I deleted my product's main feature over 8.6 mm

I'm 16, I've been designing a leather belt that works as a screen-free interface to a phone, and until today its central interaction was pulling a small audio module out of the buckle to start a voice session.

It's gone. The reason is one number, and I want to write down how I got it wrong for so long.

The number

The cavity in the buckle has 8.6 mm of usable depth. That is not a target — it's what's left after the walls:

bay depth        9.2 mm
wall  x2        -0.6 mm   (0.30 each side)
                -------
usable           8.6 mm

Now put an earpiece in it. A comfortable in-ear aims its acoustic nozzle into the concha at roughly 55°. A nozzle of length L leaving the body at angle A projects L·sin(A) out of the back of the body. So the total depth the part needs is the body thickness plus that projection:

total = body + L·sin(55°)
      = body + 0.819·L

Give the tip the ~4.5 mm it needs to reach and seal, and you've spent 3.7 mm before the body starts. Solve 8.6 = body + 3.7 and you get 4.9 mm of body — and that's the generous version, ignoring the crown, the flange, and the fact that the nozzle needs wall thickness too. Run it honestly and you land near 2 mm.

There is no 2 mm earpiece. There is no 2 mm anything that contains a driver, a cell, a microphone, an antenna and an SoC.

Ten shapes

I didn't accept this for a long time. I built ten of them, parametrically, each one measured by the same script that generated it:

• Tipless concha rest — fitted easily. Didn't seal. An earpiece that doesn't seal is a speaker you wear, and nobody wants that at this price.
• Collapsing silicone tip — fitted. But it sits at 73% axial compression every single time it docks, for hours a day, for years. I have no data saying silicone survives that, and neither does anyone else.
• Flat stem — fitted, and looked like an AirPod someone had stepped on.
• Open-ear cuff — 19.2 × 27.1 mm, clipping the ear rim. When I finally rendered it against an anatomically-scaled ear instead of in isolation, the arc was too open and the pod hung unsupported. It floated.
• Turned aluminium disc — the last one, and the good one. 19.5 × 16.6 × 8.23 mm, a concha lock, a fingernail scallop for extraction. It fitted the 8.6 mm bay with 0.37 mm to spare and every measured assertion passed.

And it was a piece of jewellery that plays music. Not an earbud. I'd have to convince someone to put a metal disc the size of a 2-euro coin in their ear, and the honest answer to "would you wear this?" was no.

The check that lied to me

Here's the part that's actually useful to anyone building hardware.

For several iterations my fit check passed a part that was 13.5 mm deep against an 8.6 mm bay. It passed because I had written the assertion against the rigid nozzle only — the parameter I was thinking about — rather than against the bounding box of the geometry that actually got built.

Parameter arithmetic is not a measurement. The fix was to stop asserting against inputs and start asserting against the mesh:

bbox = measure_world_bbox(EARPIECE)
assert bbox.z <= BAY_USABLE_DEPTH, (
    f"{bbox.z:.2f} mm part in a {BAY_USABLE_DEPTH:.2f} mm bay")

The moment that went in, three "working" designs failed instantly.

There's a related one I got backwards. I put the retention grooves on the module and the rails in the buckle wall — which cannot work, because a rail standing 0.6 mm proud of the cavity wall occupies x 9.9–10.5 from the centreline, while an 18 mm module only reaches x 9.0. They never touch. Lugs go on the module, grooves in the wall. Obvious in hindsight; invisible until something measured it.

What the belt does now

Audio is out of v1 entirely. What's left is the thing the shape is uniquely good at, and which I'd been walking past the whole time:

It measures your waist.

The strap is under tension, the tension is a function of circumference, and a strain gauge in the buckle's load-bearing flexure reads it. That sensor is under 0.3 mm thick, costs pennies, and lives in a part that already exists — it adds no hardware envelope at all.

From that one signal you also get respiration, meal detection, sitting versus standing, and coughing. A watch cannot do any of this. Not because the watch lacks sensors, but because a watch is not tight around your middle. Circumference is the number people actually track when their body is changing, more closely than weight, and nothing measures it continuously today.

Removing the cavity also freed the buckle. It was 11.8 mm thick because a bay needed to be inside it. It's purely mechanical again, and the target is 7–8 mm.

What I'm not claiming

None of this is built. It's a design-intent model with measured assertions, not a prototype.

The whole v0.8 thesis rests on one unverified assumption: that there is a readable signal in strap tension, through clothing, while walking. If there isn't, circumference and respiration both fall at once and I have to start over.

That test is one strain gauge, an amplifier, an ADC and an afternoon. It's the next thing I'm doing, and it's the cheapest possible way to find out I'm wrong.

Everything — including the shape studies that failed and the assertions that caught them — is at vyldor.com.