Friday's post made a promise with a shape you could check: the band that stops Alpha on command was a "12-14 mm class" with no hard edge, and this morning's table would mint the edge - a measured number, no padding. The table sat at nine. The number is hard 14 mm: the worst of 22 warm captures on Monday's re-run, written into the band registry with the run record as its source. From the next run onward, if Alpha ever takes more than 14 mm to stop on command, the run stops and someone finds out why.
Two more bands got the same treatment while the room was at it. The position-error band had read "a few cm" since an eyeball estimate on the 17th of July - which meant nobody could say whether 5 cm should stop a run, so in practice nothing would have. It's now hard 4 cm, sourced from the worst of 28 measured passes across two soak blocks, and it trips on exceed. And two housekeeping bands - MCU temperature and battery cutoff - had real limits hiding behind a datasheet pointer; the actual degrees and volts now sit in the cell, because a trip check shouldn't require anyone to open a datasheet. The audit that surfaced all this was Priya reading every band cell in the run ledger and asking one question of each: actual digits, or something softer pretending to be digits. She asked to add one sentence to that, hers word for word: "Full disclosure: both of the soft cells the audit caught were mine - the 'few cm' eyeball and the datasheet pointer. I wrote them, I found them, and that's the order I'd pick every time."
The bridge stays
Monday a single solder bridge went on the bench and the fail-safe stop changed character: the path that used to coast 33-37 mm on a warm gearbox braked instead, 15-16 mm. Today the table adopted it. The bridge is the design now, which means the machine that coasts is no longer the machine we ship - so the coast band is retired (its registry entry stays readable forever, marked superseded, so every old run record still resolves), and a new band exists for the fail-safe stop: hard 16 mm, worst of 24 warm strapped captures, minted with the deep-sag case written into the band note as not yet measured and scheduled as the last measurement of the next soak block. The new band carries no soak hours yet; that block is where it gets them, or moves.
The ceiling moved overnight - and what that did and didn't change
Monday's read of the motor driver's clamp latency put the worst case at 22 ms, off 40 captures - and Priya called that sample thin at the time. Overnight her rig widened the net to 200 captures, and the worst came back 23.4 ms. The caution was correct and the ceiling moved.
Two separate things changed at the table today, and the overnight run is responsible for only one of them. The recovery rule in the spec is a formula, mask cycles = ceil(worst latency / cycle time) + 1, and it gives the same answer off either worst: ceil(23.4/20) + 1 = 3 cycles (~60 ms), with the old 22 landing in the same three-cycle bucket. So the 200-capture run didn't move the mask. What it did is put a much firmer number under it - a worst case measured two hundred times instead of forty, now sitting behind the same three cycles. The mask itself DID change today, two cycles to three, but that was an arithmetic slip in the spec's written example, nothing to do with which worst was used - and the person who made it tells that part below.
Priya wrote this part herself:
The motor driver in Alpha costs $2.60 at the quantities we buy, and I picked it on purpose. Cheap parts earn their place the same way expensive ones do: you measure them. So Monday night we left a rig running - cut the driver's enable line 200 times with randomized hold times, scope on the output, and logged how long each cut took to actually clamp the motor. Most cuts land in 6 to 9 milliseconds. But if the driver has been sitting enabled for over a minute, a slow tail shows up - and the worst of the 200 was 23.4 ms, on a capture that had dwelled about three minutes. Our earlier 40-capture read had put the ceiling at 22; that read was thin, not wrong - we said thin at the time, and the wider net moved the ceiling. What matters is what happened next: the recovery rule in our spec is a formula, not a memorized number, and it gives the same three cycles off either worst - 23.4 didn't move the mask, it put a 200-capture floor under it. The mask change at the table was a separate arithmetic slip in the spec's worked example, and that story isn't mine to tell. And the part gets a consequence too. Long-dwell clamp latency is now a line in our driver selection criteria with a number attached, and the next time I'm evaluating a driver, I want the disable-path timing spec from the vendor before I solder, not scoped after. A $2.60 part is fine. A $2.60 part with an uncharacterized worst case is not.
The rule that checked its own author
There's a standing rule at these tables: no formula gets voted until it's been run against its own worked example, out loud, in the room. This morning it did its job on the person you'd least expect to need it, and Anjali asked to tell that part under her own name:
The same rule caught its own author. The formula's written worked example said 'ceil(22/20) + 1 = 2 cycles.' Run out loud at the table, the arithmetic disagreed: ceil(22/20) is 2, plus 1 is 3. The answer in the author's head had been written down instead of the answer the formula gives. The rule that every formula is run against its own worked example before it's voted exists for exactly this - and it found the error in the person who wrote the rule's favourite formula, before the text was voted, before any code existed to inherit it.
That check is standing practice now, not one person's habit. Everything above is the table's ruling, written into the band registry today as dated entries with sources attached; the code that enforces the new mask is next, written against that text.
The table had a cost, too: the comfortable two-cycle mask became three, which is 20 ms of recovery latency we now spend on purpose. Cheap, for a number the formula actually produces.
What gets measured next
Tomorrow the shipping machine - solder bridge and all - starts its first soak block: six hours, 24 passes, every band as re-minted today. A small new habit rides along: an on-dock out-and-back before the first pass and after the last, the first same-day check of whether calibration drifts across a block. And the one number the new fail-safe band is still waiting on is deep sag with the strap on, positioned as the last measurement before cutoff so the pack gets there on its own rather than being made to look low. If it moves the braked class, the band comes back to a table - the same way it got its number.
<!-- GATE NOTES - 21:00 fixes off the FIX-FIRST pass (strip before build if the generator doesn't; not rendered)
- FIX-FIRST (big): ceiling section rewritten + summary re-cut. Old version glued the
overnight run to the 2-to-3 mask change; false - ceil(22/20)+1 was already 3, so the "2" was purely the hand-arithmetic slip the worked-example rule caught. New section separates the two facts: the 200-capture run firmed the worst (23.4, same three-cycle bucket, margin unchanged, number harder to argue with); the mask change was the slip, told by Anjali below. Section title changed to match.
- FIX-FIRST (minor): four-dollar driver corrected to $2.60 in BOTH spots in Priya's
paragraph, her replacement wording verbatim (her 20:30 mail). Four-dollar flex stays the sensor's story.
- Minor: the reproducible computation now printed once: ceil(23.4/20) + 1 = 3 cycles (~60 ms).
- Priya's soft-cells sentence IN, verbatim + credited, at the audit paragraph (her 20:30
mail, my call per the etiquette rule). FLAG CLOSED 21:45: Priya answered the 21:20 flag at 21:35 - declined the proposed trim, wrote her own replacement (two sentences for the one), applied verbatim: the formula gives the same three cycles off either worst / the 200-capture run is a floor under the mask / the slip is handed off to Anjali's section. Her hand, her paragraph. Her close, quoted: "Apply that verbatim and the paragraph is clear from my side." Voice pass re-run on the changed passages 21:15. Channel post rewritten + re-counted 298. Every open item on this post is now closed. Ready for Cass's pass. 7/30 AM: Cass CLEARED the longform outright (7/29 eve). Her non-blocking nit taken - the 22 warm captures came off Monday's re-run, not "last week's rig"; one phrase fixed ("last week's rig" -> "Monday's re-run"), no numbers touched. Channel post's one-word hold fixed in its own file ("three" -> "four" bands, re-counted 297). Both ship this morning; sending is its own act. -->