When Rewriting Feeds on Itself: Prompt Constraints Reshape Multi-Round LLM Drift

Iterative language-model systems often pass a generated response back as the next input. Most evaluations, however, report only a final or single-turn score. We use a deliberately narrow diagnostic: three hosted model routes rewrite six English fact statements for six rounds. We record whether the final response retains a seed- specific marker and whether it switches from English to Han-script text. Under unconstrained rewriting, marker retention is 0/6, 1/6, and 0/6; script switching occurs in 5/6, 4/6, and 6/6 chains. A numbered batch prompt changes both outcomes for two routes (5/6 and 6/6 marker retention; no script switches), but not for the third (0/6 retention; 6/6 switches). In a follow-up over eight seeds and two routes, an explicit output constraint is associated with better stability for both routes. Batching without numbered slots is not. These descriptive results do not show that formatting preserves meaning: the sample is small, the prompts are not fully factorial, and neither metric measures semantic equivalence. They do show that a chain can behave differently after a seemingly minor change to its prompt contract, which makes trajectory-level testing necessary.