Medical AI deployment: a glossary of terms.

AMIE

Articulate Medical Intelligence Explorer. Google's research AI system for diagnostic dialogue. Reached 81.7% diagnostic accuracy in a blinded study, against 53.3% for primary care physicians. Not a product - a research demonstration that the capability curve is real. Referenced in The Mountain Top.

ARPA-H

Advanced Research Projects Agency for Health. US federal agency funding high-risk, high-reward health research. Funding the first agentic-AI clinical pilot in the United States - a signal that the federal government sees autonomous clinical AI as a near-term reality, not a distant possibility. Referenced in The Mountain Top.

Autonomous clinical AI

AI systems that can make clinical decisions - diagnose, prescribe, triage - without a human physician reviewing each action in real time. Distinct from clinical decision support, which advises but does not act. The regulatory and liability architecture for autonomous clinical AI barely exists in the US today. See Regulatory Scaffolding.

Base camp

Superposition's term for each product stage on the climb toward the harness. Each base camp is a product that earns its own revenue and solves pressing problems for its own users, while quietly laying a piece of the deployment infrastructure for autonomous medical AI. SignatureAPI is the first base camp. See The Mountain Top.

Clinical integration

The technical and operational work of connecting an AI system to a hospital's existing workflows - EHR systems, PACS, order entry, care coordination. The unglamorous, hospital-by-hospital work that frontier labs will not do. One of the core components of the harness.

Deployment gap

The distance between what medical AI can do on a benchmark and what it is allowed to do at a patient's bedside. The gap is not technical - it is regulatory, legal, and operational. Closing it requires the harness: liability architecture, clinical integrations, and regulatory scaffolding that does not yet exist. See The Harness.

E-E-A-T

Experience, Expertise, Authoritativeness, Trustworthiness. Google's quality rater framework for evaluating content credibility. Especially important in YMYL (Your Money or Your Life) topics like healthcare. Superposition's content strategy is built to satisfy E-E-A-T signals through founder-attributed, clinically grounded essays.

FDA SaMD

Software as a Medical Device. The FDA's regulatory category for software that performs medical functions independently of hardware. Current SaMD frameworks were designed for deterministic algorithms, not stochastic foundation models. The regulatory scaffolding for autonomous AI-as-SaMD is still being built. See Regulatory Scaffolding.

Frontier lab

Organizations training the largest and most capable AI models: Google DeepMind, OpenAI, Anthropic, Meta AI. In the medical context, these labs produce the clinical intelligence - Med-Gemini, AMIE, Med-PaLM - but do not build the deployment infrastructure to put it into hospitals. That is the harness, and a different business entirely.

Harness

Superposition's term for the full deployment infrastructure around autonomous medical AI. Clinical integrations, liability architecture, regulatory compliance, data standards, trust mechanisms - everything the AI needs around it to safely reach a patient. The road the car will eventually drive on. See The Harness.

Incidental finding

A clinically significant abnormality discovered on a medical image that was ordered for a different purpose. Example: a lung nodule found on a chest X-ray ordered for rib pain. Autonomous AI radiology systems will surface incidental findings at a rate human radiologists cannot match, creating new workflow and liability challenges.

Last mile (clinical AI)

The final operational steps between a working AI model and a patient receiving its output - provider trust, EHR integration, clinical workflow redesign, patient consent, and local regulatory compliance. The last mile is where most clinical AI deployments die. It is primarily an operational and trust problem, not a technical one.

Liability architecture

The legal framework determining who is responsible when autonomous AI makes a clinical error. Currently defaults to the supervising clinician under respondeat superior. No US jurisdiction has statutory AI-as-practitioner liability. Designing this architecture is one of the hardest unsolved problems in medical AI deployment. See Liability Architecture.

Med-Gemini

Google DeepMind's medical AI model family. Achieved 91.1% on MedQA, surpassing previous medical AI benchmarks and exceeding the average physician pass rate. Demonstrates multimodal clinical reasoning across text, imaging, and genomics. A research system, not a deployed clinical product. Referenced in The Mountain Top.

MedQA

A benchmark dataset of US Medical Licensing Examination (USMLE) questions used to evaluate medical AI reasoning. The standard yardstick for measuring clinical AI capability. Med-Gemini reached 91.1%; the physician pass threshold is roughly 60%. Benchmarks measure capability but not deployment readiness.

Medical superintelligence

AI that matches and exceeds human doctors in diagnosis, treatment planning, and clinical decision-making across specialties. Not a chatbot or a decision-support tool - a system capable of practicing medicine. Superposition's thesis is that this capability is arriving and the deployment infrastructure is the bottleneck. See The Mountain Top.

MIIT

Ministry of Industry and Information Technology (China). The government body that has driven China's autonomous clinical AI deployment standards. Under MIIT frameworks, over 260 hospitals across 93.5% of Chinese provinces run autonomous clinical AI systems with state-backed institutional liability.

Opportunistic finding

A clinically actionable observation that a human radiologist might note but is not obligated to report, because it falls outside the study's clinical question. Autonomous AI systems do not have selective attention - they will flag everything, creating new workflow demands and liability exposure for findings that previously went quietly unmentioned.

Pre-authorization

The insurance requirement that a provider obtain approval before a procedure or prescription is covered. A major friction point in healthcare workflows. Autonomous AI could handle pre-authorization at machine speed - if the liability and integration infrastructure exists. A target workflow for the harness.

Regulatory scaffolding

The evolving legal and regulatory framework being built around autonomous clinical AI. Includes FDA SaMD pathways, state-level legislation like Utah HB249, federal pilot programs through ARPA-H, and the liability architecture that does not yet exist. The scaffolding is being erected while the building is already going up. See Regulatory Scaffolding.

Respondeat superior

Latin: “let the master answer.” The legal doctrine holding an employer liable for employee actions performed within the scope of employment. Currently, this is how liability flows when AI errs in clinical settings - it falls to the supervising physician. The doctrine was never designed for autonomous non-human agents. See Liability Architecture.

SignatureAPI

Superposition's first product - document infrastructure for healthcare. The first base camp on the climb toward the harness. Solves real document workflow problems for healthcare organizations while building the integration layer and institutional relationships that the harness will eventually require.

Swarm

An architecture pattern where multiple specialized AI agents coordinate to solve a clinical problem - one for imaging, one for lab interpretation, one for drug interactions - rather than a single monolithic model handling everything. Swarm architectures increase the surface area of the harness: more agents mean more integration points, more liability questions, more deployment complexity.

Utah HB249

The first US state legislation authorizing AI copilots to prescribe medication. Passed January 2026. Allows AI systems to prescribe roughly 190 chronic medications under physician supervision. The physician remains the licensee and liability holder. A landmark in regulatory scaffolding for autonomous clinical AI. See Regulatory Scaffolding.

YMYL

Your Money or Your Life. Google's content quality category for pages that could impact a person's health, financial stability, or safety. Healthcare content is the canonical YMYL domain. Search engines apply stricter E-E-A-T standards to YMYL content, which is why Superposition's essays are founder-attributed and clinically sourced.