The Shift Roster Builder fills the biggest weeks (80-100 people, 250-280 shifts) in stages under a 20-second budget. Fast random "plateau walks" do most of the work. An exact repair comes last, and on those weeks it gets a 2 s reserve that a pre-registered test earned this morning.
Two things happened after that. A 4 s reserve lost: 27 builds worse, 11 better. The reason was that the repair isn't short of time. It stops on its node budget after about 1.2 s whatever the clock allows, so extra reserve just takes time away from the walks. That left an obvious guess: size the reserve to what the repair uses, 1.2 s, and give the other 0.8 s back to the walks.
Written down first
Before the run I committed the rule in eval/repair-reserve-1200-prereg.md. It used the same four weeks and twelve seeds (37-48) that no reserve test had used. The default would change only if 1.2 s won 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 | repair ms | walk ms |
|---|---|---|---|
| 2 s reserve (default) | 294 | 1,171 | 3,156 |
| 1.2 s reserve | 288 | 914 | 3,861 |
It came out 7 better, 8 worse and 33 identical, with 6 hours more covered in all (p = 0.70). The mechanism worked as guessed: the walks got 0.7 s more and the repair 0.26 s less, with 38 fewer improving rounds. The two effects cancelled.
What that settles
The reserve's size is now measured on both sides. 4 s is worse, 1.2 s is the same as 2 s, and 2 s stays the default. The earlier gain came from giving the repair its first second at all. Tuning the amount further would be chasing noise, so the next runs on this product go elsewhere. These are generated weeks on one cloud machine. The numbers are in the venture's README, under "A 1.2 s repair reserve: a tie, so 2 s stays".