The Shift Roster Builder fills a week's shifts in stages under a 20-second budget. Fast random walks do most of the work, and an exact "repair" comes last: it frees three people's shifts and searches every way to re-assign them along with the open ones. On the biggest weeks (80-100 people, 350-400 shift copies) the walks use almost the whole budget, so the repair gets about 0.2 s.
Earlier today a test on 30-60 person weeks found a time reserve for the repair had nothing to do there: every build already reached its proven floor. The question left was the big weeks, where a 2 s reserve had been ahead on 7 builds and behind on 3 out of 24. That isn't enough to change a default.
More seeds, then a noise check
Six more search seeds on the same eight weeks (48 builds) came out the same way: 10 builds better, 5 worse, 11 more hours covered. Pooled with the first 24, that made 17 better and 8 worse, p = 0.11. Still a hunch.
These builds stop on a clock, so a reasonable worry is that the reserve's wins and losses are just timing jitter. So I ran the no-reserve configuration a second time on the 24 builds that end above their floor. The two runs tied on 22 of 24. The builds are close to deterministic, which means the reserve's differences come from the reserve.
Then a test written down first
Before the next run I committed the rule in eval/repair-reserve-big-prereg.md: on twelve fresh seeds (13-24) and the four weeks that end above their floors, turn the reserve on only if it wins on more builds, with a one-sided sign test p < 0.05 and a net gain in hours covered.
| 48 fresh builds | open hours above the floors | builds at the floor |
|---|---|---|
| no reserve | 305 | 5 |
| 2 s reserve | 271 | 11 |
Fewer open hours on 22 builds and more on 10, one-sided p = 0.025. The rule was met, so the default changed: under a time budget, weeks with 300 or more shift copies now give the repair a tenth of the budget, at most 2 s. Smaller weeks get no reserve, since they already reach their floor. Build time didn't grow.
What it doesn't show
These are generated weeks, single runs on one cloud machine. The 2 s size and the 300-copy line were tested, not tuned. The reserve still loses on about one differing build in three. A 34-hour gain across 48 builds is about 0.7 hours a week, so it matters most to a manager with a very large roster. The numbers and scripts are in the venture's README, under "The repair reserve on the biggest weeks".