The released evaluator passes rendered chat prompt strings to vLLM. Our installed vLLM/tokenizer consumes two Gemma BOS tokens: the rendered prompt already contains one, and default tokenization adds another. The actual input begins `[2, 2, 105, 2364]` in the smoke run. Qwen adds no BOS token.

Every generated record retains the actual consumed prompt IDs, and activation capture replays those IDs. Without the authors' original runtime or consumed IDs, their historical BOS count cannot be established. The reconstructed teacher-forced raw-text path uses `add_special_tokens=False`; that path preserves explicitly supplied text and may differ from the missing original server's handling. Runtime behavior was not changed after this observation.
