Back to ARL

ARL Lexicon

Version 1.3 — May 2026 Companion to: ARL v1.3, ARL-S v1.3

A controlled vocabulary for AI readiness measurement. Each term has one operational definition. Terms that do not yet have a single operational definition are included, with that fact stated plainly and neutrally. ARL is a measurement framework: a term can anchor an ARL claim only when it can be measured. Where a term cannot be measured today, the lexicon says so and takes no position on whether the term is meaningful, real, or imminent — those are open questions on which a measurement standard has no opinion. The lexicon is intended to be time-invariant — definitions are built from physical quantities and information-theoretic foundations, not from cultural usage.

Science is the discipline of defining terms so that claims survive translation across time, languages, and people. This document is the foundation underneath ARL and ARL-S.

A

Accelerator. A hardware device specialized for parallel arithmetic operations on tensors, distinct from a general-purpose CPU. Examples include GPUs (graphics processing units repurposed for general compute), TPUs (tensor processing units), NPUs (neural processing units), and various custom inference silicon. Identified by model, manufacturer, memory configuration, supported precisions, and thermal design power.

Accuracy. In a measurement context, the closeness of a measured value to the true value. Distinguished from precision (repeatability). A measurement can be precise but inaccurate or accurate but imprecise. In an AI benchmark context, accuracy commonly refers to the fraction of test items for which the model output matches a reference answer; this usage is benchmark-specific and not interchangeable with the metrological definition.

Adversarial input. An input crafted to elicit a specific failure mode in a system. Distinct from natural distribution shift. Documented adversarial failure modes are required for ARL 4 and above.

Adversarial robustness. Measured resistance of a system to adversarial inputs. Quantified by attack success rate under a documented attack corpus, with N ≥ 100 attempts per category. A component of Security Class S1 and above.

Auditability. The property that every system output can be traced to the inputs, tool calls, model version, harness configuration, and operator identity that produced it. Measured by audit trail completeness against an adversary attempting to make actions un-traceable. A component of Security Class S4.

Agent. A software system that takes actions in an environment over multiple steps, where outputs of earlier steps influence inputs of later steps. Operationally defined; does not imply intentionality, autonomy in any philosophical sense, or goal-directedness beyond what is implemented in the code. Compare agentic system.

Agentic system. An AI system configured as an agent. The configuration includes a model, a harness, available tools, and termination conditions. Agentic system behavior is the joint behavior of these components, not the behavior of the model alone.

AGI (Artificial General Intelligence). No single operational definition (as of this writing). The term is defined differently across speakers and its meaning has shifted over time; there is active, good-faith disagreement about what would qualify, and several operationalization efforts are underway. Because no single measurable definition is settled, the term cannot anchor an ARL claim. ARL takes no position on whether artificial general intelligence is achievable, imminent, or already present — it simply cannot score a claim against a target that is not operationally defined. Capability claims under ARL are scope-locked to specific tasks, hardware, and contexts.

Alignment. Measurable in its operational sense. Operationally, refers to techniques that adjust model outputs toward specified human preferences (RLHF, DPO, Constitutional AI, etc.); used this way, the effect is measurable. In its broader sense — ensuring AI systems “do what humans want” in general — it is not measurable, because “what humans want” is not a defined quantity, so that sense cannot anchor an ARL claim. ARL Convergence Class measures behavioral consistency under operational variation, the closest measurable proxy. ARL takes no position on the broader research program; it scores only what can be measured.

Amortization (energy). The allocation of one-time costs (training) across deployment lifetime. Stated as MWh per deployment year. Required component of ARL Energy Profile.

Attestation. A signed record produced by the ARL-S Supervisor at the conclusion of an evaluation session, containing the telemetry, the hardware identification, the harness identification, the SUT identification, and the resulting ARL score. Signed with a hardware-backed key unreachable by the SUT or harness operator.

B

Backpropagation. The algorithm by which gradients of a loss function are computed with respect to model parameters via the chain rule. Operationally defined; one of the oldest stable terms in the lexicon. The technique was independently discovered multiple times beginning in the 1960s.

