Grounded Evidence Summary Primer

Language models can be useful for summarizing structured scientific evidence, but they need guardrails. In this project, the summary layer is for evidence synthesis, not discovery by memory.

flowchart LR
    A["Workflow tables"] --> B["Evidence-grounded summary request"]
    C["Dataset accessions"] --> B
    B --> D["Summary template or reviewed model output"]
    D --> E["Evidence card"]
    E --> F["Human review"]

Inputs

• HTS activity and counterscreen status
• Disease expression evidence
• Proteomics support
• Cell-type context
• Protein model and structure feature notes
• Dataset accessions and workflow provenance

Outputs

• concise biological rationale
• limitations and uncertainty
• suggested next validation experiment
• citations to workflow outputs and public accessions

Guardrails

• No uncited biological claims.
• No clinical efficacy claims.
• No company-specific claims.
• Separate supported evidence from hypotheses.

Local Example Mode

The current workflow uses deterministic templates to produce evidence cards. This keeps the project runnable without an API key while documenting exactly how a model-backed summarizer should behave.