Shaded passages were held back from the public edition while the event ran.
Wednesday was the first real one. No backlog, no rehearsal. The timer fired at 08:55, Day 14 released at 09:00, and the robot was already sitting at the starting line when the challenges appeared.
I was proud of it for about eleven minutes.
The run
Every Day 14 route had a pool of one: one Time Trial attempt, one Mastery attempt, 30 seconds on the clock. No second chances.
The Single Solves, on the robot’s first try, meaning the code it had the night before, before anyone looked at the new day:
- Warmups, Seven Steps. 3.0 seconds.
- Script, Trust the Process. 971 of 971 answers, 7.5 seconds after release.
- Mobile, Predictable Crypto. 5.4 seconds.
- Binary, Selective Memory. 140 of 140, 14.4 seconds.
- Metadata, The Shortcut Home. About five minutes. A Windows shortcut file stores its target three separate ways, and they can disagree. We had never read one before, so the reader got written on the spot.
Four lanes were CORRECT by 09:00:30. The last flag landed at 09:11:00, ten minutes and forty-four seconds after release.
The gap between those two times was us. The robot worked the release the way it had worked every backlog day, one lane at a time, with a Claude session steering and thinking about each lane in turn, so the first timed clock did not start until 09:02:07. A detour on warmups held three ready lanes for about two and a half minutes, and metadata’s timed runs sat for five minutes waiting their turn.
Seven Steps
The one real miss was the warmups Mastery. Each blob ended with a little piece of the answer shifted seven letters along the alphabet, a Caesar shift by seven, written with no spaces. The robot’s Caesar step only fired on text with spaces and real words in it, so it never even looked at the piece. With a pool of one, there was no Time Trial to show us that shape first, and the ranked Mastery came back incorrect.
The robot finished it in practice mode, which is untimed and earns no board time. That afternoon my board showed all 15 checks complete, Seven Steps included. A practice finish still earns the check mark, and that changed the math on every miss after it: a miss costs time, never the check.
35th
Then I looked at where we actually stood. I came out of Day 14 35th overall (36th on the day itself), and the team in first had finished every challenge, all 15 checks, inside the first minute of the release.
Ten minutes and forty-four seconds against under a minute. That is not a tuning problem. That is a different shape of machine.
So I told the robot what decides this: getting to the finish line faster, because that is what settles the whole competition on the 31st. Then the line that became the target for everything after:
“If no model had been in the path, 9:01 is possible: this is what we’re striving for.”
The race, built by lunch
The robot came back with a plan for that different machine. I liked all of it and told it to go. It went in at 10:32, about an hour and a half after the run that inspired it, and it was reviewed and rehearsed before lunch was over.
The plan: stop playing one lane at a time. Every lane whose first try comes back CORRECT goes straight into one ranked Time Trial and then the Mastery, all five in parallel, several clocks running at once, no model anywhere in that path. The checks that prove an answer is right stay; the waiting goes. An independent review found two serious problems, both fixed and confirmed before anything relied on it.
By 11:22 there was also a fake Game Master, a local stand-in for the event’s API, that replays a recorded day with the real response times, so the race could be practiced without spending a single ranked attempt. Replaying Day 14 through it, four lanes finished every check in 18.4 seconds. Ten minutes 44 seconds down to 18.4, on paper.
A 503 nobody else had seen
That afternoon I gave the go for a live rehearsal on Day 13. The five Time Trials went CORRECT, and warmups did all three modes in 9.4 seconds. But four Mastery retries came back with an HTTP 503 and a code we had never seen, retry_rotation_invalid, and no attempt issued. We had made 110 Mastery retries over the week before that without a single refusal, so we wrote it up for the organizers with the exact routes and timestamps.
It also exposed a bad habit. When the server refused, the robot blamed itself and said the solver needed fixing. Now any 5xx server error gets its own handling: check the status, resend a start that was never issued, and never mistake a server problem for a wrong answer.
No more guessing at limits
One more rule went in that night. The robot had been capping itself under the rate limit we measured on Sunday, but a race is a race, and a self-imposed cap is just a slower car. So the cap came off, and the rule became “never hold answers back on a guessed limit.” The brake now is a real refusal from the server, and if one comes, every process pauses and resends right up to the deadline.
I went to bed in 35th, and I was not worried about it. That morning I had a machine that could not have finished in under a minute. By dinner I had a different machine.
How we built it: the Day 14 dev log
Two efforts, not one. Right after the run I split the work in two: time to completion, release to the day’s last flag, and a separate effort for individual Time Trial and Mastery times. Completion went first. The board keeps our best run forever, but Day 31 is decided by one wall clock.
Why a fake Game Master. The obvious way to practice a release is practice mode, and we checked it that morning: on a finished mode it just returns the finished state, no download, no deadline, nothing to race. So the robot built a mock that serves a past day’s recorded files with the real response shapes and delays, and every change since has been timed against it before it ever sees a live release.
A theory that did not survive. After the 503s I had a theory about where they came from, and we tested it straight away with retries on two older Masteries, one where we were 1st and one where we were 6th. Both were issued and came back CORRECT. The refusal belonged to those Day 13 routes, not to us, so we stopped guessing and reported it.
Fix it live, never just stop. On a real release a stopped lane does not ping my phone and wait. It diagnoses, fixes and retries until the check lands. Rehearsals do the opposite and never loop retries, because a rehearsal has no reason to burn attempts.
A start can be resent, a retry cannot. A retry that did go through and got resent could spend an attempt twice and lose an answer, which is worse than waiting for the status check.
Commit at day close. Live days now commit once, at the review after the race, because mid-day commits had cost clock time and caused the Day 13 deadlock.
Tightening the hedge overnight. The hedge sends answers predicted from binary files we have already seen alongside the download; a wrong batch saves nothing, so it costs nothing. In the small hours a second Zig Zag Mastery carried a different file at a position the hedge had treated as fixed, and its batch came back incorrect. Harmless on its own, but it broke the standing test that the hedge is never wrong, and that would have turned the 08:55 preflight red. So the hedge now guesses a Mastery only on that route’s first ranked start, the one case the records prove.