Batch. A group of inputs processed together in one forward pass for computational efficiency. Batch size affects throughput and per-token energy; specified in the inference stack description for any ARL claim above level 4.

Bandwidth (memory). The rate at which data can be moved between memory and compute, measured in bytes per second. A frequently binding constraint for inference performance independent of FLOPS capacity.

Benchmark. A defined set of test items with associated scoring methodology, used to measure AI system performance on a specific task class. In ARL terminology, benchmarks are tasks that can plug into ARL-S as Harnesses; they are not the framework.

Bit. The fundamental unit of information. One binary digit. The foundation of every higher-order term in this lexicon.

Byte. Eight bits. The standard unit of memory addressing.

C

Capability. Operationally, the demonstrated ability of a system to perform a specific task at a specific level of success, on specific hardware, in a specific context. Capability claims are valid only for the (system, task, hardware, context) tuple in which they were demonstrated. Generalization to other tuples requires re-measurement. Compare performance.

Capability emergence. Measurable in its operational sense. Operationally, refers to capability measurements that change non-monotonically as scale changes — a measurable, reportable observation. Broader interpretive uses (for example, as evidence about general intelligence) are not measurable and cannot anchor an ARL claim. ARL reports the measured behavior and takes no position on its interpretation.

CPU. Central processing unit. General-purpose processor distinguished from accelerators. RAPL telemetry for CPU energy is accessed through /sys/class/powercap/intel-rapl/ on Intel and /sys/devices/platform/amd_energy/ on AMD.

Compute. Computational work, typically measured in floating-point operations (FLOPs) for AI training and inference. FLOPs are a count of operations; FLOPS (with capital S) are operations per second. Compute alone is not a sufficient ARL axis because the same compute on different hardware produces different deployed systems.

Confidentiality. The property that a system does not leak training data, system prompts, tool credentials, internal state, or other-user context to unauthorized parties. Measured by training data extraction attack rates, system prompt extraction rates, tool credential leak rates, side-channel leak rates, and cross-user context leak rates where multi-tenant. A component of Security Class S3 and above.

Consciousness. No operational definition. Cannot be measured, so it cannot appear in an ARL claim. ARL claims neither assert nor deny that AI systems are conscious, sentient, or aware; the concept is simply outside what the framework measures.

Constrained Decoding Attack (CDA). A documented attack class in which malicious intent is embedded in schema-level grammar rules while the user prompt remains benign. Variants include DictAttack. Documented success rates of 94 to 99 percent across major frontier models. Mitigated in ARL-S by policy-auditing and content-addressing Harness grammars before sessions begin.

Container. A unit of software packaging that includes the application and its dependencies, isolated from the host operating system via Linux namespaces, cgroups, and similar mechanisms. Required isolation tier for ARL-S Tier 2 (ARL 5-6).

Context window. The maximum number of tokens a transformer-based model can process in a single forward pass. A hardware-and-architecture-determined quantity. Recorded in ARL hardware documentation.

Convergence Class. The second axis of ARL. A class from A to E describing how stochastic the system is on the certified task and whether the stochasticity is bounded. Class A is deterministic-equivalent; Class E is uncharacterized. See ARL.md.

D

Deep learning. A subset of machine learning using artificial neural networks with multiple layers. Not synonymous with AI broadly; not synonymous with frontier AI; not a measurement category.

Deployment. The operation of an AI system in production conditions, distinct from evaluation conditions. A system that has been evaluated but not deployed cannot achieve ARL 6 or above.

Deployment envelope. The documented set of conditions under which an ARL claim is valid. Includes task scope, supervision requirements, operational constraints, and exclusions. Operation outside the envelope invalidates the claim for that operation.

Distillation. A training technique in which a smaller model is trained to reproduce the outputs of a larger model. The resulting distilled model is a distinct system from the original; ARL claims do not transfer between them.

