Volume 2 Operational Glossary

Concise definitions of technical and geopolitical terminology introduced or specialized in this volume. Section A defines frontier systems engineering vocabulary; Section B covers vertical industry mechanics; Section C details geopolitical, regulatory, and ecosystem terms. Keeping them segregated prevents the most frequent error when reviewing this volume: treating a political constraint as if it were a purely technical software design choice.

A. Frontier Engineering

Term	Operational Definition	Chapter
Probabilistic Insider Threat	The security paradigm recognizing that risk does not reside solely outside system boundaries: the foundational model itself can trigger destructive side effects without an external malicious attacker. Invalidates perimeter-only security models.	1
OWASP Top 10 for LLMs	The authoritative industry taxonomy categorizing the ten most critical security vulnerabilities in LLM-based applications. Serves as a risk prioritization framework, not a static compliance checklist.	1
Prompt Injection (Direct / Indirect)	An adversarial instruction delivered via user input (direct) or through external data retrieved by an agent (indirect: poisoned web pages, databases, or MCP tools).	1
Confused Deputy	An architectural failure scenario where a highly privileged component executes actions on behalf of a lower-privileged caller without validating the caller’s actual authorization.	1
Verifiability	The degree to which an agent’s reasoning trace and output can be independently asserted by deterministic code. The core variable added to classical trust calculus.	5
Graduated Trust	The architectural practice of replacing binary trust switches (“trusted / untrusted”) with a continuous operational spectrum calibrated to task risk and verified historical performance.	5
Actor Model / ReactiveX	The formal foundations of reactive systems: isolated state units communicating via asynchronous messages, and declarative operators composing resilient event streams.	4
Backpressure	Flow-control signaling enabling a downstream consumer to throttle an upstream producer rather than buffering unbounded work in memory until failure.	4
Dead Letter Queue (DLQ)	An isolated storage buffer where unprocessable messages are routed after exhausting retry budgets, enabling forensic inspection without halting the live stream.	4
Event Storm	A cascading failure loop where an event triggers reactions that recursively republish identical events. Mitigated by idempotency keys, DLQs, and maximum hop limits.	4
Hybrid Architecture	The deliberate partitioning of workloads between hyperscaler cloud infrastructure and on-premise/edge compute; the primary failure surface is the integration seam between both planes.	12
Feedback Latency	The elapsed time between an agentic action and the arrival of a verifiable feedback signal. Dictates the upper bound on an engineering team’s empirical learning velocity.	26
Agentic Technical Debt	The architectural debt unique to systems built on autonomous agents: unversioned prompts, missing eval suites, implicit ambient permissions, and loops lacking degradation fallbacks.	13
Cognitive Atrophy	The erosion of human domain expertise caused by delegating decision authority entirely to an agent rather than utilizing it as a cognitive exoskeleton.	6
Support-Evaluation Pattern	An architectural separation of concerns delegating helpful assistance and quality judgment to distinct agents, ensuring an assistant never evaluates its own work.	8

B. Industry

Term	Operational Definition	Chapter
Spatiotemporal Diffusion	Generative diffusion model architectures that treat time as a coherent mathematical dimension alongside spatial video coordinates.	16
Multi-Agent Reinforcement Learning (MARL)	Reinforcement learning algorithms where multiple autonomous agents learn policies concurrently within a shared, evolving environment.	28
Proactive Licensing	The strategic legal procurement of intellectual property rights for training data prior to judicial rulings or regulatory mandates.	15, 16

C. Geopolitics, Regulation, and Tech Ecosystems

Term	Operational Definition	Chapter
EU AI Act	Comprehensive European Union legislation classifying AI systems into tiered risk categories and enforcing binding obligations per tier. Regulation operationalized as an architectural blueprint.	19
General Purpose AI (GPAI)	Foundational, high-impact models subject to explicit transparency, systemic risk evaluation, and data governance disclosures under global regulatory regimes.	16, 19
GDPR / PIPL	European and Chinese data privacy frameworks governing permissible cross-border data transfer, retention policies, and architectural data localization constraints.	17, 20
Compliance-by-Design	The engineering discipline of embedding statutory compliance, data residency, and audit boundaries directly into system architecture rather than treating them as post-hoc review items.	19
NIST AI RMF	The National Institute of Standards and Technology’s Artificial Intelligence Risk Management Framework, formalizing risk mapping, measurement, and governance.	18
Constitutional AI / RLHF	Alignment methodologies: explicit machine-readable principles guiding model self-correction (Constitutional AI), and preference optimization via human feedback (RLHF).	18
Mixture of Experts (MoE)	Model architecture dynamically routing inference tokens to a sparse subset of specialized expert parameters, drastically reducing active FLOPs per token.	17
Multi-Head Latent Attention (MLA)	Attention mechanism compressing key-value (KV) cache footprints into low-rank latent projections, allowing far higher concurrent serving throughput.	17
Group Relative Policy Optimization (GRPO)	A mathematical policy optimization algorithm enabling reasoning model training from group rewards without requiring separate critic networks.	17
Technological Sovereignty	A state’s structural capacity to sustain its national AI stack without existential dependence on foreign vendors subject to adversarial extraterritorial jurisdiction.	22, 24
Supply Chain Chokepoint	A concentrated global monopoly node (e.g., ASML extreme ultraviolet photolithography) upon which all downstream advanced semiconductor manufacturing depends.	23, 24
Agentic City Maturity Model	A formal progression scale measuring the degree to which municipal urban operations (traffic, power, emergency logistics) are orchestrated by governed multi-agent swarms.	11