Zelda evaluations

21 games · six Score dimensions · two pairwise Choice models

Leaderboard and rubric

Value: Weighted Score
1.0At α=1, use returned Score / 4; otherwise transform the normalized five-level vector.
Normalized weights: . Slider values are renormalized to sum to 1.

Choice Δ is pairwise rank minus the current five-axis rubric rank; positive values place a title lower in Choice. Pareto * marks a game not dominated on all five rubric axes.

Pairwise tie and link geometry

Pairwise probability geometry

0246-4-201absolute win/loss ILR coordinate |z1|tie ILR coordinate z2

Jev interior slope +0.022; Gemini -0.366. Zero-containing Jev vectors are omitted; the dashed ε ray is a clipping artifact, not observed geometry.

○ Jev interior · ● Gemini · solid/dashed fit lines · dotted ε-clipped [1,0,0] illustration

Rubric dimension correlations

GeminiMean absolute correlation: 0.400ExplDungMechAtmoPaceExpl·+0.4+0.8+0.7+0.1Dung+0.4·+0.3+0.7+0.2Mech+0.8+0.3·+0.5+0.1Atmo+0.7+0.7+0.5·-0.1Pace+0.1+0.2+0.1-0.1·JevMean absolute correlation: 0.725ExplDungMechAtmoPaceExpl·+0.8+0.9+0.8+0.7Dung+0.8·+0.7+0.9+0.6Mech+0.9+0.7·+0.7+0.7Atmo+0.8+0.9+0.7·+0.5Pace+0.7+0.6+0.7+0.5·

Off-diagonal Pearson correlations across 21 games. Cell tone represents |r|; each cell prints signed r.

Instrument comparison

MeasurementGeminiJev
Reported cost / calls / answers$0.0262553 / 5 calls / 1,218not reported / 842 calls / 1,092
Exact repeat, Score0.000 MAE; 0/126 flips0.034 MAE; 7/126 flips
Catalog-order Score MAE (0–4)0.0750.059
Choice option-order shift+4.63 pp in p(A) when A is opt_1+4.37 pp in p(A) when A's criterion is listed second
Probability outputfloat64 JSON values; 0 exact Score/Choice zeros0.01 Hamilton grid; 597 exact Score/Choice zeros
Pairwise target entropy H(q)0.85810.4042
Davidson OOF KL0.014450.03112
Davidson OOF R²_KL0.93750.9464
Logistic-threshold OOF R²_KL0.9539not fitted in this comparison
Hodge gradient / cycle energy97.4% / 2.6%not computed: exact-zero probabilities
Nested cycle CV Δ loss (cycle − Davidson)+0.0000196 nats+0.0000013 nats
Rubric mean absolute correlation0.4000.725
Rubric standardized condition4.878.72

Pairwise R²_KL is normalized against each model's own training-fold mean vector. Cycle CV values are cycle-minus-Davidson soft loss.