# clinical-ai-validation-harness

> A validation harness for clinical AI on OpenMRS and OpenELIS — specs, plans, and visual canvases, mirrored here as full static HTML for LLM agents.

## Docs
- [Why local-first clinical AI](/clinical-ai-validation-harness/spec/specs/background/why-local-first-clinical-ai.html): The cited evidence behind the mission — offline realities, data sovereignty, right-sized open models, and WHO SMART Guidelines.
- [Project README](/clinical-ai-validation-harness/spec/README.html): What this harness is, who it is for, how to get started, and key terms.
- [Background & evidence (research)](/clinical-ai-validation-harness/spec/specs/artifacts/planning/global-health-ai-background-research-2026-06-14.html): Cited global-health grounding: WHO SMART Guidelines, LMIC realities, data sovereignty, open-model right-sizing.
- [Guardrails methodology (research)](/clinical-ai-validation-harness/spec/specs/artifacts/planning/guardrails-methodology-research.html): Prompt-injection and unsafe-answer defenses — the safety research behind the harness.

## Canvases
- [Validation roadmap](/clinical-ai-validation-harness/canvas/specs/roadmap.html): Milestones, lanes, and dependencies — start here to understand sequencing.
- [Validation research](/clinical-ai-validation-harness/canvas/specs/artifacts/canvases/validation-research.html): Evidence model, evaluation methodology, and the run-manifest traceability spine.
- [Demo-data profile & cohorts](/clinical-ai-validation-harness/canvas/specs/artifacts/canvases/demo-data-profile.html): The loaded OpenMRS 2.8 demo corpus: landscape, completeness, and validation cohorts.
- [Clinical-AI research guidance](/clinical-ai-validation-harness/canvas/specs/artifacts/canvases/clinical-ai-research-guidance.html): Research vectors, evidence levels, and maturity framing.
- [Scout comparative analysis](/clinical-ai-validation-harness/canvas/specs/artifacts/canvases/scout-comparative-analysis.html): Duke DIHI Scout and what it implies for chartsearchai, openmrs_chatbot, and Catalyst.
- [Catalyst demos](/clinical-ai-validation-harness/canvas/specs/artifacts/canvases/catalyst-demos.html): Query-to-table on two real data sources — watch a two-turn conversation on each.
- [Concept mapping & transformation](/clinical-ai-validation-harness/canvas/specs/artifacts/canvases/concept-mapping-discovery.html): Bridge rule, promotion rules, blockers, open decisions.
- [SQLMesh transformation flow](/clinical-ai-validation-harness/canvas/specs/artifacts/canvases/sqlmesh-transformation-flow.html): How the deterministic OpenMRS 2.7 → 2.8 transform is materialized.
- [chartsearchai & querystore](/clinical-ai-validation-harness/canvas/specs/artifacts/canvases/chartsearchai-and-querystore.html): Architecture of the chart-search and query-retrieval integration.
- [Answer / In-Depth parity](/clinical-ai-validation-harness/canvas/specs/artifacts/canvases/answer-indepth-parity.html): Evolving Answer and In-Depth into two truly separate, independently-measured responses — current architecture (both setups) and the roadmap.
- [Answer flow — what to simplify](/clinical-ai-validation-harness/canvas/specs/artifacts/canvases/answer-flow.html): med-agent-hub answer-generation flow and where to cut complexity.
- [Cross-project comparison](/clinical-ai-validation-harness/canvas/specs/artifacts/canvases/cross-project-comparison.html): Side-by-side architecture of chartsearchai, openmrs_chatbot, and Catalyst.
- [Upstream contribution & compatibility](/clinical-ai-validation-harness/canvas/specs/artifacts/canvases/upstream-contribution-and-compatibility.html): The ChartSearchAI relay and med-agent-hub profile integration organized into reviewable OpenMRS contributions.
- [Validator audit framework](/clinical-ai-validation-harness/canvas/specs/artifacts/canvases/validator-audit-framework.html): How validator behavior is audited and kept reviewable.

## Topics
- [The data we test against](/clinical-ai-validation-harness/topic/data-corpus.html): A realistic 5,284-patient OpenMRS demo corpus, modernized from the public 2.7 demo and bound to the CIEL clinical terminology — so validation runs against believable charts, not toy fixtures.
- [Evidence, evaluation & traceability](/clinical-ai-validation-harness/topic/evidence.html): How an AI answer is judged: scored against the patient chart, every claim traceable to a specific record, recorded on a run-manifest provenance spine — directional evidence, not a leaderboard.
- [Safety & governance](/clinical-ai-validation-harness/topic/safety-governance.html): Guarding against unsafe answers, and the change-control discipline (validator audits, PCCP-style records) that keeps validation baselines reviewable.
- [Catalyst — query-to-table for lab and program data](/clinical-ai-validation-harness/topic/lab-ai.html): Ask a clinical question in plain language; a writer/reviewer model team drafts governed SQL against a database-generated catalog, a deterministic policy enforces read-only execution, and the result is a typed table — running today on two independent data sources, OpenELIS lab data and OpenMRS HIV/ART program data.
- [Upstream & compatibility](/clinical-ai-validation-harness/topic/upstream.html): How the ChartSearchAI relay, med-agent-hub profiles, and optional Querystore source are organized into reviewable OpenMRS upstream contributions.