Clinical AI safety for the agentic era — deterministic boundaries for probabilistic systems, before and after they act.

MD/PhD — pediatric surgery, then five years in pharmacovigilance and medical affairs (Sanofi, Teva). I now apply that discipline to autonomous AI agents. I call this frame AI Vigilance — pharmacovigilance logic for runtime agent behavior. As AI moved from chat models to autonomous agents, the safety question moved from "what did the model say?" to "what is this agent allowed to do, and on what evidence?" Drug safety has long-established methods for a structurally similar problem: never trust a claim of safety — demand observed evidence, per case, with an audit trail.

• EBAC-T4 — Deterministic Trace-Bound Authorization for High-Risk EHR Writes Authorization gate for one high-risk action: execution authority per evidence cycle, ALLOW / DOWNGRADE / HALT with reason codes, HITL approval bound to exact payloads. Reference artifact, not production software.

• Clinical Verifiable Gates Deterministic verifier kernel: extractor contracts, UNKNOWN-by-default fields, PASS / BLOCK / ABSTAIN with ordered reason codes, evaluation-only Silence oracle. Installable, tested, with expected outputs published for each fixture.

• CASEF — Clinical AI Safety Evaluation Framework Qualification framework for LLM behavior under explicit constraints: spec → measurement → gates → evidence. v0.6.1 provides the contract foundation and validated schemas; the executable qualification pipeline does not exist yet. Generation is cheap. Qualification is not.

• Space Model Lab v3 Case study of governed multi-agent development: a browser-based physics sandbox built by a provider-diverse AI team (architect, PM/auditor, gated single-writer executor) under human PI approval — with a decision log, commit-hash receipts, and a reproducible invariant baseline as the public review surface.

• Modular Reasoning Framework (MRF) Retained as a historical artifact: external reasoning orchestration for fast language models — which I read in retrospect as a predecessor of the governed-orchestration work above.

1. Pre-action authorization & deterministic safety gates — execution authority granted per evidence cycle, not by role or prompt; deterministic ALLOW / DOWNGRADE / HALT and PASS / BLOCK / ABSTAIN gates for state-changing steps; evidence model: traces + constraints + audit artifacts (natural-language claims alone aren't evidence).
2. Qualification under pressure & behavior stability — reproducible crash-tests for role drift; auditable FAIL artifacts; omission-aware evaluation: missed critical actions, not only wrong ones, are treated as a first-class failure mode.
3. Governed human–agent development (PI-led) — provider-diverse specialist roles under one human PI; decision logs, per-action write gates, receipts with observed validation results; CLAIMED vs OBSERVED discipline: an audit without a commit hash is a claim, not an observation. Applied throughout Space Model Lab v3.
4. Architecture selection under hard constraints — decision protocols for LLM-only vs world-model vs hybrid approaches; hard tradeoffs: latency budgets, verification independence, cost-of-error.

• Method sanity checks: framing, assumptions, failure modes; independent cross-checking
• Engineering discipline: prototypes → documented, testable artifacts; stable interfaces; minimal reproducible examples
• Adversarial QA: stress tests for unsafe behavior, hallucinated actions, and format brittleness; definition-of-done gates

Open to open-source collaboration, research discussion, peer review, and unpaid writing/speaking where appropriate.

📫 smorev.research@gmail.com · ORCID 0009-0000-6088-1880 · LinkedIn

All of the above is personal, independent, non-commercial open-source research, done in my own time with my own resources. My current employment is in transportation & logistics (New Brunswick, Canada) and is unrelated to this research. I do not offer commercial or paid services through these projects.