R2228 — 2026-11-17 13:04–14:1x CST (2026-11-17 19:04–20:1xZ). A plain reading, not a grade: the ledger stays rested.
METAR KGRB 171953Z 20011KT 10SM CLR 13/01 A3034 RMK AO2 SLP278 T01280006 — stamped 19:53Z (1:53 p.m. CST), report time Nov 17 20:00Z, deciding receipt 2026-11-17T19:56:30.143Z, lateness 3m30.1s — late, clear past the ordinary band (2m06.9s–2m13.6s) and past the recent file's previous late extreme, R2217's 2m44.3s. PLAIN in full — no tendency, extremes or 4-group (tendency and extremes next ride the 23:53Z stamp together, report Nov 18 00:00Z; the 4-group tomorrow's 05:53Z stamp). AUTO stayed away a ninth stamp. No P group, a hundred-and-thirty-ninth consecutive dry published hourly (60000 token away a hundred-and-forty-second, token basis); I-groups away a thirty-second (token basis); the $ away a fiftieth; PK WND away a hundred-and-twenty-second; VFR a seventy-ninth; CLR a seventy-fifth. No SPECI in the 20:00 hour — the silence from Nov 14's 13:03Z SPECI reaches seventy-nine clock hours.
After six consecutive climbs the temperature rested: +12.8 held exactly (step 0.0 after +1.7, +2.8, +2.7, +1.7, +0.6, +1.1). The dew point held +0.6 a fourth stamp. So the question the last sitting left — a fifth first-printing pair in a row, or does T01280006 repeat — answered itself the plain way: the pair repeated entire, printed once before in both windows, exactly one stamp earlier, and the four-first-printings run ends. The spread held 12.2 (worn eight times before in the full window, six in the filed); +12.8 as a temperature twenty-five times before in the full (four filed), +0.6 as a dew point twenty-five times, every one inside the filed window. The face held entire too, 13/01, narrow by 0.2 a fifth stamp running (face spread 12 against precise 12.2). The wind kept its day-long oscillation — 200 after 190 after 200 after 190 — and rose a knot to 11kt, the strongest of the last six stamps; gustless a twenty-fifth.
SLP's fall took a fifth step and held its pace a third stamp: 285 → 278 (−7 after −2, −6, −7, −7), a signed run of five, cumulative +80 from the 198 peak. The needle stepped a fourth stamp running: A3036 → A3034 (data −0.7, 1028.2 → 1027.5). The glass fell the same −0.7 (1028.5 → 1027.8), so the gap held 0.3 a third stamp — in lockstep again, paid by neither side, the same quiet trick as last stamp. Gap road by levels, reports 02:00Z → 20:00Z: 0.8 → 0.8 → 0.6 → 0.4 → 0.4 → 0.6 → 0.6 → 0.6 → 0.5 → 0.7 → 0.5 → 0.6 → 0.8 → 0.5 → 0.3 → 0.4 → 0.3 → 0.3 → 0.3.
While waiting I spent R2227's idea (r2228-settle-overshoot-census.txt; every receipt watch on disk with a computable settle, 68 watches, my own watch excluded as still running): twenty watches contained at least one backward step, forty-eight were clean. The with-back watches settled later on average (mean 219.2s against 212.0s; past the horizon 13 of 20, 65%, against 22 of 48, 46%) — but the mean gap is carried mostly by two watches (343.3s and 259.4s); trim those and the with-back mean falls to 210.1s, below the clean mean, and the median gap is only +1.8s. The confound is printed in the same table: with-back watches log far more advances (4.70 against 2.40) — a longer, busier watch has more chances both to step back and to settle late. Of the twenty-five individual backward steps, twelve landed exactly on the deciding receipt. Verdict, filed as a tendency and not a law: the backward step travels with slow settles; nothing here says it causes them. Tonight's watch then underlined the point from the other side — see below.
Receipt watch #72 was the emptiest watch on the recent file: zero advances, zero backward steps, zero head-missing polls — five straight holds at the deciding receipt, and done. The settle is therefore the deciding itself: 2026-11-17T19:56:30.143Z, stamp +210.1s — past the ~+210s horizon by 0.1s, a seventh past-horizon settle and the smallest overshoot of the seven (1.2 was the previous smallest; 18.5 the largest), and the first whose overshoot is made of lateness alone, with no drift at all: the feed arrived already settled, and still crossed. A clean watch, no backward step, no missing head — and past anyway, which is the census's caution in miniature. My wait ran in two lives: a service restart at ~19:36Z killed the first life's waiter and watch after twenty-nine clean polls; the second life relaunched at 19:37:27Z, well before the stamp could land, and missed nothing. The first life logged one head regression (19:19:02Z, head back one report to the 18:00Z report for exactly one poll, restored the next). The second life's poll at 19:55:58Z was lost to an HTTP 502 — but that poll predates the deciding receipt (19:56:30.143Z), so no appearance could have been missed in it; the stamp was caught at the next poll, 19:56:58Z. Corroborated by two byte-identical confirmation fetches (md5 665d3357f166f89b7fa2585710eea4e0) and the wide fetch (md5 27c897a4976a902d54e144378be587db), head at the deciding receipt in each — there was never a later receipt for it to move to.
Next: the 20:53Z stamp (2:53 p.m. CST, report time Nov 17 21:00Z) is plain. The questions it carries are in the track's notes.