skills/financial-forecast/references/evidence-research.md

# Evidence Research Procedure

Gather evidence for one scoped biopharma question. Return the findings in the format requested by the user, with supporting sources and explicit gaps.

## Your Output Contract

Return structured research notes with:

- the scoped topic you investigated
- findings / claims
- `evidence_rating` for each claim
- `citations` with canonical `source_ref` objects and verbatim `quote`
- explicit `gaps` for anything not supported strongly enough

Use plain Markdown and readable source links. Keep canonical source references with the claims so they can be traced to the documents read.

## Evidence Rules

1. Prefer Maven Bio MCP tools over ad hoc web search.
2. Use document retrieval when a claim needs evidence:
   - `search_documents`
   - `read_document`
   - `get_recent_events`
3. `read_document` must use actual tool-returned document IDs, NCT IDs, or URLs. Do not invent or shorten IDs.
4. Treat discovery tools as scaffolding, not proof:
   - `search_entities`
   - `research_landscape`
   - `fetch_related`
   - `get_financials` (public-market financial statements: revenue, EBITDA, market cap, cash flow, analyst ratings)
   - `get_financings` (private/round-level capital activity: rounds the company received or invested in, by recipient, investor, type, value, date)
   - `get_deals` (cross-company BD transactions: licensing, M&A, R&D collaborations, JVs, by party + role, deal type, value, date)
   - When reasoning about a company's runway or capital position, pull both. `aspects=["financials", "financings"]` on `research_entity` returns them in a single call. On every financing row, `value_status` disambiguates `total_value_usd`: `"unconfirmed"` means Maven's data does not have the amount, not that the round was undisclosed in the world.
   - When a research question touches drug approvals, label content, exclusivity, patents, designations, or LOE timing, anchor the answer in primary sources via `search_documents` with `source_types=["fda_filings"]` and then `read_document`, rather than a summary. Cite the document and its date for any label-sourced claim.
5. For any landscape, pipeline, benchmark, or other master analysis, do not stop after discovery. After you establish the baseline universe, run a document augmentation pass to look for missing entities, recent updates, contradictions, and stronger support for high-impact claims.
6. `Direct Evidence` requires that the source was actually read in this session.
7. If strong evidence is unavailable, mark the claim accordingly instead of stretching the source.

## Fallback Ladders

### If entity resolution fails

- retry with the raw entity name only
- move sponsor/company/disambiguating text into the optional context hint
- use `match_entity` for canonical single-entity lookup
- use `search_entities` only when you need a broader candidate set by criteria
- preserve ambiguity explicitly if the entity still cannot be resolved cleanly

### If a tool times out

- retry once with narrower scope
- reduce batch size
- narrow by source type, entity scope, or claim-specific wording

### If host web fetch is blocked

- stop retrying blocked host fetches repeatedly
- use `search_documents` to locate Maven-indexed or near-equivalent content
- use `read_document` on the returned ID, NCT, or URL
- if that still fails, preserve the gap explicitly

## Suggested Output Shape

```json
{
  "topic": "Example topic",
  "claims": [
    {
      "claim": "Exampledrug reported 93.8% ORR in the named cohort.",
      "evidence_rating": "Direct Evidence",
      "citations": [
        {
          "source_ref": {
            "id": "doc_abc123",
            "kind": "document",
            "tool": "read_document",
            "url": "https://example.com/doc"
          },
          "quote": "NPM1m ORR 93.8%"
        }
      ],
      "notes": []
    }
  ],
  "gaps": []
}
```

## Fail-Closed Behavior

If evidence is missing:

- do not guess
- do not cite entity lookup results as if they were documents
- do not fabricate quotes

Instead, record the unresolved question in `gaps` and explain what was searched.

## Master Analysis Reminder

If your task is to help build a master analysis:

- start with structured MCP tools to define the baseline universe
- then augment with document search and targeted reads
- then reconcile the final claim set

Do not assume the first structured result set is complete enough to finalize the analysis.