Fresh sessions are an important isolation boundary, but they may not reset application-level answer caches, workspace memories, or retrieval reuse outside the conversation object. A new thread can therefore still inherit state that changes what the next case sees.
For the spike, I would run the same target case both alone and after a deliberately different predecessor, then compare the actual context records rather than requiring identical generated wording. Recording cache-hit flags and workspace/session identifiers in the trace would help explain any order dependence before it gets mistaken for a model-quality difference.
Thanks, this is a good catch. I'll update the article to cover this!
A compact example could make the update especially testable: run a target case alone, then after a predecessor that primes a different workspace fact. Compare the retrieved context and cache flags for the target in both runs.
If those inputs match but the generated wording changes, that is a different class of variation from inherited application state. Keeping those two outcomes separate would help readers avoid diagnosing every nondeterministic answer as session leakage, while still catching leakage that a fresh chat does not reset.
Ahmet Özel
Fresh sessions are an important isolation boundary, but they may not reset application-level answer caches, workspace memories, or retrieval reuse outside the conversation object. A new thread can therefore still inherit state that changes what the next case sees.
For the spike, I would run the same target case both alone and after a deliberately different predecessor, then compare the actual context records rather than requiring identical generated wording. Recording cache-hit flags and workspace/session identifiers in the trace would help explain any order dependence before it gets mistaken for a model-quality difference.