The Night Shift That Taught a Simulator to Tell the Truth
The warehouse clock ticked loud in the night. One team packed aid kits in a bright hall, but most supplies sat in a far hangar. On a laptop, a coordinator slid four big controls, worker speed, handout speed, hallway width, and the long road speed, trying to make the pretend drill match the clipboard times.
Everyone trusted that laptop for planning, but almost nobody wrote down how they set those sliders. When someone did, it was usually a tired expert guessing, checking, guessing again. It could take days, and small details, like a staging table near the line, could quietly throw it off.
This team turned tuning into a timed game. They set safe ranges for each slider, picked a single score for “how close to real,” and started the countdown. They judged it at many checkpoints, like average finish times at several stations, under different starts where the hall had more or less stock already nearby.
Instead of one person twisting knobs, they ran lots of quick tryouts at once, like a crowd of helpers each running a mini-drill and reporting a score. They tried three ways to search: tighten a grid, toss random guesses, or take small steps that seem to help. They also jumped in doubling steps across huge ranges so they did not crawl blindly.
They tried it on a tough two-building routine, lots of separate jobs, each hauling big files from far away, doing work, then writing results. They reran the real drill with a slower or faster long road, and with the staging table allowed or banned. The surprise was that the automated search often beat careful human tuning, especially when the staging table was in play.
Then a trap showed up. If one choke point ruled the whole night, like a slow handoff at storage, many slider mixes could “fit” the same final times. The searches often agreed on the choke point, but disagreed wildly on the rest, so tomorrow’s different drill could break the illusion.
Two practical lessons stuck. Using only a few very different starting setups could work fine, since each check was cheaper and let them try more slider settings. And a faster, rougher pretend drill could tune better, just because it let them test more options before time ran out. The clipboard stayed the judge, but the tuning stopped being a secret craft.