Skip to content

For AI / ML Engineers

Your models deserve
typed evidence,
not raw records.

OpenBio provides a normalized, governed evidence substrate with 16 assertion types and 6 trust layers. No custom ETL. No document parsing. Just typed evidence, ready to query.

OpenBio Agent Tools
// Typed evidence, filtered by trust layer
await openBio.searchEvidence('james-001', {
  trustLayer: 'source_fact'
})  
// → 142 assertions · all provenance-tagged
// Submit annotation — never overwrites source facts
await openBio.submitAnnotation('assertion-789', {
  type: 'hypothesis', confidence: 0.72
})  
// → source_fact unchanged · provenance preserved

The ETL Tax

Every AI team building on clinical data writes the same pipeline.

  • Inconsistent schemas across source systems

    FHIR, HL7, CSV, VCF — each requires custom parsing. Your models can’t be trained on data that doesn’t share a schema.

  • No trust metadata

    Your model can’t distinguish a source fact from a model output. It treats an AI prediction with the same confidence as a lab result.

  • Overwrites destroy provenance

    When annotations overwrite source data, you lose the audit trail. Six months later, no one knows what the original reading was.

Normalized. Typed. Governed.

The evidence substrate your models have been waiting for.

16 Assertion Types

diagnosis, lab_result, variant_call, imaging_finding, medication, procedure, and 10 more. Typed schemas. No ambiguity.

6 Trust Layers

source_fact through hypothesis. Your model always knows the epistemic status of its inputs — never confuses prediction with observation.

MCP Compatible

Query the evidence substrate via the Model Context Protocol. Works with any agent framework. Structured tools, typed responses.

No ETL Required

Ingest once, query typed assertions forever. OpenBio handles normalization, deduplication, and provenance tracking.

Every assertion carries its
epistemic status.

Source Fact

Raw data directly from a source system. Highest confidence.

Normalized

Standardized and mapped to canonical terminology.

Extracted

Derived from unstructured documents or images via NLP/CV.

Derived

Computed from other evidence assertions.

Model Output

AI/ML prediction. Confidence score required.

Hypothesis

Proposed interpretation. Under review.

Agent Access

Three tools. Every assertion in the substrate.

OpenBio exposes the evidence substrate through a tight MCP-compatible tool surface. Search with filters, submit annotations that layer above source data, and build cross-species cohorts — all without touching raw records.

  • searchEvidence()
  • submitAnnotation()
  • buildCohort()
OpenBio Agent Tools
// Typed evidence, filtered by trust layer
await openBio.searchEvidence('james-001', {
 assertionTypes: ['variant_call', 'lab_result'],
 trustLayer: 'source_fact'
})
// → 142 assertions · all provenance-tagged
// Submit annotation — never overwrites source facts
await openBio.submitAnnotation('assertion-789', {
 type: 'hypothesis',
 confidence: 0.72,
 rationale: 'Consistent with EGFR pathway'
})
// → source_fact unchanged · provenance preserved
// Build a cohort across species
await openBio.buildCohort({
 criteria: { assertion: 'EGFR_exon19_del'},
 subjectTypes: ['human', 'primate']
})
// → CohortResult · 3 subjects matched

Query James\u2019s 847 assertions.