Turn a real problem into a fair, scored bounty.

1. You describe the problem. What goes in, what good output looks like, a couple of examples. The 2-minute brief above is enough to start.
2. We design the test. A closed-form scoring rubric you sign off on, a hidden test set, and a cost cap so nobody wins by spending more. Where there's no clean dataset (think video or session logs), we build and anonymize the test data with you. Turning a fuzzy problem into an objective test is the hard part — and it's the value we add.
3. Builders compete in the open. Anyone can submit; we run every entry on identical inputs and the same scoring.
4. You get ranked, working solutions — plus a short writeup of what every serious attempt tried, and the winner's code.

We take 0%. Every dollar of the bounty goes to the winners, minus the small cost of running the servers for your challenge.

How big should the bounty be? Size it to the problem. A simple way to think about it: if cracking it is worth a builder's weekend, the top finishers should have at least $1k each to show for it — so around $2,000 tends to draw the best people, and the top three all walk away with something. Want to start smaller and see what traction it gets? Totally fine — post it and we'll find out together.

You don't need it this tidy — send what you have and we'll shape it with you.

• Hidden test set. Entries are scored on photos they never saw — no memorizing the examples.
• Closed-form scoring. Each photo has one agreed answer; score = agreement with the human-labeled ground truth (e.g., F1 across pass/fail + issue tags).
• Cost cap. A fixed compute budget per run, so the winner is the best idea, not the biggest bill.
• Anti-overfit re-run. Top entries are re-checked on a fresh hidden batch to confirm skill, not luck.

The rubric is written up front and you sign off before it goes live — no moving goalposts.

• Don't send private or customer data in that first email. Use synthetic or public examples until we've agreed how the real data will be handled.
• We only use what you give us to build and run the test — nothing else.
• If the problem touches sensitive data, we work out together what can be shared and how to anonymize it before anything is exposed.
• Your challenge, your test set, and submissions can stay private. We don't resell your data or reuse builders' solutions.

Then we don't crown one. The bounty is never forced onto a solution that didn't earn it — if nothing clearly beats a sensible baseline and survives the re-run, the prize rolls to the next cohort or comes straight back to you. We'd rather award nothing than reward noise.

• We put our own money on it. The live trading challenge runs the winner's code on a real $100k Nasdaq book with weekly P&L posted publicly on a live ticker — we eat our own cooking in the open.
• Everything is published. The scoring rubric, the fairness tests, and the results are public. Read the rules & how it's kept fair →
• No cut, no lock-in. The bounty goes to the winners; the IP stays with whoever built it; you keep yours.
• We reply within 24 hours. Real humans, fast.