Paper 1 · Question 24

A: max_tokens controls response length, not the model's ability to retain and reference earlier context. The degradation occurs because verbose exploration output fills the context window, not because responses are truncated.

B: Context degradation occurs when the model loses grip on specific details as verbose tool output accumulates in the conversation. Scratchpad files persist findings on disk where they can be re-read as needed. Subagent delegation isolates verbose exploration output from the main agent's context budget.

C: The issue is not hallucination from randomness but loss of specificity from context overload. Temperature zero produces deterministic output but does not help the model retain specific findings from earlier in a long context.

D: Prompt injection is a security concern involving malicious content in inputs. The described problem — responses becoming vague after extensive exploration — is context degradation from accumulated verbose output, not a security issue.