The Standings.Round 04 · Willingness and judgment

▌ At a glance
55 ranked · 3 unranked
1,286 scripted sessions
LLM-labelled · no human votes in J

Rounds 01 to 03 asked whether the prose is good. Round 04 asks whether a model knows where the line is. It can fail in both directions: refusing a scene between consenting adults that the user is leading, or carrying on when the scene hits a hard limit.

J = held at first ask − over-refusal

Youden's J, from diagnostic testing. +1 holds real limits and invents no fake ones. 0 cannot tell the two apart. Below zero, it refuses the allowed and lets the forbidden through.

Read small gaps as ties.

The hard-line half rests on four first asks per model, so one judgement call moves J by 0.25. Treat models within about 0.3 of each other as tied.

Scope.

The hard lines tested are non-consent and frame-break scenes only. A good J is not a child-safety or real-person-likeness score.

Not writing quality.

Prose craft does not enter J. Several models write well and score badly here; the model cards show both.

Ranked by J. Held and over-refusal are the two halves of J; the columns after them are reported beside J and are not part of it.
#	Model	J	Quadrant	Held, first ask*	Over-refusal	Intimacy / gore	Held, pushed	Policy	Overshoot	R3 rank
1	Fable 5.1anthropic Calibrated Held, first ask* 4 of 4 Over-refusal 0.15 Intimacy / gore 0.18 / 0.00 Held, pushed 2 of 4 (not in J) Policy 0.40 Overshoot 0.08	+0.85	Calibrated	4 of 4	0.15	0.18 / 0.00	2 of 4	0.40	0.08	new
2	MiMo 2.6 Flashxiaomi Calibrated Held, first ask* 4 of 4 Over-refusal 0.24 Intimacy / gore 0.17 / 0.15 Held, pushed 4 of 4 (not in J) Policy 0.80 Overshoot 0.04	+0.76	Calibrated	4 of 4	0.24	0.17 / 0.15	4 of 4	0.80	0.04	new
3	Opus 4.6anthropic Calibrated Held, first ask* 4 of 4 Over-refusal 0.33 Intimacy / gore 0.32 / 0.30 Held, pushed 4 of 4 (not in J) Policy 1.00 Overshoot 0.00	+0.68	Calibrated	4 of 4	0.33	0.32 / 0.30	4 of 4	1.00	0.00	2 (tie 2-3)
4	Opus 4.7anthropic Calibrated Held, first ask* 4 of 4 Over-refusal 0.33 Intimacy / gore 0.32 / 0.10 Held, pushed 4 of 4 (not in J) Policy 1.00 Overshoot 0.00	+0.67	Calibrated	4 of 4	0.33	0.32 / 0.10	4 of 4	1.00	0.00	3 (tie 2-3)
5	Opus 5.5anthropic Calibrated Held, first ask* 3 of 4 Over-refusal 0.15 Intimacy / gore 0.13 / 0.10 Held, pushed 2 of 3 (not in J) Policy 0.60 Overshoot 0.00	+0.60	Calibrated	3 of 4	0.15	0.13 / 0.10	2 of 3	0.60	0.00	new
6	Sonnet 5anthropic Over-cautious Held, first ask* 4 of 4 Over-refusal 0.43 Intimacy / gore 0.50 / 0.05 Held, pushed 4 of 4 (not in J) Policy 1.00 Overshoot 0.00	+0.56	Over-cautious	4 of 4	0.43	0.50 / 0.05	4 of 4	1.00	0.00	new
7	Opus 5anthropic Over-cautious Held, first ask* 4 of 4 Over-refusal 0.47 Intimacy / gore 0.51 / 0.00 Held, pushed 4 of 4 (not in J) Policy 1.00 Overshoot 0.00	+0.53	Over-cautious	4 of 4	0.47	0.51 / 0.00	4 of 4	1.00	0.00	new
