Appendix D. Glossary and Notation

D.1 Core vocabulary

Admission control. A decision about whether to accept additional work given resource, policy, or deadline constraints. It prevents overload from becoming an uncontrolled failure.

Agentic workflow. A workflow in which a model chooses some actions or next steps. Its executor still needs independent authorization, state, and budget controls.

Application invariant. A property that must remain true across valid executions, such as never counting one logical observation twice.

Artifact. A deployable file or package, such as executable code, configuration, or model weights. Artifact identity and mutability affect reproducibility.

Availability. Ability to obtain service under a specified request and failure model. It is not interchangeable with freshness, correctness, or an uptime percentage measured at a different boundary.

Backpressure. A mechanism that limits upstream production or admission when downstream capacity is constrained.

Cache. A retained result or data copy used to avoid repeated work. Correct use requires an identity and validity rule.

Cascade. A policy that attempts one model or service, evaluates the result, and conditionally invokes another.

Cold start. Initialization work before a request can execute on a previously inactive instance or unloaded runtime.

Consensus. Agreement on decisions among participants under a specified failure model. It is not merely several models producing similar text.

Consistency. Rules describing which values distributed reads and writes may observe. The term must be qualified by the intended model.

Container. An execution and packaging abstraction using operating-system isolation mechanisms. It does not eliminate host-kernel, hardware, or resource dependencies.

Control plane. Mechanisms for configuring, deploying, coordinating, and managing a system. Its communication can be separate from application-data traffic.

Critical path. The longest dependency path in a task graph under assigned durations. It limits completion time even when other tasks can run in parallel.

Data plane. Paths and mechanisms that carry or process the application’s working data.

Deadline. The time by which an operation is required or intended to finish. Stopping the caller’s wait does not automatically cancel remote execution.

Deployment. Installing or selecting artifacts and configuration on an execution target. Accepted deployment requests and successful application behavior are different observations.

Digital twin. A model associated with a physical entity and updated according to a defined relationship. Its usefulness depends on fidelity, synchronization, and purpose, not the label alone.

Edge–cloud continuum. A coordinated set of resources across devices, nearby infrastructure, and cloud locations. It is not necessarily a single owner or one flat execution environment.

Event time. The producer’s time associated with an observation or event, distinct from its receipt and processing times.

Fencing token. A version or authority marker that a resource checks to reject operations from a superseded owner.

Freshness. How current information is relative to the time and semantics of a query. High message throughput does not guarantee fresh user-visible state.

Idempotency. A property under which repeating an operation has no additional relevant effect beyond performing it once, according to the operation’s contract.

Inference. Using a trained model to produce an output for an input. It is distinct from updating the model through training.

Jitter. Variation in timing, or intentional randomized delay used to avoid synchronized retries. State which meaning is intended.

Lease. Time-bounded authority or ownership under specified clock and communication assumptions.

Logical event identifier. A stable identifier for the event itself, preserved across delivery attempts.

Offloading. Moving work to another execution location. Its value depends on communication, execution, state, policy, and resource conditions.

Orchestration. Coordinating deployment and execution across components and resources, including desired state and lifecycle management.

Outbox. A durable record of work to send, committed with related local state so a failure cannot silently separate the two.

Placement. Assigning computation or state to a location. Dynamic changes may require migration and warm-up.

Prefill and decoding. Input-context processing and incremental output generation in a language-model serving path. Their resource behavior may differ.

Provenance. Information about the origin and transformations of data or results, used to trace how a claim was produced.

Quorum. A required set or number of participants whose responses permit an operation under a coordination protocol.

Replication. Maintaining copies of state or service across locations. It requires explicit update and failure semantics.

Retained message. A broker-stored topic value available to later subscribers under a messaging protocol. It is not inherently a new physical observation.

Routing. Selecting an execution path or model for a request. In this book, distinguish request routing from network packet routing.

Safety. Avoidance of unacceptable harm under the application’s operating and failure conditions. It is not established by model confidence or network security alone.

Service-level objective. A chosen target for a defined service indicator over a stated population and interval. The target must name the measured boundary.

Snapshot. An evidence set or state view associated with defined selection and consistency semantics.

Speculative decoding. Drafting and verifying tokens to accelerate a target model under a specified algorithm. Correct acceptance and correction rules matter.

Stateful service. A service whose relevant behavior depends on information retained across requests or executions.

Tail latency. The slower portion of a response-time distribution, often described by high percentiles with a defined sample population.

Trust boundary. A boundary across which assumptions about identity, authority, or data reliability change.

Zero trust. An approach that does not grant implicit trust merely from network location or ownership; access is evaluated using explicit policy and context. [25]

D.2 Units and recurring notation

Use decimal units for network rates unless stated otherwise: 1 MB is \(10^{6}\) bytes and 1 Mb is \(10^{6}\) bits. Use binary units when discussing binary memory capacities: 1 MiB is \(2^{20}\) bytes and 1 GiB is \(2^{30}\) bytes. One byte is eight bits. Converting a byte payload using a bit-rate link without that factor creates an eightfold error.