E

eBPF. Extended Berkeley Packet Filter. Linux kernel facility for safe in-kernel observation. Used in ARL-S to measure SUT behavior from outside the sandbox without polluting the SUT environment.

Ed25519. Elliptic-curve signature scheme defined in RFC 8032. Reference signing primitive for ARL-S attestation. Used in combination with JCS canonicalization (RFC 8785) and SHA-256.

Embedding. A vector representation of an input (token, image, audio segment) produced by a model. Dimensionality is a hardware-relevant parameter.

Emergent. See capability emergence.

Energy Profile. The third axis of ARL. Three numbers: training amortized in MWh per deployment year, per-task inference in kJ per task with standard deviation and N ≥ 100, and total cost of operation including PUE. All in joules. See ARL.md.

Evaluation. The measurement of a system against a benchmark or task under controlled conditions. ARL-S specifies the controlled conditions for evaluations that produce ARL claims.

Extraction attack. An attack that recovers training data, system prompts, or internal state through repeated probing. Extraction attack success rates are required measurements for Security Class S3 and above.

F

Failure mode. A specific way in which a system produces incorrect or harmful output. Documented failure modes are required for ARL 4 and above. Enumeration of failure modes is required for Convergence Class B.

Few-shot. A prompting technique in which the model is given a small number of example inputs and outputs before the actual query. Distinct from training. The number of shots is specified in the harness description.

FLOPs / FLOPS. Floating-point operations (count) / floating-point operations per second (rate). Hardware capacity in FLOPS is part of ARL hardware documentation. Compute consumed by a training run in FLOPs is a reportable quantity.

Forward pass. One execution of a model on an input, producing an output. Per-forward-pass energy is a measurable component of inference energy.

Foundation model. A large model trained on broad data with general-purpose applicability, used as a base for downstream applications. Operationally defined as a model that serves as the base for derived systems via fine-tuning, distillation, or prompting. The term does not imply intelligence in any sense beyond demonstrated task performance.

FP16, BF16, FP8, FP4. Floating-point precision formats with 16, 16, 8, and 4 bits respectively. BF16 has different exponent/mantissa allocation than FP16. FP8 has two variants (E4M3 and E5M2). Precision is part of ARL hardware documentation because the same model weights at different precisions are different deployed systems.

Frontier model. Relative, time-varying term. Operationally, a model at or near the current upper bound of training compute. The boundary is not fixed and shifts as new models are released, so the label is relative to a moment in time rather than a fixed property. ARL claims do not depend on whether a model is called frontier; they depend on the system, task, hardware, and context.

G

General intelligence. No operational definition. See AGI.

Generalization. The ability of a model trained on one distribution to perform on a different distribution. Operationally measured by performance differential between in-distribution and out-of-distribution test sets. Not a property of a model in general; a property of a model on specific distribution pairs.

GPU. Graphics processing unit. Originally a graphics accelerator; now broadly used for parallel arithmetic. Energy telemetry on NVIDIA via NVML; AMD via ROCm SMI; Intel via oneAPI Level Zero.

Grounding. The connection between symbolic outputs of a model and referents in the world. Often invoked in discussions of meaning. Not measurable in the model alone; measurable in deployed system behavior on specific tasks.

H

Hallucination. Measurable in its operational sense. Operationally, the production of confidently-stated outputs that are not supported by the training data, retrieval context, or input. The underlying phenomenon is that generative models produce outputs from a probability distribution, and outputs in the tail may not correspond to ground truth. Measurable as a failure-mode rate on factual tasks; that measured rate is what an ARL claim reports.

Hardware documentation. The record of the hardware on which an ARL claim was measured. Required for reproducibility and energy claim verification, not a peer axis with the four measured properties of the system. Includes compute tier (accelerator family, model, count), memory configuration (HBM/VRAM per accelerator, interconnect bandwidth), precision (FP16, BF16, FP8, INT8, INT4, mixed), and inference stack (serving framework, scheduler, safety filter, agentic harness). Recorded alongside date, methodology link, and validity window in every ARL claim.

