Skip to content

Header-drift window (DIVE-3188) #586

Description

@github-actions

Rewritten by header-drift-window.yml at 2026-08-14 06:06 UTC.

A STALE CLAIM below means one named file overran its own # TIER: header by the
FAIL margin on two or more DISTINCT jobs — distinct ephemeral boxes, so the
slow-runner explanation does not cover it. Re-measure the header and widen it with
its environment and date; the replacement line is printed in the run log.

A single-box WRONG is NOT a stale claim and is listed as a stopwatch
disagreement, because that reading was measured false three times on 2026-08-10
(DIVE-3163) including a green re-run at the same sha.

tier-cal-window: stale-claim verdict (DIVE-3188) — 47 job(s) carried a graded drift count
tier-cal-window: no file was marked WRONG on any of the 47 graded job(s).

tier-cal-window: 47 run(s), median probe 143214us/iter, median corpus 1133333us/harness

RUN                     HARN WALL_S     US/HARN      CAL_US     POST_US  DELTA%  VERDICT
installed-host/core      297    288      969696      112420      112784      +0  concordant
pristine/core            297    351     1181818      143820      143835      +0  concordant
installed-host/core      298    323     1083892      139180      139263      +0  concordant
pristine/core            298    327     1097315      140816      139957      +0  concordant
pristine-confirm/core    298    394     1322147      168915      168355      +0  concordant
installed-host/core      298    362     1214765      146235      146014      +0  concordant
pristine/core            298    347     1164429      112988      114720      +1  UNPROTECTED
pristine-confirm/core    298    381     1278523      144985      146391      +0  concordant
installed-host/core      298    328     1100671      139253      139478      +0  concordant
pristine/core            298    303     1016778      113522      112897      +0  concordant
installed-host/core      296    326     1101351      144347      145101      +0  discordant
pristine/core            296    297     1003378      138888      138319      +0  concordant
installed-host/core      297    318     1070707      140633      140098      +0  concordant
pristine/core            297    317     1067340      140492      139253      +0  concordant
installed-host/core      297    305     1026936      139436      139352      +0  concordant
pristine/core            297    345     1161616      144794      145514      +0  concordant
installed-host/core      298    310     1040268      139309      138971      +0  concordant
pristine/core            298    307     1030201      137347      138819      +1  concordant
installed-host/core      299    330     1103678      144298      142522      -1  discordant
pristine/core            299    266      889632      123312      123562      +0  concordant
pristine-confirm/core    299    337     1127090      143275      141260      -1  discordant
installed-host/core      299    346     1157190      144838      144985      +0  concordant
pristine/core            299    345     1153846      112397      115556      +2  UNPROTECTED
installed-host/core      299    320     1070234      138666      138985      +0  concordant
pristine/core            299    342     1143812      145107      145696      +0  concordant
pristine-confirm/core    300    343     1143333      145956      145072      +0  concordant
installed-host/core      300    359     1196666      146485      146250      +0  concordant
pristine/core            300    340     1133333      112465      112863      +0  at-median
installed-host/core      301    335     1112956      143214      141571      -1  at-median
pristine/core            301    331     1099667      141514      141857      +0  concordant
installed-host/core      302    336     1112582      141614      142328      +0  concordant
pristine/core            302    346     1145695      145764      146323      +0  concordant
pristine-confirm/core    303    335     1105610      142242      141900      +0  concordant
installed-host/core      303    358     1181518      147119      147925      +0  concordant
pristine/core            303    378     1247524      147253      150373      +2  concordant
installed-host/core      303    337     1112211      142362      142000      +0  concordant
pristine/core            303    345     1138613      145470      144132      +0  concordant
pristine-confirm/core    303    362     1194719      148134      149000      +0  concordant
installed-host/core      303    346     1141914      143537      144343      +0  concordant
pristine/core            303    326     1075907      115517      116858      +1  concordant
installed-host/core      303    358     1181518      148208      148328      +0  concordant
pristine/core            303    350     1155115      145338      144397      +0  concordant
installed-host/core      304    359     1180921      147365      148047      +0  concordant
pristine/core            304    335     1101973      142101      142550      +0  concordant
pristine-confirm/core    304    358     1177631      147358      147805      +0  concordant
installed-host/core      304    348     1144736      144985      146720      +1  concordant
pristine/core            304    525     1726973      117904      119333      +1  UNPROTECTED

concordant 39 | discordant 6 (of which UNPROTECTED 3) | at-median 2
CONCORDANT = probe and corpus sat on the SAME side of the window median, which is the
relative budget doing what DIVE-2710 assumed. DISCORDANT = opposite sides. UNPROTECTED is
the discordant half that costs somebody a red: probe below median (reads FAST, so the
clamp floors the scale at 100% and the cap cannot widen) while the corpus sat above it.

tier-cal-window: 3 UNPROTECTED run(s). On each, the mechanism was live and applied no
relief to a corpus that was running slow. Check cal_post_us_per_iter on those rows: a post
probe SLOWER than its pre probe says the pre-probe sampled the wrong moment (TEMPORAL,
remedy = bracket the corpus); one that AGREES leaves the cost-mix reading (trap 3 in
tests/lib/tier.sh) on the table — weakly, because the corpus warms what the probe pays
for and that biases the post reading fast.

This body is rewritten every run. If the timestamp above is stale, the RAIL is
dead — that is a different failure from 'nothing recurred', and it looks the same
from anywhere else (DIVE-1924).

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions