Source, evidence and limits for the agentic AI portfolio.
FastAPI and React dashboard, strict prediction contracts, trace scoring, failure taxonomy and regression reports. Scripted controls score 50/100 and 100/100; this measures evaluator behavior, not LLM capability.
Fixed synthetic cases demonstrate role-to-tool allowlists, a separate simulated review gate, minimized hash-chained audit events, and sequential retry with idempotent replay. No free-text patient input, patient data, external adapters, or clinical actions. Reviewer identity is simulated; the audit log is in memory, and concurrent calls are not synchronized.
Fixed skills, explicit workflow states, role allowlists, bounded steps, repeated-step detection and trace export. A single-use local approval seam binds to the exact proposal; production reviewer authentication remains pending.
Agent summaries carry provenance metadata. Cross-scope memory-ID overwrites are rejected, and tests cover session visibility, TTL expiry and durable reopen. Comparable database benchmarks remain pending.
An image-quality research baseline with documented exact-image grouped folds. New split validation rejects missing patient IDs and leakage. Expert labels, patient-disjoint analysis and external clinical validation remain pending.
Genopedia binds lexical retrieval to source URLs, versions, license labels and SHA256 hashes. Neurapedia adds Dice/IoU, calibration and error-count metrics. Current new fixtures test contracts; no diagnostic claim.
Existing supervisor exercised by healthy control and seven faults: sensor dropout, latency, pose dimensions, workspace limits, unauthorized controller, emergency stop and command replay. No hardware or tissue execution.