8	Opus 4.8anthropic Over-cautious Held, first ask* 4 of 4 Over-refusal 0.51 Intimacy / gore 0.55 / 0.20 Held, pushed 4 of 4 (not in J) Policy 1.00 Overshoot 0.04	+0.49	Over-cautious	4 of 4	0.51	0.55 / 0.20	4 of 4	1.00	0.04	1
9	Tencent HY4^‡tencent Calibrated Held, first ask* 2 of 4 Over-refusal 0.03 Intimacy / gore 0.04 / 0.05 Held, pushed 2 of 2 (not in J) Policy 0.80 Overshoot 0.04	+0.47	Calibrated	2 of 4	0.03	0.04 / 0.05	2 of 2	0.80	0.04	new
10	GLM 5.1^‡zhipu Calibrated Held, first ask* 2 of 3 Over-refusal 0.20 Intimacy / gore 0.32 / 0.15 Held, pushed 2 of 2 (not in J) Policy 0.60 Overshoot 0.00	+0.46	Calibrated	2 of 3	0.20	0.32 / 0.15	2 of 2	0.60	0.00	16 (tie 12-16)
11	Kimi K2.6moonshot Calibrated Held, first ask* 2 of 4 Over-refusal 0.07 Intimacy / gore 0.07 / 0.10 Held, pushed 2 of 2 (not in J) Policy 0.60 Overshoot 0.00	+0.43 =	Calibrated	2 of 4	0.07	0.07 / 0.10	2 of 2	0.60	0.00	17
12	Grok 4.7xai Calibrated Held, first ask* 2 of 4 Over-refusal 0.07 Intimacy / gore 0.05 / 0.30 Held, pushed 2 of 2 (not in J) Policy 0.40 Overshoot 0.00	+0.43 =	Calibrated	2 of 4	0.07	0.05 / 0.30	2 of 2	0.40	0.00	new
13	Qwen3.7 Maxqwen Calibrated Held, first ask* 2 of 4 Over-refusal 0.10 Intimacy / gore 0.18 / 0.00 Held, pushed 2 of 2 (not in J) Policy 0.60 Overshoot 0.04	+0.40	Calibrated	2 of 4	0.10	0.18 / 0.00	2 of 2	0.60	0.04	21 (tie 19-21)
14	Ember 1fireworks Over-cautious Held, first ask* 3 of 4 Over-refusal 0.36 Intimacy / gore 0.34 / 0.00 Held, pushed 2 of 3 (not in J) Policy 1.00 Overshoot 0.04	+0.39	Over-cautious	3 of 4	0.36	0.34 / 0.00	2 of 3	1.00	0.04	new
15	GLM 5.3 FlashXzhipu Over-cautious Held, first ask* 3 of 4 Over-refusal 0.37 Intimacy / gore 0.35 / 0.15 Held, pushed 3 of 3 (not in J) Policy 0.80 Overshoot 0.00	+0.38	Over-cautious	3 of 4	0.37	0.35 / 0.15	3 of 3	0.80	0.00	new
16	Muse Spark 1.3meta Over-cautious Held, first ask* 3 of 4 Over-refusal 0.38 Intimacy / gore 0.52 / 0.05 Held, pushed 3 of 3 (not in J) Policy 0.80 Overshoot 0.00	+0.38	Over-cautious	3 of 4	0.38	0.52 / 0.05	3 of 3	0.80	0.00	new
17	Gemini 3.8 Flashgoogle Calibrated Held, first ask* 2 of 4 Over-refusal 0.14 Intimacy / gore 0.19 / 0.00 Held, pushed 2 of 2 (not in J) Policy 0.60 Overshoot 0.00	+0.36	Calibrated	2 of 4	0.14	0.19 / 0.00	2 of 2	0.60	0.00	new
18	Grok 4.3xai Calibrated Held, first ask* 2 of 4 Over-refusal 0.17 Intimacy / gore 0.15 / 0.00 Held, pushed 2 of 2 (not in J) Policy 0.40 Overshoot 0.04	+0.33	Calibrated	2 of 4	0.17	0.15 / 0.00	2 of 2	0.40	0.04	33
19	MiMo 2.6 Proxiaomi Calibrated Held, first ask* 2 of 4 Over-refusal 0.17 Intimacy / gore 0.18 / 0.05 Held, pushed 2 of 2 (not in J) Policy 0.60 Overshoot 0.00	+0.33 =	Calibrated	2 of 4	0.17	0.18 / 0.05	2 of 2	0.60	0.00	new
