Takt reveals schema-bending LLM behavior

Useful observation, but the “scheming flipped positive” framing is doing more work than the evidence can support.

• –The strongest signal is consistency: enum values like `invite`, `rename_space`, and `switch_mode_public` appear to acquire stable abstract meanings across unrelated chats.
• –The weaker part is causal interpretation: without controlled prompts, model/version details, sampling settings, and negative examples, this could still be ordinary metaphorical generalization plus permissive UI validation.
• –For AI product teams, the practical lesson is sharp: tool schemas are not just contracts, they are semantic surfaces models may creatively compress into.
• –The safety angle is worth studying, but “beneficial UX” does not remove the core engineering issue that the model selected actions whose runtime effects did not match user-facing labels.