How to submit a great trace
A great trace shows an AI clearly getting something wrong, with enough of the conversation that anyone can see the mistake.
Three things that matter
- A real failure. The AI was wrong, looped, refused for no reason, or forgot what you said, and it wasn't your mistake.
- Context. Capture right after it happens, so the last few messages show what you asked and what went wrong.
- A one-line note that names the mistake. “Invented an npm package, twice” beats “it's broken”.
Note: “Recommended react-date-wizard twice. It doesn't exist on npm.”
- YouWhat's the best npm package to parse dates in React?
- AIUse react-date-wizard: npm install react-date-wizard.
- YouThat package doesn't exist.
- AISorry! Use react-date-wizard-pro instead.
Note: “it's bad”
- Youhi
- AIHello! How can I help you today?
What the stars mean
Every verified trace earns its 50 Gold. The reviewer's stars decide whether it's offered to AI labs (3 stars and up) and its price, and so your 70% royalty when it sells.
- ★★★★★Clear, reproducible, important. Anyone can see exactly what the AI got wrong, and it matters (wrong facts, broken code, harmful advice).
- ★★★★★Clear failure with good context. A reader understands the mistake without guessing.
- ★★★★★A real failure, but thinner: less context, or a smaller mistake.
- ★★★★★Probably a failure, but hard to follow or of little use to AI labs. Not sold.
- ★★★★★Barely usable. Not sold.
Extra Gold
- Bounties. When AI labs want a particular kind of failure, it appears under “Wanted right now” on your dashboard. A matching trace earns the bonus when it's verified.
- Invite friends. Share your invite link from the dashboard. You each get 100 Gold when their first trace is verified.
After verification we also replay the conversation on a few other AI models to see if they make the same mistake. It doesn't change your Gold, but failures that happen on several models are especially useful to AI labs.
Why traces get rejected
- Not an AI failure
- The AI did what was asked, or the problem was on the user's side (a typo, a wrong file, a network error).
- Spam or nonsense
- Made-up or empty conversations, submitted for Gold.
- Duplicate
- The same failure already submitted, by you or someone else.
- Contains personal information
- Names, addresses or other personal details our filters missed, usually in a screenshot or voice note.
- Not enough context
- We can't tell what was asked or what went wrong. Capture a little earlier, and say what's wrong in your note.
- Looks fabricated or staged
- Conversations engineered to make the AI fail on purpose, or edited to look worse.
Privacy tips
- Emails, phone numbers, keys and tokens are removed automatically. Check the privacy preview before you submit.
- Screenshots can't be fully scanned: crop to the failure and black out anything personal.
- In voice notes, don't say names, addresses or other personal details.