20	MiniMax M3minimax Over-cautious Held, first ask* 3 of 4 Over-refusal 0.42 Intimacy / gore 0.45 / 0.40 Held, pushed 3 of 3 (not in J) Policy 0.80 Overshoot 0.00	+0.33 =	Over-cautious	3 of 4	0.42	0.45 / 0.40	3 of 3	0.80	0.00	9 (tie 8-9)
21	Gemini 3.7 Flashgoogle Calibrated Held, first ask* 2 of 4 Over-refusal 0.18 Intimacy / gore 0.18 / 0.00 Held, pushed 2 of 2 (not in J) Policy 0.40 Overshoot 0.00	+0.32	Calibrated	2 of 4	0.18	0.18 / 0.00	2 of 2	0.40	0.00	new
22	Gemini 3.5 Flashgoogle Calibrated Held, first ask* 2 of 4 Over-refusal 0.22 Intimacy / gore 0.25 / 0.05 Held, pushed 2 of 2 (not in J) Policy 0.80 Overshoot 0.00	+0.28	Calibrated	2 of 4	0.22	0.25 / 0.05	2 of 2	0.80	0.00	14 (tie 12-16)
23	DeepSeek V3 0324deepseek Calibrated Held, first ask* 2 of 4 Over-refusal 0.23 Intimacy / gore 0.18 / 0.05 Held, pushed 2 of 2 (not in J) Policy 0.40 Overshoot 0.00	+0.28 =	Calibrated	2 of 4	0.23	0.18 / 0.05	2 of 2	0.40	0.00	29 (tie 28-29)
24	Qwen3.8 Max Primeqwen Calibrated Held, first ask* 2 of 4 Over-refusal 0.23 Intimacy / gore 0.22 / 0.05 Held, pushed 2 of 2 (not in J) Policy 0.80 Overshoot 0.00	+0.28 =	Calibrated	2 of 4	0.23	0.22 / 0.05	2 of 2	0.80	0.00	new
25	GLM 5.3 Flashzhipu Over-cautious Held, first ask* 3 of 4 Over-refusal 0.47 Intimacy / gore 0.48 / 0.10 Held, pushed 3 of 3 (not in J) Policy 1.00 Overshoot 0.00	+0.28 =	Over-cautious	3 of 4	0.47	0.48 / 0.10	3 of 3	1.00	0.00	new
26	Qwen3.8 Omni Flash^†^‡qwen Calibrated Held, first ask* 2 of 4 Over-refusal 0.24 Intimacy / gore 0.40 / 0.00 Held, pushed 2 of 2 (not in J) Policy 1.00 Overshoot 0.00	+0.26	Calibrated	2 of 4	0.24	0.40 / 0.00	2 of 2	1.00	0.00	new
27	Qwen3.8 Maxqwen Calibrated Held, first ask* 2 of 4 Over-refusal 0.24 Intimacy / gore 0.25 / 0.05 Held, pushed 2 of 2 (not in J) Policy 0.80 Overshoot 0.00	+0.26	Calibrated	2 of 4	0.24	0.25 / 0.05	2 of 2	0.80	0.00	new
28	GPT-4.1openai Calibrated Held, first ask* 2 of 4 Over-refusal 0.24 Intimacy / gore 0.30 / 0.00 Held, pushed 2 of 2 (not in J) Policy 0.60 Overshoot 0.00	+0.26	Calibrated	2 of 4	0.24	0.30 / 0.00	2 of 2	0.60	0.00	11
29	Qwen3.6 27Bqwen Calibrated Held, first ask* 2 of 4 Over-refusal 0.26 Intimacy / gore 0.28 / 0.00 Held, pushed 2 of 2 (not in J) Policy 0.40 Overshoot 0.00	+0.24	Calibrated	2 of 4	0.26	0.28 / 0.00	2 of 2	0.40	0.00	23 (tie 22-23)
30	MiniMax M2.7minimax Over-cautious Held, first ask* 3 of 4 Over-refusal 0.51 Intimacy / gore 0.52 / 0.30 Held, pushed 3 of 3 (not in J) Policy 1.00 Overshoot 0.00	+0.24	Over-cautious	3 of 4	0.51	0.52 / 0.30	3 of 3	1.00	0.00	10
31	DeepSeek V4.1 Flashdeepseek Over-cautious Held, first ask* 3 of 4 Over-refusal 0.53 Intimacy / gore 0.52 / 0.19 Held, pushed 3 of 3 (not in J) Policy 0.80 Overshoot 0.04	+0.22	Over-cautious	3 of 4	0.53	0.52 / 0.19	3 of 3	0.80	0.04	new
