Skip to lesson

AnthropicClaude Certified Architect – Professional2026-100 Prerequisites

Separate model capacity from reliable knowledge

Completed by the knowledge check

Learning objective

Explain how LLMs generate output from tokens and why context windows, sampling, and training cutoffs bound what a model can know and do.

Core idea

A context window bounds the information available during generation; training supplies learned behavior rather than a guarantee of factual correctness. As an architect, you must distinguish three promises: enough capacity to process the request, enough evidence to answer it, and acceptable variability across repeated answers.

These promises need different evidence. A token count cannot establish factual accuracy, a training cutoff cannot establish knowledge of a private transaction, and a sampling setting cannot establish reproducibility. This prerequisite identifies those limits; the numbered lessons own their implementation and evaluation.

How it works

Generation is conditional, not a database lookup

A large language model (LLM) processes tokens, units that can represent words, parts of words, characters, or bytes. Its autoregressive generation predicts successive tokens from preceding context. Sampling selects among possible continuations, so a fluent answer reflects learned patterns and supplied evidence rather than a verified record lookup.

The Messages API documents temperature and top_p as sampling controls, but both are deprecated. Models released after Claude Opus 4.6 do not support setting them: temperature: 1.0 and top_p >= 0.99 are accepted for compatibility, while other values return HTTP 400. On those models, any top_k value also returns HTTP 400. Where temperature: 0.0 is supported, Anthropic still states that results are not fully deterministic. Do not build a reliability promise around a setting the target model rejects. See 5.2.1 for model limitations and 4.3.3 for repeated-run evaluation.

Capacity and knowledge have separate boundaries

Anthropic distinguishes the reliable knowledge cutoff, through which knowledge is most extensive and reliable, from the training data cutoff, the broader range of training data. Neither date guarantees a particular answer. Supply relevant external evidence for private or changing facts; see 3.5.1 for choosing retrieval, long context, or live queries.

The window holds input and output but does not guarantee attention: see 2.4.1 for budgeting with max_tokens and token-count estimates, and 2.4.2 for evidence selection against context rot. POST /v1/messages/count_tokens rejects some inputs the Messages API accepts, including most server tools and the MCP connector; it is not a universal preflight validator.

Key distinctions and trade-offs

Treat capacity, freshness, and repeatability as separate acceptance criteria. Increasing a window addresses fit. Supplying authoritative evidence addresses a knowledge gap. Repeated-run evaluation measures variation. None substitutes for the others, and more context can add distraction as well as tokens.

Architecture / decision scenario

A financial-services team must explain a policy change announced after the selected model’s training cutoff, using a private customer transaction. Its pilot has a two-second response target and requires every explanation to agree with the supplied policy.

You reject the proposal to guarantee identical answers by setting temperature: 0.0. Instead, the integration owner checks the target model’s supported settings, supplies the policy and transaction as evidence, and budgets the complete request under 2.4.1. The evaluation owner tests repeated explanations for factual agreement and latency under 4.3.3. If a larger evidence bundle fits but reduces agreement, you curate it rather than treating capacity as proof of quality.

Common exam traps

Avoid these substitutions:

  • Treating a cutoff date as a guarantee of knowledge or accuracy.
  • Treating zero temperature as deterministic execution on every model.
  • Treating a successful token count as evidence that an answer is correct.

What to remember

Defend separate limits for capacity, knowledge, and variability. Verify the selected model’s interface, supply task evidence, and evaluate outcomes rather than promising reliability from a window size or sampling value.

Knowledge check

References