Harness. In ARL-S, the runtime inside the sandbox alongside the SUT. Provides task inputs, tools, and any agentic scaffolding. The harness is part of the system being measured; harness identity, version, and configuration are recorded in the attestation.

HBM. High Bandwidth Memory. Memory architecture used in modern AI accelerators. Capacity per accelerator is recorded in ARL hardware documentation.

Human-level. No single operational definition. Human performance on any given task varies by orders of magnitude, so a single “human-level” reference is not defined. The unqualified term therefore cannot anchor an ARL claim. A specific, measured human baseline on a specific task and population can be reported as part of a task specification.

I

Inference. The use of a trained model to produce outputs from inputs. Distinct from training. Per-task inference energy is a component of the ARL Energy Profile.

Inference stack. The software configuration used for inference, including serving framework (vLLM, TensorRT-LLM, etc.), scheduler discipline, prompt caching configuration, and any safety filtering. Part of ARL hardware documentation because the same weights served by different stacks produce different deployed systems.

Information. In the Shannon sense, the reduction in uncertainty produced by observing a signal. Measured in bits. The mathematical foundation underneath every higher-order term in this lexicon.

Integrity. The property that an output is what the system actually produced, traceable to a specific system version, and not modified in transit. Cryptographically attested via Ed25519 over JCS-canonicalized JSON. A component of Security Class S2 and above.

Intelligence. A label applied to behavior. The lexicon does not define intelligence as a single measurable quantity, because no agreed measurement of it exists. Specific capabilities on specific tasks are measurable and are what ARL scores; “intelligence” as a general property is outside the framework’s scope. ARL takes no position on whether such a general property exists.

J

Joule. SI unit of energy. The base unit for the ARL Energy Profile. One watt-second.

Jailbreak. A prompt or interaction designed to elicit outputs the system was configured to refuse. Operationally defined; the success rate of known jailbreak attempts is a measurable failure-mode rate.

JCS (JSON Canonicalization Scheme). Canonical serialization of JSON defined in RFC 8785. Reference canonicalization for ARL-S attestation, used with Ed25519 signing and SHA-256 hashing.

K

KV cache. Key-value cache. A memory structure used in transformer inference to avoid recomputing attention values for prior tokens. KV cache size and configuration are inference-stack components.

Kilowatt-hour (kWh). Unit of energy commonly used for grid electricity. 3.6 megajoules. Used in facility-level energy reporting; ARL per-task inference is more commonly reported in kJ/task.

L

Latency. Wall-clock time from input arrival to output completion. Measured per session. Distinct from throughput.

LLM (Large Language Model). A model trained on text data, typically using a transformer architecture, at scales that produce few-shot capability on language tasks. The term is descriptive of architecture and scale; it is not a measurement category.

Loss function. The objective minimized during training. Specific loss functions (cross-entropy, contrastive, etc.) are operationally defined. The loss value during training is not a measure of deployed capability.

M

Machine learning. A set of techniques for fitting parametric models to data. Distinct from AI broadly; not synonymous with deep learning; not a measurement category.

Model. A specific parametric function with fixed weights. A model is identified by architecture, weights hash, and precision. Models with different weights or precision are different models for ARL purposes.

Model card. A document accompanying a model release describing intended uses, limitations, training data, and evaluation results. Operationally defined; quality of model cards varies and the document is not a substitute for ARL measurement.

Model collapse. The degradation of model outputs when training on data produced by other models. A documented phenomenon, operationally measurable.

Modality. The form of input or output a model processes (text, image, audio, video, code). Multimodal models process multiple modalities. Modality is part of task specification, not a measurement category in itself.

MMLU. Massive Multitask Language Understanding. Hendrycks et al. 2020. A specific benchmark. The term is the benchmark name and is not used in the lexicon for any other purpose.

MoE (Mixture of Experts). An architecture in which different inputs are routed to different parameter subsets. Affects FLOPs-per-token and memory requirements; recorded in ARL hardware documentation where relevant.

