The demo on Loud Enough is not a happy-path transcript. Someone talks about a leak they already reported, mold, and wanting a repair plus a written reply. The date of the first report never appears. The message that comes back leaves it as [DATE OF FIRST REPORT].
That gap is the feature.
I have watched models complete a story because the story wanted completing. On the healthcare visit app, asking for a summary before the transcript existed was how you invented a visit. Loud Enough is the same rule in a product people might send to a landlord, a manager, or a school: nothing goes out that you did not say.
Quote, then check
Every field arrives with a verbatim quote. The quote is checked against the transcript before it is shown. If it does not match, it is flagged on screen and withheld from the draft. The agent is told. It can correct itself. It does not get to keep the plausible version.
The brief on screen is the tool call, not a second summary. I did not want two models in a row, each one smoothing the last. The product write-up is listen, understand, verify, act. This log is only verify.
Blanks are honest
A bracket is uglier than a guessed Tuesday. It is also something a person can fill in. An invented date in a complaint is a different kind of product: it sounds finished, and it is wrong.
I am not going to claim this makes the model trustworthy in the abstract. It makes this app refuse one specific lie. Close the tab and the session is gone anyway. The thing that has to survive the tab is the message, and the message is not allowed to be better than the transcript.