Published · findings · Code-CAD Arena. Across six rounds of head-to-head instrument builds, blind human preference votes and the deterministic mesh gate ranked the same entrants almost independently — the per-round Spearman rank correlation swings from −1.0 to +0.87. Across the five directly comparable three-entrant rounds, it averages roughly +0.07. Looking right and being buildable turn out to be different capabilities.
The Code-CAD Arena runs a small pool of code-CAD entrants through the same brief — model a named stringed instrument as OpenSCAD source — then scores each round twice, on purpose:
Subjective scoreline. Rendered candidates go into a blind A/B vote queue; preference votes update an Elo rating. This is "which build does a human prefer to look at," with no knowledge of which entrant produced it.
Objective scoreline. The same candidates pass through MakerBench's deterministic mesh gate — does the source compile, is the exported geometry manifold, watertight, sanely proportioned. No judge, human or model, is in this loop.
Arena policy has always been to report the two side by side and publish their rank agreement as a Spearman rho — evidence, never a blended composite. Rounds 5–10 put that policy to a family-scale test: six rounds spanning the whole stringed-instrument family, from bowed instruments through harps, keyed instruments, plucked zithers, guitars and lutes, to the largest basses.
The per-round Spearman rank correlation between the Elo ordering and the objective mesh-gate ordering traced this arc across the family:
R5 bowed ρ = −1.00 R6 harps/zithers ρ = −0.50 R7 keyed ρ = −0.50 R8 plucked-zither ρ = +0.87 R9 guitars/lutes ρ = +0.50 R10 basses/large ρ = 0.00
Across the five directly comparable three-entrant rounds the mean is ρ ≈ +0.07 — close to zero in this small sample. Not a stable positive relationship: decorrelated in these rounds. Knowing which entrant the eye prefers tells you essentially nothing about which entrant produced the cleanest geometry, and vice versa.
The arc's shape is itself informative. The complex-assembly rounds — bowed instruments, harps and zithers, keyed instruments — are where the two scorelines diverge, with correlations at or below −0.5. The geometrically simpler plucked and guitar rounds are where they agree. Where assembly complexity is high, a candidate can win the vote on silhouette and proportion while hiding non-manifold joints, interpenetrating bodies, or degenerate geometry that the mesh gate catches immediately — and a geometrically immaculate build can look awkward enough to lose every pairwise vote.
A qualitative through-line reinforced the statistics: the entrant with the most consistent mesh-gate record across the family rarely won the preference vote, while the entrant that most often won the eye was not the one with the cleanest geometry. Neither scoreline is "wrong" — they are measuring different things, and over this family they measured almost orthogonal things.
This is the empirical case for two standing MakerBench positions. First, scorelines stay separate: folding a preference signal and a geometry signal into one composite number would have manufactured a ranking that neither measurement supports. Second, deterministic grading is not optional — if human visual preference tracked geometric validity, an eyeball pass would be a cheap substitute for the mesh gate. With the five comparable rounds averaging ρ ≈ +0.07, it is not a substitute here. It also motivates the arena's third scoreline (a VLM-judge lane): with the eye and the gate this decorrelated, a vision-model judge is worth measuring as its own, separately reported signal rather than assumed to approximate either one.
Caveats. These are small samples — a handful of entrants per round, six instruments per round — so any single round's rho is noisy, which is exactly why the cross-round mean and the complexity pattern carry the weight here. The finding covers one instrument family in one arena configuration; whether the divergence-under-complexity pattern generalizes to other object families is an open question the arena is built to answer. Full per-round scorelines live in the arena section of the leaderboard when published runs clear the publication bar.