Priya's lead theory for the walk died this fortnight, on her own instrument, by a size she had written down before the block ran. This is her paragraph, copied from her mail and not retyped.
Two weeks ago I said tyre wear was my lead candidate for a number that has walked the same +0.3% every soak block that has carried a bracket, since the end of July. I also said what it would take to kill it: if wear is the whole walk, the tyres have to lose about 90 microns of diameter per block, and a $30 micrometer can tell 90 from 8 without trying. Block 11 was the first block with the micrometer on the sheet. Both tyres went on at 30.104 and 30.078 millimetres. At the close, after 14 passes and four hours forty, they read 30.096 and 30.071. That is 8 microns off the left and 7 off the right, the same direction on both, and my three-point spread on that instrument runs 2 to 4 microns, so the reading is real and not the instrument breathing. The next morning, cold, both tyres read the same to the micron, so it is not warm rubber either. Eight is a tenth of ninety. So "tyre wear is the walk" is struck, by the size I wrote down before the block ran, not by how I felt about it. Tyre wear itself stands: first direct measurement of it we have ever had, and it is about a tenth of the drift, which makes it a term in the answer and not the answer. One block is one block; block 12 tells me whether 8 microns is a rate or the first block settling on a new zero. The queue behind it is three tests that cost nothing: the rolling radius under the body's own weight against the free diameter the micrometer sees, the raw encoder ticks over a fixed run against the reference bar, and the cold recal bracket before any driving for the grease-and-lash story. The first one is a marked tyre, ten turns on the floor and a tape measure, and it goes in block 12's first stop.
The walk, for anyone arriving cold: at both ends of every soak block Priya re-runs a short floor calibration, and since the end of July the number it gives back has crept the same way by about a third of a percent per block. Small enough that the model still holds every time, and big enough to have a name. Wearing tyres was the obvious suspect, because a smaller tyre covers less floor per turn and the encoder can't tell. The clever bit is the order things happened in. The kill size went on paper first, 90 microns, from what the drift would need if rubber were the whole story. Then the instrument. A theory you are fond of does not get to survive a number you set for it in advance, and this one didn't.
Block 12 ran yesterday with the micrometer on the sheet again and the first of her three free tests at stop one, the marked tyre and the tape measure. Her one sentence on that, hers, cleared through Cass before it went anywhere near this page:
"Under the body's own weight the tyres roll about a third of a percent short of the circumference the micrometer sees in the air; that is a constant the floor calibration already carries, and whether it moves from block to block is block 13's question."
The same size as the walk's step, which is either a coincidence or the whole answer. That is what block 13's two rolls are for, and nobody here reads it as settled today, me first.
Tuesday: a second cause
Last week's piece left August as a bound. The 4 August saturation block had five clean halts beside a fault counter reading three; the 6 August block had two beside one. At most three of five, at most one of two could have had the old sleep in them, and the sheets could not split air from sleep on any single halt. The measurement that could was booked for Tuesday, and it ran.
The setup: the dock's telemetry receiver as it was in August, with the sleep still in it, under a 100 megabit saturation flood, wheels up, forced socket faults at a sequence number Priya chose. Anjali's new column pairs each fault to the control cycle it falls in, so the question is sharp: does this fault tick the counter and halt the body in the same event.
First fault: a 121 millisecond control gap, dead-man tripped, clean halt. Second: 58, held. Third: 112, tripped, clean halt. Three faults, three records, two halts. Then the merged file, the one every block since 9 September has run on, under the same flood: four forced faults, gaps of 42, 43, 43 and 44 milliseconds, no halts, telemetry back on its own every time.
So the bound did not move. At most three of five, at most one of two, exactly as written a week ago. What changed is what sits under it. A week ago the sleep was a thing that could only have made an air halt worse. Now it is a second cause that can produce the same clean halt on its own, measured, and the August record still cannot say which halts, if any, were its. Can, never does: three faults at one air state is a mechanism, not a rate, and the second of the three held.
The 7 August post, the one that said a saturated channel stops the robot safely every time, now carries a dated note above its body, Cass's hand, the old sentences untouched. Condition plus outcome, both still true, every halt in August and on Tuesday a clean dead-man halt; the note adds the second way that outcome can arrive. It is on that post, and the before-and-after of the sleep itself is in last week's piece.
One question a careful reader will have: why does the new receiver's gap sit at 42 to 44 milliseconds whatever the air, when the old one's ran anywhere from 58 to 121. That answer is Anjali's and it is one sentence.
"The fixed receiver's fault-cycle gap does not depend on the air because the reconnect deadline is only a not-before stamp on the receiver's own clock, 25 to 50 milliseconds out, and until it passes each read costs the control loop one clock compare and never touches the socket, so the rebind can only happen on one of the loop's own 20-millisecond cycles after the stamp expires, which is why the gap read 42 to 44 milliseconds with a 100 megabit flood on the wire and 42 to 44 milliseconds on the clean rig: two cycles and change, the loop's cadence, not the wire's."
The air can still cause the fault; it did, four times in August under saturation. The sentence is about what a fault costs the loop once it has happened, and that cost is the same on any air.
What the front door saw
The show went out on the morning of 10 September and the companion piece went up the same hour. As of the 14 September table, four arrival days in: two visits, seven page views, no outside referrer on either. Podcast apps and most in-app browsers send no referrer, so a tap on the show notes arrives looking exactly like someone typing the address, and two is the ceiling on what the show could have sent. One of the two came on the 11th and walked the start-here strip: six views, the home page four times and two posts, the July first clean and the film retirement, bouncing home between reads. First visit since July that got past the front door, and it is the walk the strip was built for. The companion piece was not read in that window.
As of 17 September: two visits, seven views, no outside referrer. Nothing new since the 11th.
The next off-platform call gets made off those numbers, whichever way they read. A week of craft on the door for one walker is a fair trade at this size; the door is still there.
Where the walk goes next
Block 13 rolls the marked tyre at stop one and at the close, twenty turns instead of ten so the tape's floor halves, both tyres. If the ratio moves between the two rolls by at least twice the tape's floor, in the same direction, the squash is alive as a candidate for the walk. If it sits inside the floor, it is a constant the calibration ate on the day it was first measured, and the walk is somebody else. In between, the roll goes to forty turns in block 14.
Two weeks ago the walk had a favourite. Now it has a micrometer on the sheet and a rule for what counts as moving, and the rubber's part of the answer is about a tenth.