Technology
A conceptual architecture for source-linked, human-approved review.
The architecture below describes how structured case data, approved sources, AI-assisted drafting, explainability, human review and governance fit together.
Architecture
Six layers, from structured case data to governance.
Select any layer to expand what it is responsible for. The stack always ends with a human approval step before anything reaches governance and audit.
Conceptual six-layer architecture
Text alternative for the architecture diagram: Layer 1, Structured Case Layer: Neutral case fields organized for review rather than free-text notes. Layer 2, Approved Source and Citation Layer: A controlled set of approved source categories with citation references attached to guidance. Layer 3, AI-Assisted Draft Layer: Draft review language prepared for pharmacist editing, always labelled as an AI-assisted draft. Layer 4, Explainability and Abstention Layer: Factor views and an explicit abstention path when information is insufficient. Layer 5, Human Review and Approval Layer: A separate approval control operated by the reviewing pharmacist. Layer 6, Governance and Audit Layer: Audit history, model card, data card and feature-status registry.
This architecture is a design in development, shown here for orientation and not deployed as a live system.
Approved source layer
Source categories are placeholders pending corpus approval.
No guideline recommendation is quoted, summarized or invented anywhere in this build.
Canadian dyslipidemia guidance
core reference for statin-related review
Ontario medication-review material
provincial review-practice reference
Health Canada guidance
national regulatory reference
Model Lab
Candidate directions only — no benchmarking has been completed.
Historical Random Forest baseline
Historical exploration record
Recorded as a prior internal exploration only. It does not run inference in the current platform build.
Why considered: Kept as a record of prior internal exploration only, to be transparent about where the current design came from.
Logistic or ordinal benchmark candidate
Interpretable comparison direction
A candidate comparison approach under consideration for future benchmarking work.
Why considered: Considered for its interpretability as a possible future comparison point against other approaches.
XGBoost or CatBoost benchmark candidate
Structured-data comparison direction
A candidate gradient-boosting comparison for future benchmarking work.
Why considered: Considered for handling structured, tabular case fields as a possible future comparison point.
TabPFN benchmark candidate
Tabular foundation-model direction
A candidate tabular foundation-model comparison for future benchmarking work.
Why considered: Considered as an emerging tabular foundation-model direction worth future exploration.
No metric is displayed
Explainability and uncertainty
Illustrative surfaces for reasoning and abstention.
Why this result
- Dose context recorded for reviewSupports review
- Tolerance observations presentSupports review
- Co-medication interaction check pendingRequires attention
- Follow-up interval not documentedRequires attention
Text alternative: synthetic factor emphasis, listed strongest first — Dose context recorded for review, supports review; Tolerance observations present, supports review; Co-medication interaction check pending, requires attention; Follow-up interval not documented, requires attention.
This panel does not reproduce the historical SHAP analysis and does not make a causal claim.
Confidence and abstention
Abstention: insufficient information
When required review fields are missing, the workspace withholds a review summary and asks a qualified professional to collect the missing information first.
Actual calibration and conformal validation are ongoing work.
Governance and audit
Every review leaves a demonstration-grade trail.
A model card, data card, feature-status registry and a session-local audit history sit behind every reviewed case in the guided demonstration.
See it in context