The seam you will never find

Five minutes into a recording, the glasses quietly seal the file and start a new one. No tone. No vibration. No pause in the audio. The next segment begins with a one to two second overlap of what the previous one ended with, and the app stitches the pieces back together, using the overlap to remove the duplicated words.

If we have done this correctly, you will use the product for years without ever finding out it happens.

Why split at all

Four unglamorous reasons, all about failure rather than features:

• Smaller files sync faster over a Bluetooth link that is only awake for a moment.
• Speech-to-text services handle short files more reliably than long ones, and the failure modes are gentler.
• If something goes wrong — battery, corruption, a dropped transfer — the blast radius is one segment, not one afternoon.
• A sealed file is a safe file. Anything already closed cannot be damaged by whatever happens next.

Why five, not ten

This is the part worth writing down, because it is a small decision that demonstrates a rule.

Since the split is imperceptible, the user experience of five minutes and ten minutes is identical. There is no perceivable benefit to the larger number. But there are real engineering costs to it: bigger files, slower transfers, higher transcription requirements, more to lose when something fails.

When a parameter is invisible to the user, take the conservative value. There is nothing on the other side of the trade to defend it.

Ten minutes would have been chosen only out of a vague sense that fewer files is tidier. Tidiness we cannot see is not a reason.

The overlap

One to two seconds of duplicate audio at each boundary costs essentially nothing — a few kilobytes — and it removes an entire class of defect: the clipped word at the seam, the half-syllable that becomes a nonsense token in the transcript, the sentence that reads wrong for no visible reason.

When the app merges the text it finds the overlapping region and drops the repeat. The reader gets continuous prose. Nobody has to know that the sentence they are reading was assembled from two files recorded on a device with no screen.

Where else this shows up

The same principle produces the other invisible rules in the product. Two recordings less than sixty seconds apart merge into one note, so pausing to think does not fragment a thought. Anything under two seconds is discarded, so a pocket press never becomes an entry. A ring buffer catches the 300 milliseconds before the microphone stabilises, so the first word survives.

None of these are features in the sense that they can be demonstrated. They are the absence of four specific annoyances that you would otherwise have felt daily and probably never quite been able to name.