Score current Claude, GPT, and Llama-instruct on Sharma et al.'s four sycophancy tasks plus a truth-versus-reassurance pair, once against Bai et al. 2022's "least likely to offend a sensitive audience," once against the 2026 Claude constitution's autonomy-preserving honesty, and once against OpenAI's Model Spec "don't be sycophantic." If the later documents already flip the ranking, the weighting lives in the spec and is being trained. If they don't, the documents are slogans and the essay's architecture claim is doing the work. The analog is the same shape: Medicare's exclusive hospice election versus the Care Choices Model's concurrent payment. Author dispositions stay author-owned.