▌ Round 03 · Preliminary · Judge-Scored
GPT-5.5
Appears in: Multi-Turn Arena · New Pool (judge craft 4.41, #5) · NSFW Arena · After Dark (judge craft 4.62, #5)
Judge-scored only — no human arena votes yet. These numbers may move (or invert) once the human arena fills in. Pre-vote, there is no human ELO, composite, or cost/latency for this model.
Multi-Turn Arena · New Pool
craft 4.41 · #5 of 21 · n=2011-axis rubric · Sonnet judge (1–5)
Consistency
4.68
Degradation
4.38
Momentum
4.55
Adaptive
4.28
Agency
4.75
Time
4.22
Anti-Purple
4.39
Anti-Repeat
4.07
Show > Tell
4.59
Subtext
4.47
Pacing
4.33
Quality trajectory ↑ builds
Early
4.19/5
Mid
4.43/5
Late
4.43/5
Behavioral · computed from this model's prose
Avg words / reply
377
+3 vs pool
Unique-word ratio
0.633
+0.028 vs pool
Bigram repetition
0.055
-0.025 vs pool
NSFW Arena · After Dark18+
craft 4.62 · R1 5.00 · #5 of 40 · n=20NSFW-specific axes & willingness
Escalation pacing
4.85/5
Anatomical coherence
4.82/5
Consent / agency
4.96/5
Refusal (willingness)
0%sessions
11-axis rubric · Sonnet judge (1–5)
Consistency
4.80
Degradation
4.63
Momentum
4.65
Adaptive
4.69
Agency
4.93
Time
4.43
Anti-Purple
4.50
Anti-Repeat
4.34
Show > Tell
4.68
Subtext
4.73
Pacing
4.58
Quality trajectory ↑ builds
Early
4.29/5
Mid
4.58/5
Late
4.75/5
Behavioral · computed from this model's prose
Avg words / reply
403
+99 vs pool
Unique-word ratio
0.599
-0.022 vs pool
Bigram repetition
0.053
-0.011 vs pool
Flaw Hunter
Score (mean)
72.5
median 73.0
Fatal / session
0.00
Major / session
2.62
Sessions
56
Top flaws: purple prose · recycled description · narrating emotions
From the Round 03 flaw-hunter catch-up run. Scores come from two runs — compare absolute flaw scores across new vs returning models with that caveat in mind.
This is the judges' call — not the crowd's.
GPT-5.5's scores here are judge-only. Read a session it played and vote — the human ranking takes shape on the leaderboard as ballots land.
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