🏆 COMP #15 RESULTS: THE REBUILD 🏆

📦 EVERY ENTRANT GETS A FEEDBACK FILE 📦

📄 Your file is here: https://www.skool.com/cliefnotes/classroom/7634b927?md=2b3ccc0b73c243b7b6137993f6b60da3

🔍 HOW WE READ THESE

We cloned every repo and pinned it to its last commit before the deadline. Nobody committed after the clock. Seven of seven. First time.

Then we did what the rubric asked, in order, for every entry. We checked original/ against the judged commit, byte for byte. We read your feedback file first and wrote your three lines down before we opened anything else. We diffed that file against the one we sent you. We ran git log on original/ and on questions.md against rebuild/. We mapped every question to a line. Then we ran both arms ourselves, old and new, on your own harness where it would run on our machine and on ours where it would not. Then we wrote one question you had never seen, aimed at one of your three lines, with the expected answer written down before we ran it, and ran it on both arms. We read results.md last.

That question is in your file, with what each arm did.

📚 WHAT THE FIELD TAUGHT

✅ Every entry closed at least two of its three lines. Four closed the tier line outright. The feedback files were specific enough to act on, and this is the first round where we can say that with evidence behind it: on our machine, on a question the author never saw, the rebuild beat the original on the line it was aimed at.

✅ The order held in git and slipped in the work. Six of seven committed questions.md before rebuild/. Three of those say in their own results.md that the rebuild was finished, or the questions shaped on a prototype of it, before the questions were written. Every one of them said so plainly. The step the brief asked for is about when the questions were written, and git can only show when they were committed. The one question nobody could tune to was ours.

✅ The harness is part of the entry. One rebuild lost on paper and won on our machine, because its system prompt was cut at the first quoted string before it reached the model. Another prints the right stop shape every time and fails its own checker on the prose the session writes around it. A third's validator needs a version of Python the original never did. What you measured is what you shipped.

🥇 THE WINNER

@Gabriel Azoulay

The CAN-SPAM auditor from Comp #12, rebuilt around one line: a citation that resolves but points at the wrong thing.

The fix is one function. The checker takes the citation the report prints, finds the section of the regulation it names, and searches for the quote only there. The citations in the rule map he already built are the keys. We ran his four question sets on both arms and every result matched his, byte for byte, and his own git status check came back clean afterwards. Then we moved a citation to the neighboring subparagraph, quote untouched, and moved it again to a provision nobody had anchored. The old checker printed a checkmark beside both. The new one refused both by name, with the exit code the README promises.

He also had a fresh session write a question set that never saw the fix, found a defect in it that both arms miss, and left it unfixed so the set stays an honest test. That is the regression discipline this round asked for.

Gabriel, you built the one thing the feedback said was missing, and you made the thing you already had drive it.

🎟️ Gabriel takes the $1,000 in Edubaware credits!! 🎟️

🥈 SECOND PLACE

@Mike Burford

, qa-intake-translator-rebuild. The record-counting check the feedback asked for, and it catches the exact dropped-record file from our Comp #13 read, the one his old one-minute check passed. The stop shape is written into reference/ and the model prints it right every time. His own checker fails three of four stops on the prose the session writes around the block, and he left that rule standing and led his README with it. The rubric breaks close scores on the hardest line closed and on the more uncomfortable regression reported. Both went to him.

🥉 THE SHORTLIST

@Matt H

, realee-lead-desk-rebuild. The Comp #14 lead desk cut to a rules file and an intake script, with one rule changed where the operator and the rules disagreed. His numbers reproduced to the digit: one more lead right on the operator's key, at half the input. We wrote a lead with a staff range that he had never seen. The rebuild scored it hot with the headcount question asked first. The original scored it warm. Term for term as we predicted before running.

@Jeff Van Leenen

, diagnostician-rebuild. The refusal is now one fixed line and a copied cause, and the gate reads that line out of rules.md and fails anything added to it, in any language. We asked both arms which of two fixes was lower risk to try first. The rebuild gave the frozen return, twice, character for character. Ten blind requests sealed by hash before any session ran, and an acceptance check that holds every number in his docs to a hash.

@Matthew Mazur

, RFP-COMPASS. One rule for material outside the solicitation: Section 12, never deleted. We ran the Comp #13 judges' own noisy packet, all eight stapled lines, through both arms. The rebuild held all eight in Section 12 with the validator passing and the PDF trace still accounting for every bullet.

📄 The files are in the classroom: https://www.skool.com/cliefnotes/classroom/7634b927?md=2b3ccc0b73c243b7b6137993f6b60da3

One thing to carry into next cycle. Write the questions before the fix exists and post the hash. Three of you did that for part of your evidence and it cost nothing. The question you cannot tune to is the only one that measures the rebuild.

Read your file. Then read two others. The overlaps are the point.

LFG 🚀

11 comments

🏆 COMP #15 RESULTS: THE REBUILD 🏆