The previous report derived its baseline arithmetically. This one runs the job
three times over the same 90 days against identical data, writing real summaries
each pass: production's calculation, then R1, then R2. 18,900 summaries per run,
56,700 in total. Window 2026-05-10 to 2026-08-07, deliberately stopping a week
short of the backup so the partial days around the cut cannot distort it.
Reportable population, 8,350 client-days across 96 clients:
baseline 1,024 days off, 87.74% clean, $58,531.75
+ R1 752 days off, 90.99% clean, $55,528.17
+ R1 + R2 105 days off, 98.74% clean, $852.38
R1 8,054 unchanged, 272 to balance, 0 to unbalance, 0 balanced-day edits
R2 7,701 unchanged, 647 to balance, 0 to unbalance, 0 balanced-day edits
Across 90 days of trading neither fix altered a single day that already
balanced, replicating the 30-day finding on three times the sample.
The longer window surfaced two residual clusters the short one could not see:
NGMV across five days in late May and NGEB across four spanning May to July.
NGEB is the known ezCater fee question; NGMV and N-30012 are undiagnosed and
worth a look before this ships.
Also records a deployment finding: the transactor was sized for a toy database
(2 GB object cache against 27 GB of data) and the final 7,958 client-days of a
pass were crawling at ~35/min. Resized to 8 GB cache and 16 GB heap, they
finished in 90 seconds. Worth checking production's sizing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Bryce
89baac19f3