The Unofficial Gemini 3 Showcase (2): Can We Trust AI Judges? I Checked My Ranking and I’m Not Top 50…
I Checked My Ranking and I’m Not Top 50…
While waiting for the final results, maybe you checked my previous post about The Unofficial Gemini 3 Showcase and noticed your ranking wasn’t as high as you’d hoped.
Original source: https://www.kaggle.com/competitions/gemini-3/discussion/668733
Now you’re wondering: Is this AI scoring even legit?
I ran some experiments to find out.
The Key Finding

Ah, you might be curious about how it works? Here is the diagram
Previously, each app got scored alone → rankings.
Now I put 4,000 app pairs head-to-head: if the “winner” disagrees with the original ranking, that’s a flip.
(Yep, I also checked if showing A before B matters. Next episode!)
When the score gap is small, AI judges are unreliable.
| Score Gap | Flip Rate | Verdict |
|---|---|---|
| < 1 pt | 20-35% | 🔴 Basically a coin flip |
| 1-2 pts | 15-20% | 🟡 Shaky |
| > 2 pts | < 8% | 🟢 Reliable |
So if you’re not in the top tier: it might just be noise, not skill.
Do the 8 AI Judges Agree?

Yes — correlations hit 0.70-0.88. Not perfect twins, but definitely siblings.
-
👀 Eagle-eyed readers might notice: Gemini 3 Pro’s data looks a bit… lonely? My API budget said “no more” before I could collect enough samples (scoring takes a lot of resources since it has to “read” all the source code).
-
But here’s where YOU come in! I built a quick 1-min-eval app on aistudio — just drag in your project (or paste your AI Studio URL), and let Gemini 3 judge it.
-
If you want, you can download json and create a PR here, we can make build it together (thinking of a better way) or just post it in the comments
How to Use This Leaderboard
| ✅ Get a rough sense of where you stand | ❌ Judge your worth as a vibe coder |
|---|---|
| ✅ Find your weak dimensions | ❌ Predict official results |
| ✅ Discover great projects to learn from | ❌ Lose sleep over it like I did |
| ✅ Try the “Shuffle” button when selecting a tier |

I’m also working on personalized takeaways for each app — what you did well, what you could improve, and ideas to inspire your next project.
If you want something specific? Drop a comment below and let me know what would be most helpful!
More data (updating soon): topvibecoder.com/kaggle-gemini-3

Comments
Select any text to comment on a specific part. Existing inline comments appear as small numbered bubbles. Powered by GitHub Discussions.