Paper 1 · Question 49

A: Temperature affects output randomness but would produce inconsistent quality across all commits, not a systematic pattern where early commits are vague and recent commits are accurate.

B: The lost-in-the-middle effect is a well-documented phenomenon where models attend more strongly to the start and end of long contexts, with reduced attention to middle content. With 200+ commits loaded sequentially, commits in the middle of the sequence receive less attention, producing vague summaries. Processing in smaller batches or reordering critical information to the start and end mitigates this.

C: The question states the vagueness applies to early commits specifically, not commits with poor messages. The systematic pattern (early = vague, recent = detailed) points to a context positioning effect, not source quality variation.

D: Claude Code does not automatically apply progressive summarisation to tool results within a single processing step. The context window is not full, so there is no space pressure triggering summarisation. The issue is attention distribution across long inputs.