--- name: synthesize-evidence 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." --- # Synthesize Evidence This primitive is the bridge from research findings to defensible claims. ## Use When - a claim needs stronger support - a workflow has facts but not enough provenance - you need to downgrade or tighten a statement based on what the documents actually say - a narrative depends on quantitative claims such as incidence, prevalence, utilization, outcomes, cost, treatment windows, response rates, or safety rates ## Core Tools - `search_documents` - `read_document` - `get_recent_events` - `search_documents` with `source_types=["fda_filings"]` (when claim provenance is a regulator label) ## Output Contract Return structured claim objects: - `claim` - `evidence_rating` - `citations` - `notes` - `gaps` Example citation object: ```json { "source_ref": { "id": "doc_abc123", "kind": "document", "tool": "read_document", "url": "https://example.com/doc" }, "quote": "verbatim supporting text" } ``` ## Numeric Claim Audit Mode Use 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. Before writing prose: 1. Extract a checklist of required numeric claims from the user request. 2. For each claim, run targeted `search_documents` queries rather than one broad topic search. 3. Read the strongest source candidates with `read_document`, usually `format="sections"` or `format="citations"`. 4. Return a working claim table with: - `claim_needed` - `value_found` - `source_ref` - `quote` - `evidence_rating` - `status`: `found`, `adjacent`, `conflicting`, or `missing` - `notes` 5. Only synthesize numbers marked `found` or clearly `adjacent`; preserve `conflicting` and `missing` items as gaps. Do 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. ## Quality Bar - use verbatim supporting quotes - prefer primary or authoritative documents when available - use the actual document ID, NCT ID, or URL returned by the tool result; do not invent plugin-local aliases - 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 - if host web fetch is blocked, treat `search_documents` plus `read_document` as the standard fallback path - downgrade evidence strength when the support is indirect - if the support is missing, record the gap instead of forcing a claim - do not write final narrative prose until high-signal numeric claims have been audited into structured claim objects