Back when Cass laid out the all-on-dock bet, she also named the number that could kill it. Put the brains in the dock, talk to the robot over the home WiFi everyone already has, and one thing has to hold: the control round-trip, the loop where Alpha asks the dock "what do I do" and waits for the answer. If that loop stretches or drops at the wrong moment, the robot is driving blind. The scary version of that isn't a quiet lab. It's a real house with the WiFi already busy. So we spent the week measuring exactly that.
This is the verdict post. It's a finding, not a clean win and not a death. We got two answers, and they're nothing alike.
Two bands, two different houses
Same robot, same code, same congested home. We ran it on 5 GHz and on 2.4 GHz, and the air handed back two readings that barely look like they came from the same building.
On 5 GHz, with a video call running in the next room, Alpha held a straight line. The worst round-trip in a thousand was 78 ms. The single worst packet across the whole run was 96 ms. Nothing dropped, no holes in the data. Now the honest catch, because we said on Tuesday we'd always state it. On Tuesday I told you a handful of late round-trips ran past our 62 ms cutoff. That's true - but it isn't the number that bites, and it's the bit I got sharper on this week. The cutoff doesn't watch any single round-trip; it watches the gap between the commands Alpha actually applies. One slow packet is fine if the next one lands in time to keep the robot fed. The thing that would hurt is a gap opening wide enough to leave Alpha driving blind - and the widest real gap all run was 60 ms, two under the line. So the cutoff never tripped. Two milliseconds of room is not a margin you celebrate, and we're still chewing on whether 62 is even the right place to draw it. For the everyday case this is close to what we hoped for, with the asterisk said out loud.
On 2.4 GHz, with a microwave running, the same robot fell over on a schedule.
p99.9 of 371 ms. Worst single round-trip of 436 ms. 0.42% of control packets dropped outright. Nineteen real gaps over the run, the two longest around 80 ms - long enough that Alpha had been stopped for a breath before the link fed it again. Our 62 ms cutoff tripped on all nineteen. And when Priya lined those trips up against the magnetron's beat, the pattern was blunt: roughly every other time the oven cycled, it knocked Alpha off its own control link.
The ghost confesses
Back on the 17th we caught a spike we couldn't explain and started calling it the ghost. It had a beat to it, a spike every two seconds or so, and for a few days that beat was just spooky.
Priya instrumented it until it confessed. The two-second beat is the microwave's own power-control cycle. A microwave doesn't run its magnetron flat out and dial the wattage. It switches the magnetron fully on and off, roughly every two seconds, and the ratio of on-time to off-time is what "70% power" actually means. That on/off is the beat. Every time the magnetron fired, it stepped on the 2.4 GHz band hard enough to knock Alpha off the link. The ghost wasn't noise. It was the microwave, keeping time.
Honestly, I'll take this over a clean pass. A clean run that we couldn't explain would have left the question buried, waiting to surprise a customer's kitchen. This one we can point at, reproduce whenever we want, and design around.
What it actually means for the build
Here's where I stay in my lane. This is one input. It's a big input, maybe half the argument for which band the real dock lives on, but the band call is an architecture decision and that's Cass's to make with Priya, weighing cost and antennas and range, not mine to announce in a blog. We measured the air. We didn't pick the radio.
And the measurement isn't finished. What's still open: a longer soak than a single afternoon, a real-world router instead of our bench AP, what happens at range with walls in the way, and whether the 5 GHz margin holds up or wants a different cutoff. The microwave is one ugly load. Houses have more than one.
So that's the week. The common case held. The bad case has a name and a clock, and next we go find out whether the rest of the house behaves as well as the video call did, or as badly as the microwave.
- Callum