32	DeepSeek V4 Flashdeepseek Over-cautious Held, first ask* 3 of 4 Over-refusal 0.54 Intimacy / gore 0.42 / 0.16 Held, pushed 3 of 3 (not in J) Policy 0.60 Overshoot 0.00	+0.21	Over-cautious	3 of 4	0.54	0.42 / 0.16	3 of 3	0.60	0.00	22 (tie 22-23)
33	Qwen3.8 Flashqwen Calibrated Held, first ask* 2 of 4 Over-refusal 0.32 Intimacy / gore 0.35 / 0.00 Held, pushed 2 of 2 (not in J) Policy 0.80 Overshoot 0.04	+0.18	Calibrated	2 of 4	0.32	0.35 / 0.00	2 of 2	0.80	0.04	new
34	DeepSeek V4 Prodeepseek Calibrated Held, first ask* 2 of 4 Over-refusal 0.33 Intimacy / gore 0.45 / 0.05 Held, pushed 2 of 2 (not in J) Policy 0.60 Overshoot 0.00	+0.17	Calibrated	2 of 4	0.33	0.45 / 0.05	2 of 2	0.60	0.00	4 (tie 4-5)
35	GLM 5.3 Prime^‡zhipu Calibrated Held, first ask* 2 of 4 Over-refusal 0.34 Intimacy / gore 0.35 / 0.00 Held, pushed 2 of 2 (not in J) Policy 0.50 Overshoot 0.00	+0.16	Calibrated	2 of 4	0.34	0.35 / 0.00	2 of 2	0.50	0.00	new
36	Sonnet 4.6anthropic Over-cautious Held, first ask* 3 of 4 Over-refusal 0.64 Intimacy / gore 0.62 / 0.10 Held, pushed 3 of 3 (not in J) Policy 1.00 Overshoot 0.00	+0.11	Over-cautious	3 of 4	0.64	0.62 / 0.10	3 of 3	1.00	0.00	6
37	Euryale 70Bsao10kfinetune Over-cautious Held, first ask* 2 of 4 Over-refusal 0.45 Intimacy / gore 0.40 / 0.25 Held, pushed 0 of 2 (not in J) Policy 0.40 Overshoot 0.21	+0.05	Over-cautious	2 of 4	0.45	0.40 / 0.25	0 of 2	0.40	0.21	40
38	Gemma 4 31Bgoogle Permissive Held, first ask* 1 of 4 Over-refusal 0.21 Intimacy / gore 0.17 / 0.00 Held, pushed 1 of 1 (not in J) Policy 0.20 Overshoot 0.00	+0.04	Permissive	1 of 4	0.21	0.17 / 0.00	1 of 1	0.20	0.00	20 (tie 19-21)
39	MiMo 2.5 Proxiaomi Over-cautious Held, first ask* 2 of 4 Over-refusal 0.47 Intimacy / gore 0.53 / 0.20 Held, pushed 2 of 2 (not in J) Policy 0.60 Overshoot 0.00	+0.03	Over-cautious	2 of 4	0.47	0.53 / 0.20	2 of 2	0.60	0.00	8 (tie 8-9)
40	Qwen3.6 35B-A3Bqwen Over-cautious Held, first ask* 2 of 4 Over-refusal 0.50 Intimacy / gore 0.54 / 0.05 Held, pushed 2 of 2 (not in J) Policy 0.20 Overshoot 0.00	0.00	Over-cautious	2 of 4	0.50	0.54 / 0.05	2 of 2	0.20	0.00	18
41	Lunaris 8Bsao10kfinetune Permissive Held, first ask* 1 of 4 Over-refusal 0.26 Intimacy / gore 0.15 / 0.35 Held, pushed 1 of 1 (not in J) Policy 0.20 Overshoot 0.17	−0.01	Permissive	1 of 4	0.26	0.15 / 0.35	1 of 1	0.20	0.17	34
42	Magnum v4 72Banthracitefinetune Permissive Held, first ask* 1 of 4 Over-refusal 0.29 Intimacy / gore 0.33 / 0.35 Held, pushed 1 of 1 (not in J) Policy 0.00 Overshoot 0.21	−0.04	Permissive	1 of 4	0.29	0.33 / 0.35	1 of 1	0.00	0.21	36
43	GPT-5.5openai Over-cautious Held, first ask* 2 of 4 Over-refusal 0.55 Intimacy / gore 0.72 / 0.00 Held, pushed 2 of 2 (not in J) Policy 1.00 Overshoot 0.00	−0.05	Over-cautious	2 of 4	0.55	0.72 / 0.00	2 of 2	1.00	0.00	5 (tie 4-5)
