On Wednesday we wrote down, in advance, what counts as failure on a long run: ten limits, each with a number and a source behind it. On Thursday, day one of using that rulebook, pass six of twelve, one of the limits tripped.
It was the best thing that happened all week.
What tripped
Some background. When Alpha loses its control link, it doesn't brake - it cuts drive and coasts out on gearbox friction alone, because a fail-safe that depends on an active system isn't a fail-safe. We'd measured that coast distance cold: 31 millimetres, and Wednesday's rulebook set the limit right there, no padding.
Thursday, mid-block, a scheduled coast check came back at 34 to 37 millimetres. Six stops, all past the band. The run stopped, as written, and Priya spent forty minutes on the rig finding out why.
Priya's part, in Priya's words
The finding is hers, so here's her telling it. I didn't get to touch this paragraph:
"The band stays cold because the band's job is to trip. We set the coast limit at 31 mm - the number we could actually defend, no padding - and a warm gearbox rolled 34-37 mm and tripped it. Forty minutes on the rig turned that from a scare into a measurement: 3 mm spread, one direction, every stop, and the commanded braking completely untouched. Random is a fault; tight and one-way is physics. The coast stop is friction-only, warm grease is thinner, so the machine rolls a few extra mm before friction wins. Worst coast 37 mm against 130-150 mm of clearance. If we'd padded the band to be comfortable, that number would still be a secret."
- Priya Nair, hardware
Cold numbers, warm machines
The bit worth sitting with: the band and the gearbox weren't disagreeing about the truth. They were measured at different temperatures.
Every number in our rulebook was pinned somewhere specific - a cold bench, a datasheet, a short run. A machine four hours into real work isn't that machine, and its friction-only stop rolls a few millimetres further. If you've ever noticed a car feel different in winter, you already have the intuition: same machine, different physics regime.
Why does a warm gearbox roll further? Priya's working read - it gets more soak blocks before anyone calls it settled - is the grease: thinner warm, so a stop that relies on friction alone travels a few extra millimetres before friction wins.
For scale, the worst case added up the way our worksheet adds it: the link can be silent for up to 70 milliseconds before the stop even begins, and the machine covers about 14 millimetres in that window. Add the worst warm coast, 37, and the total travel is around 51 millimetres - against 130 to 150 millimetres of clearance. Roughly 80 to 100 spare. Nothing came near anything. What tripped was a number crossing a line we drew on purpose, tight, so that a three-millimetre shift in the machine's behaviour couldn't slip past. It didn't. A padded band would have let it.
The reading we didn't get
One more thing went on Thursday's sheet. The forty-minute diagnosis ate the time slot reserved for the reflex-latency spot-check - the measurement that confirms the emergency stop still fires inside its 70 millisecond budget. So that row on the paper record says NOT-MEASURED, and it stays that way. We don't backfill a blank cell with "it was probably fine."
Friday morning, that check went first, before the day's first pass: 40 to 46 milliseconds against the 70 budget. Measured again at the end of the block - same class. The blank row became a number with margin, one day late and labelled the whole time.
Then Friday happened
Block two ran today, and it ran long: sixteen passes, four hours ten minutes, next to yesterday's twelve and three hours five.
Before it rolled, Anjali's morning test suites caught one real bug in the run ledger - a deleted run could have had its number quietly re-issued to a later one, which would eventually have corrupted the bookkeeping every hour of soak data depends on. Fixed and green before a single pass counted. The pack sat charged from 09:00 and waited for that green; hours don't start until the ledger's trustworthy.
Then the day confirmed Thursday. Two more warm coast checks came back at 33 to 36 millimetres - same tight spread, same one direction. Both formally tripped the cold band; both stops took minutes to confirm against Thursday's record - Thursday already paid the forty minutes, so Friday only had to check the match: same spread, same direction, commanded stops untouched. Late in the block, with the battery deep in sag, the coast measured 35 to 37 - inside the same warm class, no new behaviour under low charge. Commanded stops read 12 to 14 millimetres all block, including under deep battery sag - and that's stated as a measurement, not a verdict, because reviewing this piece caught something the table missed. The coast band got a hard edge on Wednesday; the commanded band only got a "12 mm class", a description with no line that can trip. It shouldn't be that way, and at next week's table it gets a hard number of its own - no padding, same deal the coast band got. The link stayed clean throughout: worst gap 54 milliseconds, nothing over the 70 window, zero unattributed halts - which now makes over seven real hours without one.
Running total after two days: about seven and a quarter hours, twenty-eight passes, one genuinely new fact - the warm coast class - and zero surprises the rulebook didn't have a slot for. The question the soak exists to answer is how long a clean run stays clean. So far the most interesting answer is a few millimetres of warm coast, found on day one, by a limit that refused to be comfortable.