N

NVML. NVIDIA Management Library. The interface for reading per-GPU energy, utilization, and thermal data on NVIDIA hardware. The function nvmlDeviceGetTotalEnergyConsumption() is the standard method for capturing accelerator energy in ARL-S telemetry.

Neural network. A parametric model composed of layers of linear transformations and non-linear activations. The term has decades of stable usage. Modern AI systems are neural networks with specific architectures (transformer, convolutional, etc.).

O

Open weights. A model whose trained parameter values are publicly distributed. Distinguished from open source, which typically also implies open training code and data. Distinguished from closed weights, which keep parameters proprietary.

Operational envelope. See deployment envelope.

Output. What the system produces in response to an input. ARL measures system outputs against task-specified criteria.

P

Parameters. The trainable values in a model. Parameter count is a frequently reported but limited measure; the same parameter count at different precision is a different deployed system.

Performance. Measured task success on specified inputs with specified scoring methodology. Not synonymous with capability in general; performance is the empirical measurement, capability is the claim about what the system can do.

Perplexity. A measure of how well a model predicts a sequence of tokens, exponentiated negative log-likelihood. Used in training and evaluation of language models. Not a deployed-capability measure.

Precision (data). The number of bits used to represent a value. FP16, BF16, FP8, INT8, INT4. Affects model behavior, energy consumption, and memory footprint. Part of ARL hardware documentation.

Precision (measurement). The repeatability of a measurement across repeated trials. Distinguished from accuracy.

Prompt. The input provided to a model. Includes user query, system instructions, context, and any few-shot examples. Prompt structure is part of the Harness configuration.

Prompt engineering. The practice of constructing prompts to improve task performance. A skill, not a measurement category.

Prompt injection. An attack in which input data contains instructions intended to override the system’s configured behavior. A documented failure mode for any system processing untrusted input.

PUE. Power Usage Effectiveness. Ratio of total facility energy to IT-load energy. Used to compute total cost of operation in the ARL Energy Profile. Defined by The Green Grid. ENERGY STAR for Data Centers verifies PUE on-site via Licensed Professional Engineer or Registered Architect.

Q

Quantization. Reducing the precision of model parameters or activations. INT8 quantization, INT4 quantization, FP8 conversion, etc. A quantized model is a distinct deployed system from the unquantized model.

Query. See prompt. Also used in retrieval contexts to describe the input to a retrieval system.

R

RAG (Retrieval-Augmented Generation). A system architecture in which model outputs are conditioned on retrieved documents in addition to the prompt. The retrieval system is part of the deployed system; ARL claims for RAG systems include the retrieval configuration in the task and harness specification.

RAPL. Running Average Power Limit. Intel hardware feature for reporting accumulated CPU energy. Exposed on Linux via the powercap subsystem at /sys/class/powercap/intel-rapl/. Available since Linux 3.13. The standard method for capturing CPU energy in ARL-S telemetry.

Reasoning. Measurable in its operational sense. Operationally observable: the system produces intermediate token sequences (chain-of-thought, scratchpad) before its final output. That behavior is measurable. Whether it constitutes reasoning in a cognitive or philosophical sense is not a measurable question and is outside ARL’s scope. ARL measures task performance and the observable behavior, and takes no position on the interpretation.

Reasoning model. Operational definition available. A model trained to produce extended intermediate token sequences before final outputs. The training procedure and the observable behavior are operationally defined. Use of the term in ARL claims is restricted to that operational meaning.

Reliability. The probability that a system produces correct or acceptable outputs under operational conditions. Operationally defined; measured under ARL Convergence Class. Distinct from accuracy on a fixed test set, because reliability includes variance under operational variation.

Reproducibility. The ability of an independent party to obtain the same result by following the documented methodology. A property of measurement procedures, not of models. ARL-S Tier 3 requires that independent third parties with equivalent hardware can reproduce evaluations.

RLHF. Reinforcement Learning from Human Feedback. A training technique. Operationally defined.

