Friday I wrote that range was the next exam and we'd post the count good or ugly. This is the count. It's both.
Here's what we did. The whole streamed-control loop - the planner on the dock, the drive commands going out over 5 GHz air, the body driving them, telemetry coming back - has only ever run in one room, close to the access point. This week Priya carried it outward. Same room. Then one wall, near the AP. Then a mid-hall mark, about six metres and a wall away. Then the far point: through a wall, down a hallway, signal on the floor.
What streams and what doesn't
Round-trip latency (p99.9) climbing as you walk away from the AP: 34 ms and 41 ms at the two same-room marks (the bench-side and the far side of the room), 52 ms through one wall near the AP, 61 ms at the mid-hall mark. Worst applied command gap at those same four marks: 46, 55, 66, 74 ms. At the far point the link breaks up and stops giving you a clean number at all.
The plain version, mark by mark:
- Same room: clean. Both same-room marks (34/46 and 41/55 ms). This is the floor the film demo runs on, and it passed with room to spare.
- One wall, near the AP (RSSI around -62): clears the bar, but thin. 52 ms p99.9, worst gap 66 ms against a 70 ms window. Four milliseconds of margin. That's inside the line, not comfortably inside it, and I'm not going to call it comfortable.
- Mid-hall, six metres and a wall out (-66): marginal. 61 ms p99.9, worst gap 74 ms - already inside the slide toward the cliff. One clean halt here.
- Through the wall, down the hallway (-70 and below): out. Streamed control doesn't survive this. The link goes into outage bursts and the body stops.
So the located range is: same room and through one wall near the AP. Not down a hallway. Not across a floor. Anyone reading "works across your house" into that is reading something we didn't measure and didn't find.
The part I'm actually here to tell you
At the far point the link didn't just get slow. Anjali ran the raw trace: 19 moments where a round-trip never came back. Thirteen were short enough to stay inside the 70 ms link window and were carried straight across - the last command just held, same as the 48 ms worst gap we measured last week, and nothing stopped. Six ran long, 100 to 410 ms. Four of those tripped the link dead-man, the watchdog on that 70 ms window. The other two landed inside a burst where the body was already coasting to a stop, so they didn't spawn a fresh halt.
Each of those four, the link watchdog tripped and Alpha coasted to a stop on its own held velocity, no dock involved. Add the one halt at the mid-hall mark and that's five clean stops across the run. And this is the first time that layer has fired at a genuine physical limit - not a factory default we could switch off, like the roam-scan holes back on the 8th, but link budget: the signal simply too low to hold, nothing left to turn off. It stopped the body anyway. Five for five.
The holes weren't a flaky link, either. Between bursts the far point was boringly clean - p99.9 around 40 ms, same as a good short run. The outages arrive in bursts a second or two long, and they line up with the signal sagging. At that distance the radio starts stacking retries into short outages when the level dips. We can't patch that away. It's link budget, physics on this floor with this radio.
The bar was set before the number
The far point failed. I want to be exact about that, because Cass was. She wrote the pass/fail bands for this run last Friday, before anyone had a reading: the far point passes at zero halts, p99.9 under 55, worst gap under 70. It gave four dead-man halts where her bands allowed none, and its outages ran 100 to 410 ms, far past the 70 ms line. That's her fail clause, word for word, and she's calling it a fail with her name on it rather than reaching for "marginal."
What the far point was for was to find where control gives out and whether it survives a whole real floor. It answered both. The cliff is located - roughly -62, where it's clean but thin, sliding to -71, where it's broken. And a whole floor of streamed control, on this radio, is a no. We don't get to make that claim, so we're not making it.
The edge is where a safety design earns its keep, and this one did. We can tell you where Alpha drives on a link and where it stops itself instead, because we walked it out there and watched it stop.