How the arena works

Every ranking traces back to a simple, auditable loop: a real specimen, identical inputs, anonymous pairwise votes, and a statistical model that turns those votes into a score.

Reference

A real organism is captured or sourced — CT scan, photogrammetry, or a licensed reference image.

Generate

Every model reconstructs a 3D asset from the same single input, under identical conditions.

Pair

Two outputs for one task are sampled anonymously and shown side by side.

Vote

You inspect both in the live viewer and pick the better reconstruction — or call a tie.

Bradley–Terry

Pairwise votes fit a Bradley–Terry model, turning wins into a latent skill score for each model.

Rank

Ratings and 95% intervals produce the leaderboard, refreshed as votes accumulate.

The Bradley–Terry model

P(i beats j) = σ(β_i − β_j)

σ is the logistic sigmoid; β_i and β_j are the fitted latent skill of models i and j. The model that wins more often than its rating predicts pulls its own β up; the leaderboard is these β values rescaled to an Elo-like range.

Ranking

Online Elo (K=32, ties = 0.5) updates live on every decisive vote — the fast, approximate signal shown next to each generator.
Bradley–Terry MLE over the full pairwise win record is the authoritative leaderboard, fit by minorization–maximization and rescaled to an Elo-like range. Bootstrap resampling gives 95% confidence intervals.
Significance: a paired bootstrap reports P(A ranks above B) for every pair, so we can say whether A is meaningfully ahead — not just nominally.
Rank (UB): models whose 95% bootstrap confidence intervals overlap share a rank; a model outranks another only when its lower bound strictly exceeds the other's upper bound. Equal-CI models are co-ranked rather than arbitrarily ordered.
Ties are credited as a split — one win in each direction — so they inform Bradley-Terry symmetrically without a separate tie parameter. Comparisons marked "both bad" are recorded for audit but excluded from the fit entirely. Comparisons are drawn within a task, so the two outputs share a prompt and format.

Vote integrity

Anonymity: generator identities are never sent to the browser during voting; sides (A/B) are randomized to neutralize position bias.
Gold attention checks: ~10% of comparisons are calibration pairs with a known-good vs. an obviously-bad asset. A voter's trust is the Laplace-smoothed pass rate (gold_passed + 1)/(gold_seen + 1). Votes from sessions below trust 0.5 are excluded from the Bradley–Terry leaderboard (still recorded for audit).
Rate limiting: at most 60 votes per 60s per session.
Dedup: a session cannot re-vote the same pairing for the same criterion.
Human verification: captcha is enabled (Turnstile/hCaptcha integration point).

Bias audit

Gold/decoy assets are excluded from matchmaking and from all rankings — they exist only to measure voter trust.