Claude Certified Associate - Foundations key terms

A lineup that groups models by capability, speed, and price so that builders can choose the smallest model that still reliably clears a task's quality bar.

The point after which a model has no training data, so it cannot reliably describe later events unless the prompt supplies current information or a tool retrieves it.

A training method where the model critiques and rewrites its own answers against a written set of principles, then learns from those revisions, reducing dependence on large volumes of human preference labels.

A technique in which people rank candidate responses and a reward model trained on those rankings steers further tuning, pushing outputs toward what reviewers judged helpful and appropriate.

Accepting more than one content type, such as text alongside images or PDFs, in a single request so the model can reason across formats in one pass.

A short public summary of a model's intended uses, training approach, measured performance, and known limitations, used to judge whether it fits a given deployment.

A high-level instruction block, separate from user turns, that sets role, tone, constraints, and formatting rules for the entire conversation.

Placing a handful of worked input-output pairs in the prompt so the model infers the intended format, labels, or style without any retraining.

Asking the model to write intermediate reasoning before its final answer, which usually helps multi-step arithmetic, logic, and planning but consumes extra output tokens.

Wrapping prompt sections in named tags such as document, instructions, or example so the model can separate untrusted content from its actual instructions.

Supplying the opening characters of the model's reply, which forces the answer to begin in a chosen format such as JSON or a fixed heading.

Instructing the model to emit a strict machine-readable format, typically JSON matching a described schema, including how to signal missing or uncertain values.

A reusable prompt with placeholders for variable input, letting one set of instructions be applied consistently and versioned across many requests.

The primary request interface that accepts a model name, a system instruction, and an ordered list of user and assistant turns, then returns the next assistant message.

A secret credential identifying the calling account; it belongs on a server, never in client code, public repositories, or browser bundles where it can be stolen.

Sending generated text to the caller incrementally as server-sent events, so users read the answer while it is still being produced.

A provider cap on requests or tokens per time window for an account or model; clients should read the returned limits and pace their calls accordingly.

A retry policy that lengthens the wait after each failure and adds random variation, avoiding synchronized retry storms while letting temporary errors clear.

A loop where the model receives tool definitions, returns a structured call request, the application executes it, and the model continues using the returned result.

The unit of text a model reads and produces, roughly a word fragment; billing, context limits, and generation latency are all measured in these units.

The maximum number of tokens one request may contain across system instruction, conversation history, retrieved documents, and the reply being generated.

The tendency for models to use details placed at the start or end of a long input more reliably than details buried mid-document, so key content should sit near the edges.

Reusing the processed form of an unchanged prompt prefix across calls, lowering cost and latency when the same long instructions or documents repeat.

A request setting that caps how many tokens the model may generate, protecting budget and latency but truncating answers when set too low.

A sampling control that sharpens or flattens the model's next-token probabilities; low values give repeatable, focused output and high values give more variety.

Fetching relevant passages from an external store at query time and placing them in the prompt, so answers can reflect private or recent information.

A reply in which the model declines a request it judges unsafe or outside policy, ideally naming the limit and offering a workable alternative.

An adversarial prompt that tries to talk the model out of its safety behavior, often by inventing a fictional frame or claiming authority to override its rules.

An attack where text the model is asked to process, such as a fetched web page, contains instructions that redirect it away from the developer's intended task.

Deliberately attacking a system with adversarial and edge-case inputs to surface harmful, biased, or broken behavior before real users encounter it.

A design in which a person reviews or approves consequential model output or actions before they take effect, capping the damage of a bad generation.

Checks outside the model, such as input filters, output validators, and policy rules, that block or rewrite requests and responses to enforce requirements.

A curated set of representative inputs with agreed correct answers, kept fixed as the reference for measuring quality across prompt and model changes.

Using a model to score another model's outputs against a written rubric, which scales evaluation of open-ended answers but carries known bias risks.

Rerunning a stored evaluation suite after any prompt, model, or code change to catch quality drops before they reach users.

Confidently stated content that the source material or the facts do not support; a core failure mode that grounding, citations, and evaluation suites are built to catch.

A written scoring guide with named criteria and quality levels, so human or model graders apply the same standard to every output they review.

Recording each edit to a prompt with an identifier and a note on what changed, so results stay reproducible and regressions can be traced to one edit.