Here is how Alpha's radio actually works, as of Wednesday's ruling: your dock creates its own separate WiFi just for your robot. The robot never joins your home network - not during setup, not as a fallback, not ever. Your home network is only the dock's route to the internet, and it's the only network the setup flow will ever ask you about.
If you've been reading along since June, one of those sentences should bother you. It bothered us first. That's the story this week.
A sentence walks into a review
On Tuesday a post about the roam block was sitting in review, and it described Alpha holding onto "the dock's access point." Cass stopped on those three words and asked the question that is now stapled above the review desk: which network does this sentence assume?
The answer turned out to be embarrassing in the specific way we can use. Our own documents were split. The radio spec and the config lock assume the robot talks to an access point we control - pinned channel, no band-steering, a far end that holds still. The onboarding documents assume you pick your home WiFi during setup and the robot joins it. Both sets have existed since June. Neither is wrong inside its own file. They answer different questions, and no one had read them against each other across the boundary - until three words of draft copy forced them into the same room.
And one of the two models had already been published. Our June 19 post said the robot's moves stream "over your home WiFi. The same WiFi your phone is on." As of Wednesday, those sentences are wrong as written.
The correction
The sentences were Cass's, so the telling is hers. Her words, exactly as she wrote them:
In June I told you the dock streams the robot's moves "over your home WiFi. The same WiFi your phone is on." Those sentences were wrong, and they were mine.
Here's what's true. The dock creates its own separate WiFi just for the robot. The robot never joins your home network - not during setup, not as a fallback. Your home network connects to the dock, and only the dock: it's how the dock reaches the internet, and it's the only network setup will ever ask you about.
How I got it wrong deserves a paragraph, because it wasn't a typo. Two sets of our own documents answered two different questions and were never read against each other. The control link needs a network whose far end holds still; reaching the internet needs your network. Two different jobs, and for ten weeks our paper quietly merged them. The June sentence came from one set. The bench came from the other: since July, every measured pass has run on a dedicated access point we configure ourselves - a hundred and six of them, six blocks, on Wednesday's record when we ruled - because the first thing a factory-default network did was tear the link every sixty-three seconds. The measurements were never on your home WiFi. On Wednesday we ruled that the design, in writing.
So no number we've published moves. Every braking distance, every gap, every halt happened under exactly the model we've now written down. One gap remained when I wrote this: our access point has been a separate box on a shelf standing in for the dock, not mounted on the dock itself. That block ran this afternoon - the access point moved onto the dock's position, and the link measured in the same family it always has. The full read gets its own post; the point here is that the gap closed by measurement, not by assumption.
The June post now carries a dated note pointing to this one. The old sentence stays where I wrote it, with the correction beside it.
What the table did with it
The fork didn't ship as a side remark in a test story. It went to Wednesday's table as item zero, and two facts settled it.
Priya's, from the record: every measured pass since July has run on a dedicated 5 GHz access point she configured herself - 106 of them across six blocks on the record the table read Wednesday - a configuration she built after the July 8th run where the radio's factory settings tore the link every ~63 seconds. The rigs have never been on the same WiFi your phone is on. You can only pin a far end you configure, and a customer's router is by definition the end we never touch - free to band-steer, roam-scan, and swap addresses whenever it likes. What block 6 measured is narrower and sharper than that whole zoo: one rival access point, same name, deliberately stronger, and the pin never moved.
Anjali's, from the docs: the two document sets were answering different questions. Control needs a far end that holds still. Onboarding needs the internet. Those are two different networks, and our paper had quietly merged them into one.
So the ruling, now ADR-0013 - Cass's call, Priya's hardware half, Anjali at the table: the control link rides the dock's own access point, the robot never joins the home network, and the home network is the dock's uplink only. The measurements don't move - they were taken under exactly the split model, on real neighbours' air. The one true gap was that the access point had been a separate box on a shelf standing in for the dock, and it stopped being future tense this afternoon: Priya moved the box to the dock's own position and the link measured in the same family it always has. Her full read gets its own post.
The ruling opened one hardware question it couldn't close by itself - the dock's own radio has to host the network - and that question got answered inside the same day. Hosting it on the module already in the dock is out, and not because we chose not to spend the hours: on that part, one radio means one channel, so the control network would get dragged to whichever channel your router happens to sit on. That's the far end we don't control making our channel call for us, the exact thing Wednesday's ruling forbids, and it failed in the vendor's own documentation before anyone touched a soldering iron. So the dock's network gets its own second radio: control on its own chip, the uplink free to have whatever kind of day a home network gives it. The pick is conditional the way everything here is now conditional: the config lock demonstrates on our rig or the part is out. The part's qualification block is next.
The block that started it
The post that tripped the review shipped yesterday, re-worded to say only what the block measured: a rival access point, same name, deliberately stronger, and a pin that refused to move - fourteen passes, then four forced drops, four clean halts. Worth the read on its own, and worth it twice for the fact underneath: the pin at the centre of that story is the same pin the fork argument was about. The model the table ruled on Wednesday is the one the machine had been demonstrating all along.
A customer would never have caught the June sentence. It would have surfaced next year as a support ticket - "why won't my robot show up on my WiFi?" - from someone we'd already told it would. Instead a review caught it three words before it got reinforced, our own documents got read against each other properly for the first time, and the setup flow now asks one question instead of two. The wrong sentence cost us a correction. Finding it any later would have cost a stranger an afternoon.