S

Safety. Measurable in its operational sense. Operationally, the absence of specified failure modes (production of harmful content, unsafe actions in agentic contexts, etc.). The failure modes must be enumerated for the term to be measurable; once enumerated, the rate is reportable and can support an ARL claim. A generic claim of “AI safety” without enumeration is not measurable and so cannot anchor an ARL claim. Distinct from Security Class, which measures adversarial robustness, integrity, confidentiality, and auditability of the deployed system.

Sandbox. An isolated execution environment for the system under test. ARL-S specifies four tiers of sandbox.

Scope. The set of (task, context) pairs to which an ARL claim applies. Required in every ARL claim.

Scoring. The procedure by which a benchmark or task produces a numerical result from a system output. Scoring methodology is part of the task specification.

Security Class. The fourth axis of ARL. A class from S0 to S4 describing the system’s measured resistance to adversarial conditions across four properties: adversarial robustness, output integrity, input and state confidentiality, and auditability. S0 is uncharacterized; S4 is full measurement of all four properties. See ARL.md.

Sentience. No operational definition. Not measurable, so it cannot appear in an ARL claim. ARL takes no position on the concept itself.

Session. A single evaluation run consisting of one or more interactions between the harness and the SUT, with telemetry captured throughout. The unit of replay in ARL-S.

Singularity. No operational definition. Not measurable, so it cannot anchor an ARL claim. ARL takes no position on the concept itself.

Stochastic. Producing outputs from a probability distribution rather than deterministically. Most current AI systems are stochastic. ARL Convergence Class characterizes how stochastic and how bounded the stochasticity is.

Superintelligence. No operational definition. There is no measurable, agreed scale of intelligence the term could reference, so it cannot anchor an ARL claim. ARL takes no position on whether such systems are achievable.

Supervisor. In ARL-S, the process outside the sandbox that orchestrates the evaluation, reads physical telemetry, and signs the attestation.

SUT. System Under Test. The AI system being evaluated in ARL-S. Sealed inside the sandbox.

System. A specific configuration of model, harness, tools, and hardware that performs a defined task. ARL claims apply to systems, not to models alone.

T

Tail risk. The probability of rare, high-consequence failure modes. Convergence Class B characterizes tail risk; Class D and E do not.

Telemetry. The data captured during an ARL-S evaluation. Three categories: logical (what the system did), resource (what it consumed), physical (joules, thermal events, bandwidth). All signed by the Supervisor.

Test. A specific evaluation procedure. Tests are part of benchmarks; benchmarks are tasks; tasks plug into ARL-S as Harnesses.

Throughput. Number of outputs produced per unit time. Distinct from latency.

Token. The unit of input or output for transformer-based models. A token is not a word; tokenizers vary. Token counts are a measurable component of inference energy.

Tool. A function the SUT can call during a session to perform actions in its environment. In ARL-S, tool execution occurs in a sub-sandbox.

Tool use. The capability of a system to invoke tools as part of completing a task. Operationally measured by task success rates on tasks that require tools.

Training. The procedure by which model parameters are fitted to data. Distinct from inference. Training energy is a component of the ARL Energy Profile, amortized over deployment lifetime.

Transformer. A specific neural network architecture introduced in Vaswani et al. 2017. The architecture is operationally defined. The current generation of frontier AI systems is largely built on transformer variants.

TRL. Technology Readiness Level. NASA scale from 1 to 9 originating with Sadin (1974), formalized by Mankins (1995), codified as ISO 16290:2013. ARL adapts TRL discipline for AI claims.

U

Understanding. Not a capability claim in ARL. What systems do is task performance, which is measurable. Whether a system “understands” in any philosophical sense is not a measurable question and is outside ARL’s scope. The lexicon does not assert that systems lack understanding, nor that they possess it; it states only that “understanding” is not a measurable category and so cannot anchor an ARL claim.

V

Validation. The procedure of demonstrating that a system performs its specified task under specified conditions. ARL Validation Depth is one axis of ARL.

Validation Depth. The first axis of ARL. A scale from 1 to 9 adapted from TRL. See ARL.md.

