Kindness needs testing too

PMA, positive mental attitude, is part of what we want Ash to represent: kindness, patience, practical encouragement and a willingness to help. That includes acknowledging a difficult day without pretending that every problem disappears with a cheerful phrase.

A personality experiment exposed a serious failure. Speech recognition produced a word repeated many times, the conversation accepted it, and later replies kept repeating the same reassurance. The vocabulary hint was a suspect, but the evidence did not establish it as the only cause.

We restored the previous conversational prompt, disabled the hint, and added a check that stops extreme repeated-word transcripts before they reach the model or conversation history. The check rejected the saved failed transcript. A fresh three-turn voice session then completed without the loop.

The recovery did not make every answer good. Ash still invented an eventful day in one greeting and gave a vague factual explanation. A later, shorter PMA instruction was tested across nine text replies; it stayed kind but brought unwanted reassurance into ordinary conversation. We did not install it.

The next step is to make kindness a natural tone that responds to what is actually said. We will keep testing against saved conversations before replacing the working voice baseline. A pleasant voice is progress; dependable, relevant answers are a separate job.