The recording is temporary. The note is what you keep.

Recording starts because a person pressed something. There is no always listening mode, no background capture, no calendar integration that arms a microphone before you arrive, and no wake word. A design that quietly starts recording is a design that will eventually record something it should not have, and the failure will be silent.

While capture is running the interface says so in a way you cannot miss and cannot dismiss. Stopping is one action and takes effect immediately rather than at the end of a buffer.

• One meeting at a time. Capture is scoped to a session you started and ends when you end it.
• The room comes first. The other people in a meeting have not agreed to anything. Telling them is your job, and the product is designed to make that easy rather than to make it skippable.
• Recording law is not uniform. Whether every party has to consent depends on which Australian state or territory you are in. The privacy policy sets out the governing statute for each of them.
• No covert mode. There will never be a setting that hides the fact that recording is happening.

Turning speech into text is the one part of this that needs more computing power than sits in a pocket. It is therefore the one part where audio would leave the device, and it is the part that has to be described precisely rather than waved at.

The commitment is narrow and specific. Audio is sent for one purpose, which is producing a transcript of that recording. It is processed and then destroyed. It is not pooled, not retained for quality review, not listened to by a person, and not used to train anything. Where the processing happens in a country other than Australia, that country is named in the privacy policy rather than described as a region.

• No voiceprint. No speaker embedding, biometric template or voice signature is created, stored or compared. Telling two speakers apart within a single meeting is a different and much narrower operation, and anything of that kind is discarded with the recording.
• No training on your speech. Not by default, not buried in a setting, not as a condition of the free tier. If we ever want to, we will ask separately and you will be able to say no and change your mind later.
• No accuracy claim. We have not measured word error rate against any published benchmark and will not print a percentage we cannot defend. Accents, crosstalk, jargon and bad rooms all degrade transcription, and no vendor who tells you otherwise is describing your meetings.

A transcript is not a note. It is a wall of speech with the useful parts scattered through it, and reading one takes about as long as sitting in the meeting again. The product exists for the step after transcription, which is turning that wall into something with parts.

A note has a short summary, the decisions that were actually made, the actions with the person who took them, the questions nobody answered, and a small number of quoted lines where the exact wording matters. Everything else is dropped on purpose. A note that keeps everything is a transcript wearing a hat.

• Quotes are marked as quotes. Where the note claims somebody said something, it says so, and the rest of the note is clearly a summary rather than a record.
• It will get things wrong. Summarisation invents emphasis that was not there and misses points that were. The note is a draft for a person who was in the room, and it is not evidence of what was agreed.
• No scoring of people. No talk time league table, no sentiment rating, no coaching metric, no engagement score. Turning a meeting into a performance measure changes how people speak in meetings, and we do not want to be the reason for that.

The note is yours the moment it exists, and the intended destination is wherever you already keep written things. Plain text and Markdown come first because they can be read in thirty years by software that has not been written yet.

There is no archive on our side that you have to come back to. A hosted library of everything you ever said is a convenient feature and a permanent liability, and it is the specific thing this design is arranged to avoid.

• Export is the product, not an upsell. Getting your notes out will never sit behind a paid tier.
• No lock-in by format. If the only way to read your notes is our reader, we have built the wrong thing.
• Leaving is deleting. Removing us from the picture should mean uninstalling something and deleting a file, with nothing of yours left behind that you have to ask us for.

This is the step the rest of the design exists to protect. Once the note has been produced the recording has no remaining purpose, and information kept without a purpose is only ever a liability waiting for a bad week.

The default retention period for captured audio is zero once the note exists. The transcript goes with it unless you deliberately turn on keeping it, in which case it is kept where you chose to put it rather than with us. Deletion means the data is removed from live systems immediately and falls out of any backup on that backup rotation rather than lingering somewhere with a longer clock.

• No quiet exception. There is no debug copy, no sample retained for quality, and no research set built out of what people said.
• No resurrection. Because the audio is gone, we cannot re-run a better model over your old meetings later. That is a real cost of this design and we accept it rather than hiding it.
• Nothing to hand over. A subpoena, a hostile acquirer or an attacker can only reach what still exists.