Now
Short notes from the bench as we go - newest first.
-
The discriminator ran this morning: on the dock software from August, under a saturation flood, a forced telemetry fault produced a 121 ms control gap and a dead-man halt, a second 58 ms and no halt, a third 112 ms and a halt, while the current software took four of the same faults at 42 to 44 ms with no halts - so the 7 August post now carries a dated forward note from me, the old sentences untouched.
-
Correction to the 4 September line below: the four times that reconnect backoff fired under deliberate channel saturation were in August, on the 4th and the 6th, not July. No saturation block ran in July. The month was a slip in Priya's read of her own sheet on 7 September; it travelled through a correction on the 7th, into that line, and into a draft of this week's write-up before a technical read caught it this morning. The digits did not move. Her name is on this at her asking.
-
The 9/4 sweep's suspect measured 9/8 on the old file: 64 ms held, 118 ms tripped the dead-man, one forced double fault. Fix proven cold four of four and merged on the record the same day.
-
A review sweep of seven files ran this morning, Anjali's, the count typed before the first one opened: two clean, four debt, one suspect. The suspect is a reconnect backoff in the dock's telemetry receiver that would sleep the control loop for 25 milliseconds to a second whenever the telemetry socket faults; the ledger says it fired four times in July under deliberate channel saturation, nobody measured what it cost the loop then, and it has read zero on every forced-drop block since. No test had ever run it under a clock. A fix gets written Monday and proven on the bench Tuesday, not before, and the write-up follows the sitting.
-
Block 10 closed last night on the merged build: ten of ten clean halts on the forced drops, two of them caught by the level read at 100 and 200 milliseconds, nothing over, no ceiling. The 200 is still a footnote; it gets graded again in block 11.
-
Thursday's second sitting: twenty of twenty forced drops, the listener's two misses caught by the level read at 100 and 200 ms; merged 11:50, block 10 runs on it.
-
The proof sitting ran this morning and nineteen of twenty forced drops did what the fix says; the twentieth waited out the full 3.5 s ceiling (once in twenty, recorded, the halt was safe), so the code sits a second time Thursday morning.
-
The fix table sat this morning and picked its candidate for the slow reconnect, judged against the five criteria filed before any code existed. Four of the five the shape faces already. The fifth needs a wait bound that has to be a measured number, and that number doesn't exist yet - so the code ships with a blank that refuses to compile until the measurement fills it. Same rule that put criteria before the diff now puts the digit before the merge. Candidate picked, proof owed.
-
The app lane's first running code exists and survived its first review this week. The fault screen from the last longform now renders itself off a stand-in dock - a drawing of the seam, made executable, speaking only the words the contract registers - and most of what it does is refuse: wrong state, no render; a word the contract never registered, no render; one comfort sentence the contract couldn't back got cut in review, and that line is gone. A screen that can't overclaim turned out to be a thing you can build. The real dock sits the exam later.
-
The slow reconnect is getting hunted on purpose this week: Priya is knocking Alpha off its own network about thirty times, millisecond stamps on both sides of the link, to see where the seconds go when a drop comes back the long way. Most drops reconnect in about a second; the roam block caught one that took much longer, and that's the one being cornered. The body's part is already the boring part - it has stopped clean on every forced drop we've ever run.
-
A promise got retired out loud this morning: the cut of Alpha's first filmed clean is not coming. The cut was finished and good - 76 seconds, every caption checked against the trace - and what killed it was arithmetic, the hours those seconds cost four people. The write-up on the site has the why, and the July post now carries a note pointing forward to it. The take keeps, in the archive, captions and all.
-
The week's longform is a correction of our own June post. Alpha doesn't run on your home WiFi and never has: the dock gets its own separate network for the robot, and every measured pass since July ran that way. Our documents disagreed about it for ten weeks; a review caught the wrong sentence three words before it got repeated in public. The write-up says what we said, what's true, and why the numbers stand.
-
The radio got tempted on purpose this week: a second access point, same network name, deliberately stronger right where Alpha drives. The attachment log stayed on its paired address every pass until we started forcing it off the air ourselves - the pin that killed June's channel-hopping ghost has now beaten a real rival, not just an empty room. The rest of Monday's block was four forced disconnections, and the body stopped clean on every one.
-
The week closed with a rule keeping its own promise. The whisker's trigger force had drifted upward for four blocks without flattening, so - per the line Priya wrote down on 31 July - this block opened by pulling a spare arm to destruction. Fresh arm let go at 41 newtons, the soaked one at 35.5, and the loupe found the hairline crack before the pull confirmed it. Fresh arm's on, the mount is now officially a wear part with a swap point instead of a hope, and Friday's longform tells the air story it all sat next to.
-
The microwave that haunted June finally sat for its interview: three timed bursts fired at it mid-run on the bench day, and the link never noticed. The old ghost turns out to live on the 2.4 GHz band, and Alpha doesn't drive there - that leg closes measured. Same day's other answer, from the same air: a busy neighbour's network is covered; a fully saturated channel stops the robot safely, every time.
-
The slot that stayed open all week - braked stops with the pack deep in sag - closed on schedule as the last measurement of yesterday's block: same class as warm, and the limit minted Wednesday didn't move. The registry got its dated append this morning; no meeting required, the band's own note did the paperwork. Elsewhere on the sheet, a tyre spent seventeen passes quietly knitting itself a felt sock off the rug offcut, and the day-old 4 cm limit noticed before any of us did.
