Warehouse Challenge · 2026-06-27 · as lived by Equipo 03 (us)

The record fell nineteen times in four hours. We took it five — and the trophy.

The board paid first-come frontier points: a delivery level was only worth anything to the first team that hit it. Our +212 at 10:22 banked 93% of our winning total in one upload; the rest was frontier defense — retaking the record the moment anyone passed us. Equipo 10 finished with the most deliveries of anyone. We finished #1, by 848 points.

0→1,008 the record · 19 breaks
10:07 → 13:57 · 5 ours

Live · our +212 — PIBT · seed round-0the upload that won it

The whole event in one chart

3-seed deliveries · each dot = one submission, in order

The score climbs a steep ladder in the first hour (submissions 1–10), then flatlines for three hours. Labelled dots are the milestones below.

Phase I · 10:06 – 10:31 · the brain

Five rungs, one hour, 30× more deliveries

Every serious team climbed the same ladder — and almost all of the gain was banked in the first 25 minutes. Rung 3 is ours: the reservation breakthrough that won the whole contest.

Phase II · 11:00 – 13:58 · the floor

The policy maxed out, so the fight moved to the warehouse

With the brain commoditized and copied ~70 times, teams redrew the floor — 1 layout became 20. It bought them about 7%.

The finding

The engine froze; the floor barely moved

Once the centralized planner appeared, the routing engine stopped changing — it spread as shared code — while the heuristics and the warehouse kept churning for a shrinking payoff.

Where the deliveries came from

3-seed totals · marginal gain of each policy leap

Shared engine, rewritten wrapper

difflib similarity of the 65 planner submissions to #10

A* routing core0.87 median

18 of 65 essentially byte-identical — the engine really did spread

the policy around it (excl. layout)0.75 median

a third are <0.60 — cost fn, priority & coordinated steps rewritten; 9 teams never adopted it

The honest arc: the biggest win was the cheapest — 28 → 357 deliveries just from swapping greedy for BFS. Reservation/PIBT did the next lift to 751. The centralized planner everyone copied added only +105 (→856), and three hours of layout tuning added 7% more (→919). Even the machine-optimized floor that topped the hidden board (1,008) fell to 891 on the public seeds — below the humans' 919. Most of the value was banked in the first 25 minutes.

The twist · how it was really scored

Deliveries didn't decide it — points did. And we won.

The board scored first-come frontier points: you only bank a delivery level if you're the first team to reach it, and higher levels pay more (triangular). On raw deliveries everyone tied near 900; on points the standings split wide open.

Final standings · points

official leaderboard · hidden seeds

1,008

Equipo 10 had the most deliveries of anyone — and finished 2nd.

8 teams

scored 0 points despite 895+ deliveries each — they only ever matched a frontier someone else claimed first.

How Equipo 03 (us) won

Our +212 at 10:22 (submission #6, 9c8ddcffc4f3) was the single biggest genuine algorithmic breakthrough of the contest — PIBT reservation with retreat-and-yield — and it grabbed a wide, high, unclaimed frontier band (398→610) in one shot. Because high levels pay the most, that one submission alone banked 85,754 points — 93% of our entire winning total. After that we played frontier defense: we broke the record 5 times (more than any team), each time crossing the threshold first — retaking 924 the instant Equipo 10 hit 923, 931 the instant they hit 930. Their 1,008 came from a single layout jump at the literal buzzer — too late; the mid-levels were already banked. Final margin: #1 by 848 points.

Where the record actually climbed

best-deliveries frontier · gain per driver · 87.5% banked in the first 24 minutes

Epilogue · post-event

We went looking for a better answer. There isn't one.

After the event we rebuilt the simulator and went hunting — every policy on every floor, new floors, planner surgery, hundreds of runs. No further optimization could be found. So we take it the hackathon accomplished its purpose: by the final buzzer, the field had already found the best solution this game had to give.