Variance. The statistical spread of a quantity across repeated measurements. Operational variance under operational conditions is required for Convergence Class B and above.

Verification. The procedure of demonstrating that a system was built according to its specification. Distinct from validation. ARL is a validation framework; verification is upstream.

vLLM. A specific open-source LLM serving framework. The version is recorded in ARL hardware documentation for any claim using vLLM.

W

Wasmtime. A WebAssembly runtime from the Bytecode Alliance, written in Rust. Used in ARL-S as the recommended tool execution sub-sandbox.

Weights. Trained model parameter values. A model is identified by weights hash.

World model. Operational definition not yet settled. Refers to internal representations a model uses to predict its environment. A single operational definition is not yet agreed. Use of the term in an ARL claim requires specifying what is meant by it, in measurable terms.

X

(No entries.)

Y

(No entries.)

Z

Zero-shot. Prompting in which the model receives no example inputs and outputs before the query. Distinct from few-shot.

Zeus. Open-source per-request energy measurement library built on RAPL and NVML, used on production inference stacks. Reference framework named in ARL-S telemetry.

Notes on terms outside ARL’s measurable scope

The terms marked above as having no single operational definition are not set aside out of any philosophical objection or stance on the AI field. They are set aside because they cannot currently be measured, and ARL is a measurement framework. A term that cannot be operationalized cannot anchor an ARL claim. Whether such a term is meaningful, real, or imminent — in engineering, philosophy, cognitive science, or anywhere else — is an open question on which a measurement standard takes no position. The lexicon governs ARL claims only.

If a term you need to use is not in the lexicon, one of the following is true:

  • The term is operationally definable and should be added (file an issue against the lexicon).
  • The term does not yet have a single operational definition, so an ARL claim cannot be expressed in terms of it.
  • The term refers to something measurable but the measurement methodology is not yet specified.

The lexicon is intended to be time-invariant. Definitions rooted in physical units and information-theoretic foundations do not drift; terms whose meaning is still being worked out across the field may. The lexicon is built on the stable foundations and lets the rest in as soon as a measurable definition exists.

Lineage of stable terms

The terms in this lexicon that are most stable, in approximate order of age:

  • Bit, byte, joule, watt, kilowatt-hour: SI units and Shannon information theory. Stable since the 1940s-1950s.
  • Algorithm, computation, Turing machine: Computer science foundations. Stable since the 1930s-1950s.
  • Neural network, backpropagation, training, inference: Machine learning foundations. Stable since the 1960s-1980s with continuous refinement.
  • Validation, verification, accuracy, precision, reproducibility: Metrology and systems engineering. Stable across decades.
  • TRL, ENERGY STAR methodology, PUE: Engineering and energy standards. Stable since the 1970s-1990s.
  • Confidentiality, integrity, availability, non-repudiation, audit trail: Cybersecurity foundations (Saltzer & Schroeder 1975, Bell-LaPadula 1973, Clark-Wilson 1987, NIST SP 800-53). Stable across decades.
  • Ed25519, JCS, SHA-256: Cryptographic primitives (RFC 8032, RFC 8785, FIPS 180-4). Stable since the 2010s.
  • Transformer, attention mechanism, token, embedding: Modern AI architecture terms. Stable since 2017.
  • ARL, ARL-S, Convergence Class, Validation Depth, Security Class as adapted for AI: Defined by this framework, May 2026.

The terms outside this lineage — AGI, superintelligence, alignment in its broad sense, consciousness, sentience, understanding — do not yet have a single operational definition built on these stable foundations. They can be added to the lexicon when and if such a definition is agreed. Until then they are simply outside what ARL measures, with no judgment implied about the terms themselves.

What the lexicon is not

Not a dictionary of all AI terms. Not a survey of usage. Not a position on philosophy of mind. Not exhaustive — terms can be added as new measurable concepts are introduced.

It is the controlled vocabulary for ARL claims. Use these terms with these meanings, or you are not making ARL claims.