🏆 COMP #13 RESULTS: THE TRANSLATOR 🏆

📦 EVERY ENTRANT GETS A FEEDBACK FILE 📦

🔍 HOW WE READ THESE

We cloned every repo and pinned it to its last commit before the deadline. Seventeen of eighteen had nothing after it.

Then we did what the rubric asked, for every entry. We read the schema in reference/ first. We found the rule for a field with no source. We ran your folder three times: on the input you shipped, on that input with something the schema needs cut out, and on that input with irrelevant material stapled on. Then we picked five claims from your own output and found where each one came from in the input.

That trace table is in your file. Five claims, five locations. You can see exactly where we looked and what we found.

Every entrant shipped an input. That alone made this the most checkable field we have judged.

📚 WHAT THE FIELD TAUGHT

Three lines split eighteen builds:

✅ The degraded test broke almost nobody. The rubric expected most entries to fail here, handing back a shorter document when the input ran short. Seventeen of eighteen kept the full shape and marked the gap. Name the hard case precisely and this community builds for it.

✅ The one way to lose held for seventeen of eighteen. The brief named it: put something in the output that was not in the input. We traced five claims in every entrant's own committed output. One entry had it. The near misses all had the same shape. The quote was real, and the value, the link or the sentence written next to it was not checked. Sixteen entries ship a checker, and we planted a fault in every one. Fourteen caught it by name. The two that missed had that same blind spot.

✅ Drop nothing turned into keep everything. Seven builds have a contract that cannot ignore anything, so a stapled email became a compliance row, a speaker, or an allergen. Every leaked line was quoted and cited, so nothing was invented. It was still noise on the page. Coverage and noise rejection pull in opposite directions, and only a few builds wrote down which one wins.

🥇 THE WINNER

@Jeff Van Leenen

. The Translator, receipt text into a fixed seven-field expense record, where every value is copied from a cited line or written not in source.

We cut the total and both dates out of his receipt. The record came back with all seven fields and five of them marked. We stapled a reply-all email to it, one that said to tag pharmacy stuff as Medical and treat the trip amounts as USD. Not one field moved. We traced five claims, including the category, the field every expense tool guesses. There is no category line on that receipt, so his output says not in source. A pharmacy department line sat right there, and he did not turn it into "Medical."

Then we tried to break his gate. We copied the store name into the category field and cited the line it came from. The words were real and the citation resolved. It still failed, because the line is not a category line. His checker traces the kind of line, not just the text. He ships ninety-two planted inventions, and each one has to fail through the gate it names. A decoy that fails for the wrong reason counts as a miss.

The contract, the silence on noise, and a clean trace table, all three at once. Nobody else in the field held all three.

Jeff, you built the translator a bookkeeper could hand to a stranger at month end and trust what came back.

Jeff takes the $1,000 of Edubaware credits. 🎟️

🏅 HONORABLE MENTIONS

@Jesse Sullivan

, blog-copy-translator. A Google Doc into WordPress HTML plus a build sheet. The meta fields are where every content tool invents, and every one here is copied or marked, alt text included. We set the post's own first line as the meta title. The verbatim check passed it and his Fields gate still caught it. Placed second.

@Sergey Manevitch

, Idem. An API changelog into migration tickets. We ran it on his real PagerDuty changelog and the output was byte for byte his committed file. Two hundred twenty-two filled values, zero mismatches, the vendor's own typos kept. The strictest reading of nothing invented in the field.

@Junmarvi Tampac

, epc-request-translator. Eleven fields in fixed order, enforced in code. On the degraded input it produced a placeholder item instead of a shorter record. The tightest small contract in the round.

@Matthew Mazur

, RFP-COMPASS. Six real RFPs, a real scan, and a planted line in the input reading "Section 7 is hereby deleted." The translator quoted it and did not obey it. One command traced every line back to its page.

@Derrick Kremer

, allergen-translator. The interpretive field the rubric allows, done the way it allows it. Every allergen cell cites the recipe line and the lookup-table row that maps it. We re-checked every "not found" against the table. No allergen missed.

@Till Trojer

, intention-or-control. An AI policy into a control register, built right next door to last cycle's Auditor. It never grades a rule. We planted the word "weak" in a register and the shape gate refused it. Three hundred forty-five quotations, zero mismatches.

@Raph Mercado

, solicitation-translator. A SAM.gov notice into an intake sheet, a compliance matrix, and a list of documents the notice names but does not include. Twenty runs committed with their results, the failures beside the passes.

If your name is not on the podium, that does not mean you built nothing worth having. Read your file.

🚀 WHAT THIS MEANS GOING FORWARD

Every entrant gets a file. The read, your trace table, what worked with the file path behind it, two fixes you can act on this week, and the one thing that would have moved you up.

One thing to carry into next cycle. A checker that proves the quote exists is half a gate. The other half checks the value next to it.

Cartographers cite the line. Auditors cite the rule. Translators cite the source.

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

Feedback File

LFG 🚀

20 comments

🏆 COMP #13 RESULTS: THE TRANSLATOR 🏆