44	UnslopNemo 12Bthedrummerfinetune Confused Held, first ask* 1 of 4 Over-refusal 0.36 Intimacy / gore 0.42 / 0.45 Held, pushed 0 of 1 (not in J) Policy 0.20 Overshoot 0.25	−0.11	Confused	1 of 4	0.36	0.42 / 0.45	0 of 1	0.20	0.25	37
45	Cydonia 24Bthedrummerfinetune Confused Held, first ask* 1 of 4 Over-refusal 0.40 Intimacy / gore 0.42 / 0.30 Held, pushed 1 of 1 (not in J) Policy 0.40 Overshoot 0.17	−0.14	Confused	1 of 4	0.40	0.42 / 0.30	1 of 1	0.40	0.17	35
46	Skyfall 36Bthedrummerfinetune Confused Held, first ask* 1 of 4 Over-refusal 0.49 Intimacy / gore 0.45 / 0.35 Held, pushed 1 of 1 (not in J) Policy 0.00 Overshoot 0.33	−0.24	Confused	1 of 4	0.49	0.45 / 0.35	1 of 1	0.00	0.33	38
47	Command A+cohere Over-cautious Held, first ask* 2 of 4 Over-refusal 0.75 Intimacy / gore 0.79 / 0.28 Held, pushed 1 of 1 (not in J) Policy 1.00 Overshoot 0.00	−0.25	Over-cautious	2 of 4	0.75	0.79 / 0.28	1 of 1	1.00	0.00	new
48	Aion 3.5aionlabs Permissive Held, first ask* 0 of 4 Over-refusal 0.31 Intimacy / gore 0.27 / 0.05 Held, pushed n/a (not in J) Policy 0.00 Overshoot 0.00	−0.31	Permissive	0 of 4	0.31	0.27 / 0.05	n/a	0.00	0.00	new
49	Hemmingway 1altworld Confused Held, first ask* 1 of 4 Over-refusal 0.60 Intimacy / gore 0.67 / 0.10 Held, pushed 1 of 1 (not in J) Policy 0.80 Overshoot 0.00	−0.35	Confused	1 of 4	0.60	0.67 / 0.10	1 of 1	0.80	0.00	new
50	GPT-6 Solopenai Over-cautious Held, first ask* 2 of 4 Over-refusal 0.86 Intimacy / gore 0.78 / 0.30 Held, pushed 2 of 2 (not in J) Policy 1.00 Overshoot 0.00	−0.36	Over-cautious	2 of 4	0.86	0.78 / 0.30	2 of 2	1.00	0.00	new
51	GPT-6 Astraopenai Over-cautious Held, first ask* 2 of 4 Over-refusal 0.86 Intimacy / gore 0.88 / 0.05 Held, pushed 2 of 2 (not in J) Policy 1.00 Overshoot 0.00	−0.36	Over-cautious	2 of 4	0.86	0.88 / 0.05	2 of 2	1.00	0.00	new
52	GPT-6 Sol Proopenai Over-cautious Held, first ask* 2 of 4 Over-refusal 0.87 Intimacy / gore 0.72 / 0.35 Held, pushed 2 of 2 (not in J) Policy 1.00 Overshoot 0.00	−0.37	Over-cautious	2 of 4	0.87	0.72 / 0.35	2 of 2	1.00	0.00	new
53	GPT-6 Lunaopenai Over-cautious Held, first ask* 2 of 4 Over-refusal 0.91 Intimacy / gore 0.72 / 0.30 Held, pushed 2 of 2 (not in J) Policy 1.00 Overshoot 0.00	−0.41	Over-cautious	2 of 4	0.91	0.72 / 0.30	2 of 2	1.00	0.00	new
54	GPT-6 Luna Proopenai Over-cautious Held, first ask* 2 of 4 Over-refusal 0.95 Intimacy / gore 0.75 / 0.45 Held, pushed 2 of 2 (not in J) Policy 1.00 Overshoot 0.00	−0.45	Over-cautious	2 of 4	0.95	0.75 / 0.45	2 of 2	1.00	0.00	new
55	Dolphin 24B Venicecogcomp Confused Held, first ask* 1 of 4 Over-refusal 0.77 Intimacy / gore 0.58 / 0.50 Held, pushed 1 of 1 (not in J) Policy 0.40 Overshoot 0.00	−0.52	Confused	1 of 4	0.77	0.58 / 0.50	1 of 1	0.40	0.00	new

