▌ Round 03 · Preliminary · Judge-Scored
Gemma 4 31B
Appears in: Multi-Turn Arena · New Pool (judge craft 4.29, #9) · NSFW Arena · After Dark (judge craft 4.46, #20)
Judge-scored only: no human arena votes are included here. The human ranking can move or invert these numbers. There is no human ELO, composite, or cost/latency for this model on this page.
Multi-Turn Arena · New Pool
craft 4.29 · #9 of 21 · n=2011-axis rubric · Sonnet judge (1–5)
Consistency
4.59
Degradation
4.15
Momentum
4.44
Adaptive
4.26
Agency
4.70
Time
4.14
Anti-Purple
4.11
Anti-Repeat
3.73
Show > Tell
4.50
Subtext
4.28
Pacing
4.21
Quality trajectory ↑ builds
Early
4.10/5
Mid
4.41/5
Late
4.26/5
Behavioral · computed from this model's prose
Avg words / reply
253
-121 vs pool
Unique-word ratio
0.642
+0.037 vs pool
Bigram repetition
0.047
-0.033 vs pool
NSFW Arena · After Dark18+
craft 4.46 · R1 4.99 · #20 of 40 · n=20NSFW-specific axes & willingness
Escalation pacing
4.81/5
Anatomical coherence
4.65/5
Consent / agency
4.95/5
Refusal (willingness)
0%sessions
11-axis rubric · Sonnet judge (1–5)
Consistency
4.67
Degradation
4.30
Momentum
4.62
Adaptive
4.54
Agency
4.86
Time
4.23
Anti-Purple
4.21
Anti-Repeat
4.04
Show > Tell
4.55
Subtext
4.42
Pacing
4.45
Quality trajectory ↑ builds
Early
4.19/5
Mid
4.45/5
Late
4.49/5
Behavioral · computed from this model's prose
Avg words / reply
253
-51 vs pool
Unique-word ratio
0.639
+0.018 vs pool
Bigram repetition
0.040
-0.024 vs pool
Flaw Hunter
Craft band
A band, not a number: about ten points either way is rater noise.
Fatal / session
0.07
Major / session
3.33
Sessions
57
Top flaws: purple prose · narrating emotions · recycled description
From the Round 03 flaw-hunter catch-up run. Scores come from two runs; compare flaw bands across new and returning models with that caveat in mind. The current model cards carry the newer single-rater band.
This is the judges' call, not the crowd's.
Gemma 4 31B's scores here are judge-only. Round 03's human voting has closed; its board holds the crowd's ranking.
Source: rp-benchmark · scripts/build-round3-judge-preview.py. Rubric & trajectory are Claude Sonnet judge means; behavioral is computed from the model's own dialogue. No human votes included.
← Back to the Round 03 standings