final narrative prose until high-signal numeric claims have been audited into structured claim objects\n"
}