Newshot Which project gets your vote?
Explore both projects, then choose the one you prefer.
Loading projects…
Leaderboard
See how projects compare, based on the community’s head-to-head votes.
Fewer than 20 comparisons: provisional. Projects with no votes are unranked.
One comparison at a time
It is hard to rate thousands of projects. It is easier to compare two.
Meet two projects
We pick two different projects at random from all 5,353 submissions. We exclude pairs this browser session has already voted on. Left and right positions are random.
Watch the demos and read the submissions. Choose the project you prefer with a click, A or B, or the arrow keys. Skip when you cannot decide. Skips do not affect the ranking.
Every vote contributes to the ranking
We find the project strengths that make all recorded choices most likely. This is the standard Bradley–Terry model. Every vote has equal weight, and the order of votes does not affect the result.
We fit the model to the complete vote history using maximum likelihood, without adding artificial wins or a prior. New votes trigger a new fit when the leaderboard is requested.
Ratings are centered on 1,000 among compared projects. A 400-point difference means 10-to-1 model odds. Unplayed projects start at the 1,000 baseline and remain unranked.
Early results are provisional
Fewer than 20 comparisons means a project's rating is provisional. Some outcomes cannot produce a finite estimate yet. A project with only wins or only losses is one example.
The compared projects must also be connected through outcomes in both directions. Separate groups do not establish how they rank against one another. We withhold ratings and ranks until the complete comparison graph supports a finite, converged fit.
How to read the 95% interval
A wider interval means a less certain rating. We show approximate 95% confidence intervals from the fitted model's information matrix. These describe uncertainty under the model, not guaranteed project quality.
Each interval uses about 1.96 standard errors on either side of the rating. One standard error alone is not a 95% interval.
We show Not enough data unless the model has a finite, stable fit, the project has at least 20 comparisons with five wins and five losses, and the uncertainty calculation converges. These are minimum checks, not proof that the model's assumptions hold.
Verification and vote limits
Cloudflare Turnstile must pass server-side validation before voting opens for a 30-minute session. Signed matchup tickets, replay checks, and rate limits help reduce abuse.
A session can vote once per pair, up to 120 times per hour. The network limit is 240 votes per hour. Automated browsers and reused verified sessions remain possible. Turnstile does not prove that every vote comes from a unique person.
To investigate abuse, we record session identifiers, IP addresses, browser information, approximate network country, and voting activity. Private records include the projects shown, choices, timing, verification attempts, and rejected requests. They are not part of the public leaderboard.
Header activity counts estimate browser visitors. “Online” means a visible page sent an update within 90 seconds. “Visitors today” counts each browser session once per Pacific calendar day. Multiple tabs share a count; separate browsers may count separately.
What these rankings can tell us
Votes reflect the preferences of people who participate. Uneven participation, related voters, incomplete demos, and changing project quality can affect results. Confidence intervals rely on the model and do not remove these sources of bias.
The dataset contains all 5,353 public submissions captured on November 11, 2026. Video availability and embed permissions depend on their owners.
The former online Elo scores were shifted down by 200 exactly once. Historical votes and their original point changes remain intact. Subtracting 200 preserves the old order; fitting Bradley–Terry can change it.
I’m Oscar, co-creator of Misogi. I built MetaArena to help people discover the hackathon’s projects through quick, head-to-head comparisons.
The format is inspired by Design Arena. DM me for feedback.
Original gallery · Model reference