24 proposed concepts
A working Armenian AI glossary with definitions and revision status
Twenty-four core concepts prepared for consistent bilingual use, with every Armenian form explicitly marked as provisional.
A useful termbase governs concepts, not isolated word swaps. Each entry below has an English concept, a short definition, and an Armenian form produced for this edition. The Armenian forms are proposals from the Gemini writer route and may change when usage evidence or community corrections reveal a better choice.
01
What we know now
- 01
NIST's trustworthy-AI glossary is designed to support shared, use-case-neutral language.
- 02
The starter list separates a concept's definition from the choice of Armenian term.
- 03
Every Armenian term in this version remains proposed, with no claim of native or institutional approval.
02
Why this matters for Armenia
A governed glossary reduces accidental variation across Armenian articles while keeping disputed or unnatural forms open to evidence-based correction.
03
core and lifecycle
AI, machine learning, training, inferencemodel and interaction
model, parameters, tokens, context, prompts, instructions, outputdata and adaptation
dataset, embedding, retrieval, fine-tuningevaluation
benchmark, metric, uncertaintysystems and controls
confabulation, agent, tool, guardrail, multimodal, open weightsCategories help browsing; they do not change the concept definitions.
04
How to read the glossary
The English term names the concept used in source material. The definition describes that concept in plain language. The Armenian term is a publication choice for consistent drafting, not a declaration that other forms are wrong.
A term can be stable in one context and misleading in another. Writers should preserve the source's meaning, expand an abbreviation on first use, and prefer a brief explanation over forcing an unfamiliar coinage into every sentence.
- Concept first, word choice second.
- Define an abbreviation on first use.
- Keep variants visible when usage is unsettled.
- Revise the term and affected articles together.
05
What “proposed” means
Proposed means the entry can be used consistently in this private preview but has not received native-language or institutional approval. The label prevents a model output from acquiring false authority through repetition.
A correction should identify the concept, current form, suggested form, context, and evidence. Agents then compare meaning, actual Armenian usage, morphology, clarity, consistency, and downstream effects before issuing a versioned change.
- No entry claims official language authority.
- Popularity alone does not prove meaning.
- A credible meaning error can quarantine affected copy.
- Revision history should preserve the reason for change.
06
Use definitions to resolve near-neighbours
Training and inference are different stages. A model and an agent are different system boundaries. A benchmark is a test arrangement; a metric is the rule used to calculate a result. An open-weight model exposes weights under stated terms, which does not automatically make every component open source.
These distinctions matter more than surface fluency. When an Armenian sentence blurs them, the writer should rewrite the explanation rather than rely on a familiar but imprecise borrowed word.
- Model is not agent.
- Benchmark is not metric.
- Training is not inference.
- Open weights do not automatically mean open source.
07
Revision workflow
Each accepted change should update the termbase version, record the old and new form, state the reason, and identify affected content. Mechanical checks can find spelling and variant drift, but they cannot establish naturalness or semantic fit.
This starter glossary is therefore complete as a versioned proposal, not complete for all time. New concepts should enter only with a definition and source; existing concepts should change only with a visible record.
- Propose with context and evidence.
- Compare meaning and real usage.
- Test in full sentences across content types.
- Version the decision and update dependent copy.
08
Working directory
A machine-based system that produces predictions, content, recommendations, or decisions from inputs.
Methods that use data to improve performance on a task without writing every rule directly.
A learned mathematical representation used to produce an output from an input.
The process of adjusting a model from examples or feedback.
Using a trained model to produce an output for a new input.
An organized collection of examples or records used for training, evaluation, or another defined purpose.
A learned numerical value that helps determine a model's behavior.
A unit of text or other input that a model processes; it may be shorter or longer than a word.
The amount of input and generated material a model can consider in one interaction.
Input that asks or guides a model to perform a task.
Higher-priority task guidance supplied by the application running a model.
The content, prediction, action proposal, or other result produced by a system.
A numerical representation that places related items near one another in a learned space.
A workflow that retrieves relevant material and supplies it to a generative model before an answer is produced.
Additional training that adapts an existing model to selected examples or objectives.
A defined task, dataset, and procedure used to compare system performance.
A rule for calculating or summarizing an observed result.
Plausible-seeming output that is unsupported or inconsistent with the relevant evidence, often called hallucination.
A system that observes state, selects steps, and may use tools while pursuing a bounded goal.
A capability an agent or application can call to read information or perform an action.
A rule or mechanism intended to prevent, detect, or contain specified unwanted behavior.
Able to process or produce more than one form of information, such as text and images.
A model whose learned weights are available under stated terms, without implying that all code or data is open.
What is not known or not reliably determined about a claim, measurement, or system output.
09
When proposing a better Armenian form
A strong correction helps agents compare meaning in context.
- 01
Name the English concept and current Armenian form.
- 02
Provide the suggested form in a complete Armenian sentence.
- 03
Explain the meaning difference or naturalness problem.
- 04
Link public usage or a trustworthy terminology source when available.
- 05
Identify other entries or articles that the change would affect.
10
Limits of this edition
The 24 Armenian forms are provisional Gemini output and do not claim native or institutional approval.
Definitions are short working explanations, not complete technical standards.
The glossary covers a starter set and will need versioned additions and corrections.
SRC
Source desk
Direct links to the material behind this selection. Seeing the source matters as much as reading the synthesis.