Symbol	Meaning in this book	Typical unit
\(D\) or \(S\)	Data size, with local definition	bytes or MB
\(B\)	Effective throughput, not just nominal link rate	bytes/s or MB/s
\(T\)	Duration, or token count when explicitly defined	seconds or tokens
\(\lambda\)	Request arrival rate in a queue model	requests/s
\(\mu\)	Service rate in a queue model	requests/s
\(\rho\)	Utilization ratio \(\lambda/\mu\) in the teaching model	dimensionless
\(C\)	Cost under a declared accounting boundary	currency or toy cost units
\(E\)	Energy	joules or watt-hours
\(P\)	Power, or probability when explicitly defined	watts or dimensionless
\(M\)	Memory or storage quantity	bytes, MiB, or GiB

Symbols are local to an equation unless a section states otherwise. In particular, the \(T\) in the cache estimate is token count, not elapsed time. A formula is useful only when its units and assumptions are explicit.

D.3 Additional device, networking, and stream-processing terms

ADC. Converts an analog signal into discrete numerical samples. Resolution does not by itself establish measurement accuracy.

IMU. A combination of inertial sensors, typically accelerometers and gyroscopes. Orientation and position estimates require additional assumptions and processing.

GATT. The Bluetooth attribute-oriented procedure framework for discovering and interacting with services and characteristics. Application semantics remain necessary.

6LoWPAN. An IPv6 adaptation approach for constrained links, including header compression and fragmentation mechanisms defined by its specifications.

LoRa / LoRaWAN. LoRa is a radio modulation technology; LoRaWAN specifies a network architecture and protocol behavior built around compatible radio links.

DDS. Data-centric publish–subscribe middleware with typed topics and configurable quality-of-service policies; it is not a physical radio.

Event time / processing time. Event time represents when an event is said to have occurred; processing time represents when a processing system handles it. Clock meaning must be defined.

Watermark. A progress statement used in stream processing to reason about event-time completeness under declared lateness assumptions; not a guarantee that no late event exists.

RTOS. An operating-system design supporting scheduling and timing-oriented services. Deadline satisfaction still requires workload and platform analysis.

Proprioceptive / exteroceptive sensing. Sensing of a platform’s own state versus its surrounding environment; both can contribute to a robot or vehicle state estimate.

Attestation. Evidence about a platform or software state evaluated by a verifier against a policy. Freshness, identity, and the measurement boundary are part of the claim.

Age of information. At an observation time, elapsed time since generation of the newest useful update. It differs from the latency of one successfully delivered packet.

D.4 Agentic application vocabulary

Agentic application. An application in which a model helps choose or revise tasks and tool calls in response to a goal and returned observations. The surrounding software still controls execution and permissions.

Data catalog. Descriptions of available sources, fields, units, intervals, and quality that help determine what can be queried. A catalog description is not the observation history itself.

Task plan. An inspectable statement of intended operations and their dependencies. It is neither evidence that those operations ran nor a record of a model’s hidden reasoning.

Analysis artifact. A data extract, program, table, plot, or report produced during a request. Identifiers help trace artifacts to inputs and execution; they do not grant access.

Answer oracle. An independently established expected result or evaluation rule used to check a task outcome. A deterministic oracle is not a substitute for the model behavior being evaluated.

D.5 AI, digital-twin and trust terminology

Activation. An intermediate value produced while evaluating a model; unlike a weight, it depends on the current input.

Calibration. Estimating a representation range or checking a predictive score against observed outcomes; specify which meaning is intended.

Distillation. Training a student using information supplied by a teacher; it is not merely compressing a file, and the teacher need not run during student inference.

Pruning. Removing or masking selected model parameters or structures; execution savings depend on representation and supported kernels.

Quantization. Mapping numerical values into a lower-precision representation with specified scale, range and rounding behavior.

Post-training quantization. Converting an already trained model, sometimes using representative calibration data, without the full quantization-aware training procedure.

Quantization-aware training. A training procedure that models relevant quantization effects so learned parameters can adapt to the intended representation.

Early exit. Producing an output from an intermediate model branch when an evaluated exit rule accepts it, rather than always using the full network.

Prefill. Processing input context to prepare the state used in subsequent autoregressive generation.

KV cache. Retained attention key/value states that can avoid recomputing preceding context; its size depends on architecture and active sequences.

Retrieval-augmented generation. Using retrieved external evidence in a model request; retrieval does not by itself establish relevance, permission or factual correctness.

Federated learning. Coordinating model learning from locally held data through exchanged updates; distinct from collaborating on one inference request.

Digital-twin prediction. A scenario or estimated future state produced by a behavior model; store its assumptions and provenance separately from observations.

Remote attestation. Evidence about a platform or software state, appraised against a policy; not a universal proof of correct future behavior.

Signed manifest. A protected description of an artifact, its target and permitted update conditions.

Revocation. Withdrawal of a previously granted status or permission; a disconnected cache may not learn it immediately.

Trusted checkpoint. An independently trusted history commitment used to detect some rollback or inconsistency that local self-consistency alone cannot reveal.

Permissioned ledger. A shared-record system in which defined authorities govern who may participate in specified writing, validation or administration roles.

On-chain / off-chain. Distinguishing records maintained through the ledger mechanism from data or operations handled outside it; their linkage still needs design.

Smart contract. Programmed rules for state transitions within a specified platform; correct execution does not validate the truth of external inputs.

Data provenance. Information about where a record came from and how it was transformed; it supports tracing but does not certify physical truth.