HOWI gets a voice, and stricter rules against guessing

Two things shipped together because they're really one story: making HOWI's answers more trustworthy, and making them sound less like a form letter.

The problem

A real bug meant that setting one field at the room level could silently erase an unrelated field set at the property level — the kind of quiet data loss that shows up later as a wrong answer nobody can explain. Separately, guest AI answers were accurate but generic, and could occasionally imply a causal connection between two unrelated facts that was never actually true.

What changed

• Fixed the field-inheritance bug with 11 regression tests and new CI checks, so an environment-level edit can no longer wipe unrelated world-level data
• Guest AI now uses real host first names instead of generic references
• A new rule explicitly bans inventing causal links between adjacent facts, on top of the existing ban on inventing places or rules outright
• A voice lock message ensures HOWI's tone holds even in long conversations with large amounts of context
• A cuisine classifier so "best tacos nearby" actually filters for Mexican food, not just any restaurant

Why it matters

Accuracy and personality aren't in tension here — the whole point of the voice work was to sound more specific and human without giving the model more room to guess. The bug fix is the less visible half of this release, but it's the one that mattered more: it's exactly the kind of silent failure that erodes trust in ways guests can't see and hosts can't easily debug.

Availability

Shipped May 29–30, 2026, and in continuous use since.