The examination source explicitly identifies D . Non-determinism is a normal LLM property, but that fact does not justify assuming every inconsistency is unavoidable. Systematic diagnosis begins by capturing the complete execution conditions for successful and unsuccessful runs.

A useful trace should expose the actual user input, system prompt, prior messages or context, tools and tool results where applicable, model identifier, request parameters, output, stop reason, and relevant execution events. Current Anthropic agent infrastructure exposes session, span, and agent events specifically to provide observability into session state and execution progress. This makes comparison possible: the engineer can determine whether apparently "same" calls actually differ in accumulated context, tool state, model configuration, parameters, or external data.

A changes a variable before establishing the cause and, on newer Claude releases, temperature may not even be configurable. B generates additional samples but does not ensure the underlying conditions are understood. C prematurely classifies unexplained behavior as acceptable nondeterminism.

The correct sequence is observe, compare, form a hypothesis, reproduce, modify, and evaluate .

Relevant Claude Developer topics: observability, tracing, debugging, model nondeterminism, request-state capture, failure reproduction, and production diagnostics .

===============