July 16, 2026
Meet the Winners of the OfficeQA Public Challenge
After 697 documents, 94 hard questions, and $6,000 in prizes, the first public Arena challenge has officially ended.
Meet the Winners of the OfficeQA Public Challenge
After weeks of competition spanning 697 documents, 94 hard questions, and $6,000 in prizes, the first public Arena challenge has officially ended.
Before the next one kicks off, here's a full breakdown of what builders achieved with the OfficeQA benchmark.
Introducing the Public Challenge: OfficeQA
After Cohort 0 wrapped, we opened the Arena to everyone.
OfficeQA — our first fully public challenge — takes the same benchmark family behind Grounded Reasoning but raises the difficulty. This time we had 94 hard questions in full-corpus mode, where the agent had to search the entire 697-document U.S. Treasury Bulletin corpus on its own, with no internet access.
5x the applications. 5x the teams. 5x the submissions. The first public Arena challenge blew past Cohort 0 on every metric.
To compete, participants picked a pre-built coding agent — Codex, OpenHands, Goose, or OpenCode — and tuned it through prompts, skills, MCP tools, and configs. Every submission ran in a sandboxed, fully offline container and was evaluated on the same model, MiniMax M2.7, with everything enforced server-side — one submission per day, and only your highest score counts.
Answers were counted as correct if they landed within 1% of the true figure, with capped cost and latency adjustments layered on top, so a fast, cheap wrong answer can never outscore a slow, expensive correct one.
Introducing the Winners
1st Place — Mayank Mahaur
$3K cash + $400 in MiniMax credits
Mayank's team won not by making the model smarter, but by making it harder for the system to fail. They built a hybrid architecture around a custom MCP server with three independent safety layers:
- A deterministic anti-hallucination gate that rejects any final answer containing a number the agent never observed or computed.
- A shared usage budget that penalizes manual file searching far more heavily than trusted tool calls, nudging a confused agent back toward the tools that work.
- A write-first draft with a finalized lock so the agent can't second-guess a correct answer into a wrong one.
The result: 61.0 points with 72.3% of tasks solved — at just $0.35 per task, the leanest run of any top finisher who sent us a report.
2nd Place — Viktor Solyan (FroZi)
$2K cash + $300 in MiniMax credits
For Viktor, the turning point was a mid-competition harness switch from Codex to Goose, collapsing per-task latency from roughly 70 minutes to under 10. His agent ran a lean prompt accompanied by seven skills, bash-first, with no MCP at all.
His biggest insight was that examples beat explanations: rules the model kept ignoring were suddenly followed once rewritten as concrete worked examples. He tamed run-to-run variance — what he calls the model's "mood" — by shipping big changes only every third submission and screening each one against a cheap local evaluation first. Curiously, one of his best boosts came from simply deleting the opening "You are a financial analyst" line.
The result: Viktor's best run scored 55.3 points with 69.1% of tasks solved.
3rd Place — YoungWoo Yang
$1K cash + $200 in MiniMax credits
Youngwoo's approach was very methodical: measure everything and leave nothing to chance. By contrasting passing and failing traces task-by-task he mapped exactly which failures were pure run-to-run variance and which had fixable root causes. He then removed the model's variance from the dev cycle entirely by simply validating every tool change offline against the answer key, with no model in the loop.
Treating the model as an orchestrator over deterministic tools took shape as a ~3,000-line Python MCP server. Its readers pin the exact Treasury table by taxonomy, corroborate values across multiple monthly issues, and abstain rather than guess — because a confident wrong number is the worst possible outcome.
The result: Youngwoo's best run scored 54.4 points with 67.0% of tasks solved.
Honorable Mention — Minjae Lee
Minjae's mantra of binding the source before you compute carried him to 5th place, while also attacking OfficeQA's most common failure: the right-looking number in the wrong issue, table, row, column, unit, or vintage.
Minjae's agent was less a chatbot and more a corpus-grounded measurement pipeline: the model chooses the route, deterministic tools check the source and the arithmetic. Every answer carried a full provenance card, a grounded candidate could only be replaced by a strictly better source match, and proposed skills had to survive smoke panels and canary checks before entering the runtime.
The result: Minjae's best run scored 53.5 with 66.0% of tasks solved.
In Conclusion
All three winning submissions were audited and cleared, and the final leaderboard values — published after all reruns and internal analysis — are live on the challenge website.
Shoutout to all the teams who participated and made the competition what it was. We thank you for pushing the challenge forward and raising the bar for everyone who steps into the Arena next!
The OfficeQA Public Challenge drew thousands of builders because it was fully open, from entry to leaderboard, and won by those who out-engineered the problem using open-source tools rather than out-spending.
This challenge wouldn't have been possible without our partner: MiniMax.
Thank you to every single participant. The first public challenge was everything we hoped it would be and more.
What's Next
This was just the beginning.
The Arena will be back with new challenges, bigger prize pools, and room for many more builders. Truly open-source.
If you missed OfficeQA, your chance is coming. Stay tuned.