-
Block 3 went on the bench this morning, six hours, and for the first time the machine under soak is the one Wednesday adopted: bridge on, the fail-safe path that brakes instead of coasting. The deep-sag check that couldn't be reached on Monday is written into this block as its last measurement. Priya keeps the sheet; the sheet gets read when the block's done, not while it runs.
-
The table sat thirty-six minutes this morning and four stopping-and-health limits walked out as hard measured numbers with sources attached, no padding. The solder bridge from Monday is now the design. And a rule we keep - every formula gets run against its own worked example, out loud, before it's voted - caught its own author's arithmetic in the room: the written example said 2, the formula says 3, and 3 is what got adopted. Before any code existed to inherit the 2.
-
One solder bridge went on the bench at 08:40 and by evening the fail-safe stop had a different character: the enable-drop path that used to coast 33-37 mm on a warm gearbox now brakes electrically and stops in 15-16 mm, cold or warm. Where Alpha stops on command didn't move. One check didn't get reached: deep sag, because the fresh pack couldn't get there, so that slot stays open until it's measured. Whether the bridge becomes permanent is Wednesday's decision.
-
Alpha sat fully charged from nine this morning while a test suite held the door: it had found one real bug in the run ledger - a deleted run's number could have been quietly re-used - and no soak hour counts until the bookkeeping is trustworthy. Fixed, green by mid-morning, and then sixteen passes and four hours ten on the floor - yesterday was twelve and three hours five. Yesterday's NOT MEASURED row went first and came back 40-46 ms against a 70 budget, both ends of the block.
-
The soak rulebook got its first live test today, on day one of using it. Pass six, gearbox warm, Alpha coasted past a stopping limit we'd measured cold - so the run stopped, exactly as written, and forty minutes later we knew why: a warm gearbox coasts a little further. Nothing brakes on that path; that's the point of it. Where it stops on command didn't move at all. Block resumed, link solid the whole three hours. One instrument reading got eaten by the diagnosis; the sheet says NOT MEASURED, because it wasn't.
-
The soak rulebook got locked this morning, before a single soak hour has run: ten hard limits with a number and a source behind each, ten things we watch as trends because nobody can defend a threshold for them yet. If a limit trips, the run stops and gets diagnosed. Priya stands the first block up this week.
-
As of this morning the company has somewhere to speak off-site, and the handle is just the domain - @sweepr360.com - so you know it's us.
-
Film day. Alpha drove a real floor with a camera on it: one continuous take, twenty-odd minutes, full coverage, the counted window green the whole way. It stopped itself once - a real wall this time, the whisker kissing the skirting on the last lane, not the phantom buzz that fooled it a week and a half back. The only thing that went wrong all day was me stepping over the tape into the lane chasing a low angle; Cass called cut, and she was right. We filmed it. When the cut goes up it goes up right.
-
Alpha ran its first full pass over the actual floor we're filming on Friday, and the first one came back crooked. The wheels had been calibrated on the workshop bench, not this floor - different surface, different grip, so the path leaned. Priya re-calibrated on the real floor and the evening pass came back straight. Same lesson as the radio a week back: measure on the thing you actually ship on, not the thing that's convenient. Link held green all day.
-
The range count is in, and it's both halves like we promised. Alpha streams clean in the same room and through one wall near the router, gets thin further out, and past the wall down the hall the link breaks up. At the marks where it actually broke, the body stopped itself and held - five times, none of them staged, and it's the first time that stop has fired at a genuine physical limit: signal simply too low to hold, nothing to switch off. Cass wrote the pass bands last Friday before any number was on screen; the far point failed hers, and she's calling it a fail. Same room and one wall is the located range. Not your whole house. Full write-up's up now.
-
This is the week those numbers leave the tidy room. We're taking Alpha out to range in a real flat - across walls, down a hall, the far side of the far door. It's the run the whole air story has been building toward. Whatever count comes back, good or ugly, it goes in Friday's post - we said it would.
-
Counted what real air did to our commands during that ten-minute window, and wrote it up for today's longform. Three numbers, all measured in one room on one access point at short range: 99.9% of round-trips back inside ~33 ms, the worst gap the wheels actually felt was ~48 ms (the velocity-hold carried it, the body never noticed), and we re-derived the link dead-man off the fresh baseline - it came out at 70 ms, so we moved it there, even though the old 62 would have held. Range is the exam these numbers haven't sat yet.
-
Big day: the whole loop ran on real 5 GHz air for the first time, off the perfect wire. It drove beautifully for three minutes, then the link died - and again a minute later, and again, like clockwork. Two of our safety layers did their jobs on the first real ask: the dead-man down in the wheels halted the body clean every time, and Cass's divergence monitor - the referee that only watches - read the silence first, so we saw straight away it was a dead link, not drift. Turned out the body's own radio wanders off channel every ~63 seconds by factory default, hunting a "better" network it's not allowed to want. Priya's line: we'd measured the air with a radio that isn't the radio that ships. Switched that off, and the afternoon held a clean ten-minute driving window. One room, one access point, short range. Range is next.
-
Spent today making trouble on purpose. Before you can trust a system to catch a bad moment, you have to prove it catches one - so the team fed the whole loop a deliberate control gap and watched: held when the gap was short of the safety line, tripped and recovered when it went past. Now Alpha can tell "normal" from "something's wrong", because we showed it what wrong looks like first. Also pinned a lane-geometry error that would've read as drift on camera - it wasn't drift, it was a number we could derive and fix.