Trust & factuality bundle: RAG grounding, uncertainty + error-rate display, compliance docs

[Liohtml]

Part of #51. Three closely-related trust/factuality improvements grouped as one
tracking issue (can be split into sub-PRs).

---

D. Compliance & positioning documentation

Document the intended-use boundary so contributors and users understand what
MedCheck is and isn't — and so the project stays on the safe side of the
medical-device line (see epic #51).

Research basis (verified against primary sources):

• US: a tool that analyzes a medical image fails FDA Cures-Act CDS Criterion 1,
  and a patient-facing tool fails Criterion 3 → it is, on its face, a regulated
  SaMD. The CDS exemption is categorically unavailable to image-analysis software.
• EU: MDR Rule 11 / MDCG 2019-11 — diagnostic-information software is ≥ Class
  IIa, imaging typically IIb; no CDS carve-out.
• A "not a medical device / educational only" disclaimer does not by itself
  exempt software — intended use (claims, labeling, functionality) governs.

Deliverables:

• docs/intended-use.md: explicit intended use + dual positioning (research/
  developer track vs. patient-education track), and the "do / don't" boundary
  (no autonomous image findings presented to patients as truth; explain
  existing reports + terminology + question prep).
• A model card (docs/model-card.md): models used, data provenance, known
  limitations, hallucination risk, no validation claims.
• Map MedCheck against FUTURE-AI principles (Fairness, Universality,
  Traceability, Usability, Robustness, Explainability) and note gaps.
• Strengthen the README/disclaimer wording accordingly.

---

E. RAG grounding for factuality (Tier 0 — no training)

Reduce Vision-LLM hallucination by retrieving relevant reference material (curated
finding descriptions, anatomy templates, normal-vs-abnormal exemplars) and
grounding the prompt — no model training required. Research shows training-free
RAG (LaB-RAG, RadioRAG) is competitive and is the best risk/reward first step for a
small team; even SOTA report-generation models hallucinate, so grounding + human
review is mandatory.

Deliverables:

• Optional retrieval over a curated, license-clean reference corpus (e.g. the
  anatomy templates already in prompts/anatomy/, expandable).
• Inject retrieved context into vision_analysis.build_prompt(); cite which
  references were used in the output (traceability).
• Config to enable/disable; lazy deps for any vector store (or a simple in-repo
  embedding/BM25 to avoid heavy deps initially).
• Tests: retrieval returns expected refs; prompt includes grounding; disabled = no-op.

---

F. Uncertainty + error-rate display

The evidence-based mitigation for patient harm (false reassurance/alarm) is to
always pair any AI output with its uncertainty and error-rate context. A
controlled study (npj Digital Medicine 2025) found disclosing error rates
measurably reduced harm from patient-facing AI imaging output.

Deliverables:

• Surface per-finding confidence prominently and consistently in all report
  renderers (already on StructureFinding).
• Add standard, honest limitations + error-rate context blocks to reports
  ("AI image analysis can miss findings or flag things that aren't there; this
  is not a diagnosis").
• In the patient report mode (Add patient-friendly plain-language report mode #56), present uncertainty in plain language and
  never imply certainty the model doesn't have.
• Feed confidence/limitations into the validator (Add findings cross-validation pipeline step (LLM findings vs. ML signals) #52) and reconciliation (Add report reconciliation: compare official radiologist report against image analysis #58)
  framing.

---

Acceptance criteria (bundle)

• D: intended-use + model-card docs merged; README disclaimer strengthened.
• E: optional RAG grounding wired into prompts with citation/traceability.
• F: uncertainty + error-rate context shown consistently across report modes.
• Each sub-item independently testable; all additive (defaults unchanged).

Effort

D ~0.5 day · E ~1–2 days · F ~0.5 day. Light deps.

References (session research)

FDA CDS Final Guidance (Criteria 1 & 3); EU MDR Rule 11 / MDCG 2019-11; FDA General
Wellness guidance; FUTURE-AI (BMJ 2025); WHO AI-for-health ethics (2021/2024); npj
Digital Medicine 2025 (error-rate disclosure); LaB-RAG / RadioRAG.

[Liohtml]

Assessment

Value: High — but this is a large umbrella feature that bundles several distinct capabilities. Worth breaking apart.

The three main requests:

1. RAG grounding — tethering LLM responses to a curated radiology knowledge base
2. Uncertainty + error-rate display — showing confidence scores and model limitations in reports
3. Compliance docs — HIPAA/GDPR/MDR documentation artifacts

My take on each:

1. RAG grounding — Significant complexity and uncertain ROI for radiology-specific tasks. The LLMs supported (Claude Opus, GPT, Gemini) are already trained on extensive medical literature. A domain-specific vector store could help with rare conditions or institutional protocols, but building and maintaining it correctly is substantial work. I'd hold this for a later milestone.

2. Uncertainty + error-rate display — This is already partially implemented! StructureFinding.confidence (0.0–1.0) is populated by every LLM provider and included in the JSON output. The gap is surfacing this in PDF/HTML reports and adding population-level calibration data ("this model has X% sensitivity for ACL tears"). The second part requires clinical validation data we don't have yet. The first part (surfacing confidence in report templates) is a quick win.

3. Compliance docs — Relatively straightforward. The docs/intended-use.md and docs/model-card.md are already started. EU MDR Article 13 information, HIPAA BAA template language, and a "limitations" section in the model card are the main deliverables. This doesn't require code changes — mostly documentation work.

How I would prioritise

1. Quick win (1–2 days): Surface StructureFinding.confidence prominently in PDF/HTML report templates. Add a "Confidence Legend" section to reports.
2. Medium effort (1 week): Expand docs/model-card.md with known limitations, population-level caveats, and intended-use restrictions for compliance.
3. Defer: RAG grounding — only worth it if there's a clear use case (e.g., institutional protocol adherence) that the base models demonstrably fail at.

Concerns

The "error rate display" risks giving false confidence if it's not backed by rigorous clinical validation on a real patient cohort. Showing "87% sensitivity for meniscal tears" without a cited study is worse than saying nothing. Be careful about generating authoritative-sounding statistics that aren't actually validated.