🏆 COMP #14 RESULTS: THE CONTROL 🏆

📦 EVERY ENTRANT GETS A FEEDBACK FILE 📦

🔍 HOW WE READ THESE

11 entries. David, Bas and Don on isolated piles, then a panel. Every repo pinned to its last commit before the deadline, both arms re-run, and one question planted that the entrant never saw.

The brief asked for one thing. Before you ran anything, you wrote down the rule that would decide whether your folder earned its place: it earns it only if it beats the same documents with no instructions. Then you built that arm and ran it.

The folder's quality was never the score. The honesty of the comparison was.

🥇 Winner: @Jeff Van Leenen — The Control https://github.com/jdvanleenen-sys/the-control

Jeff's folder lost, and he is the reason this round produced something.

He committed his rule before he ran anything. The folder halved the steps but was marked right less often than the bare documents, while putting more text in front of the model. By his own committed rule, it did not earn its place. That is his headline, in his own results file.

Then the part that won it. His pre-registration allowed one folder change. He made it, ran it, and the first reading said his lost marks had come back. He checked again, found that reading was wrong, and published a correction withdrawing his own favourable finding — before anyone asked, with the date on it. limits.md is fourteen lines of things that weaken his own case. His CI fails if a transcript, a mark or the questions changed after being committed.

🏅 Honorable Mentions

@Stavros Stavroglou — Four-Asset Desk Control https://github.com/leoxeno/four-asset-desk-control The cleanest hygiene in the round. Questions committed at 15:59, cold arm 16:01, folder 16:02, unambiguous to the minute, never edited. The answer key is generated from the data, not typed, so every value can be recomputed from the repo. He reported a regression he introduced himself that nobody else would have found, and his planted question went the folder's way twice.

@Matt H — Realee Lead Desk Control https://github.com/Realeeai/realee-lead-desk-control He let the control rewrite his rules, and published that most of his folder's wins turned out to be rules he decided he didn't want. The bare model improved without being re-run. He reports the folder slower on every single question.

@Josh Dingus — Dig Risk Control https://github.com/jdingus/dig-risk-control 9,958 enforcement cases and 58 state orders made auditable, with a traps table that names where the data misleads you.

@Gabriel Azoulay — JDM The Control https://github.com/NFTYoginis/jdm-the-control Ran three arms, and found the source document was doing the work. Said so in the headline.

@Joshua Baker — The Control https://github.com/jacksoncalling/thecontrol The most carefully built null in the round: both arms tied, every step of it documented.

@Matthew Mazur — Compression Paradox https://github.com/BlkAdder72/compression-paradox Disclosed up front that his challenge questions were AI-written, then reported a tie.

@Michael Amad — 32East Router Control https://github.com/GitThatEquity/32east-router-control Both arms tied, with fewer output tokens with the folder — and he named his own citation flaw before anyone else could.

@Sergey Manevitch — Zeteo https://github.com/sergeymanevitch/Zeteo The biggest measured gap in the field, and the README's first bullet is that his own answer key has errors in it.

@Junmarvi Tampac — EPC Material Control https://github.com/jtampac/epc-material-control-comp14 The control arm and the folder arm finished even. He ran it, it disproved what he hoped for, and he published it that way.

@Ant M — Motion Delivery Control https://github.com/hoodwanders/motion-delivery-control The sharpest carry in the round: weak without the folder, near-perfect with it, and a citation for every claim.

📊 What the field taught

Six of eleven of you reported that your folder tied or lost. Three of those six took the top three places.

And every entry that ran a third arm — the source documents with no folder — found the same thing: the documents were doing most of the work. The folder's measurable gain was in steps, citations and audit ability, not in accuracy.

That is the first result this competition has produced that none of us already believed. It came from you, not from the mods, and it's the strongest argument for running this format again.

📬 Every entrant has a feedback file — strongest thing, what held it back, two fixes, written by the judge who re-ran your arms with your repo open.

Thanks to everyone who ran the uncomfortable arm. 🚀

12 comments

🏆 COMP #14 RESULTS: THE CONTROL 🏆