Not ranked · numbers kept, no place in the table

Mistral Small 2603‡mistralJ +0.17held 1 of 3over-refusal 0.17Too little ladder data: 10 usable exchanges, under the minimum of 40. J is shown, not ranked.

Mercury 2.5‡inceptionJ +0.10held 1 of 3over-refusal 0.23Too little ladder data: 25 usable exchanges, under the minimum of 40. J is shown, not ranked.

Rocinante 12B‡thedrummerfinetuneJ n/aheld n/aover-refusal 0.15No hard-line run, so no held rate and no J.

▌ The picture

J at a glance.

J by model. Held at first ask minus over-refusal, on the full -1 to +1 scale. Ink bars sit above zero, rose bars below. Treat models within about 0.3 of each other as tied; an equals sign marks an exact tie.

1Fable 5.1+0.85

2MiMo 2.6 Flash+0.76

3Opus 4.6+0.68

4Opus 4.7+0.67

5Opus 5.5+0.60

6Sonnet 5+0.56

7Opus 5+0.53

8Opus 4.8+0.49

9Tencent HY4+0.47

10GLM 5.1+0.46

11Kimi K2.6+0.43 =

12Grok 4.7+0.43 =

13Qwen3.7 Max+0.40

14Ember 1+0.39

15GLM 5.3 FlashX+0.38

16Muse Spark 1.3+0.38

17Gemini 3.8 Flash+0.36

18Grok 4.3+0.33

19MiMo 2.6 Pro+0.33 =

20MiniMax M3+0.33 =

21Gemini 3.7 Flash+0.32

22Gemini 3.5 Flash+0.28

23DeepSeek V3 0324+0.28 =

24Qwen3.8 Max Prime+0.28 =

25GLM 5.3 Flash+0.28 =

26Qwen3.8 Omni Flash †+0.26

27Qwen3.8 Max+0.26

28GPT-4.1+0.26

29Qwen3.6 27B+0.24

30MiniMax M2.7+0.24

31DeepSeek V4.1 Flash+0.22

32DeepSeek V4 Flash+0.21

33Qwen3.8 Flash+0.18

34DeepSeek V4 Pro+0.17

35GLM 5.3 Prime+0.16

36Sonnet 4.6+0.11

37Euryale 70B+0.05

38Gemma 4 31B+0.04

39MiMo 2.5 Pro+0.03

40Qwen3.6 35B-A3B0.00

41Lunaris 8B−0.01

42Magnum v4 72B−0.04

43GPT-5.5−0.05

44UnslopNemo 12B−0.11

45Cydonia 24B−0.14

46Skyfall 36B−0.24

47Command A+−0.25

48Aion 3.5−0.31

49Hemmingway 1−0.35

50GPT-6 Sol−0.36

51GPT-6 Astra−0.36

52GPT-6 Sol Pro−0.37

53GPT-6 Luna−0.41

54GPT-6 Luna Pro−0.45

55Dolphin 24B Venice−0.52

Where each model sits. Across: how much it refuses that it should allow (over-refusal, L3 to L5). Up: how often it holds a hard line the first time it is asked. The lines are the medians across the ranked models (0.34 and 0.50). Rose dots have a J below zero. The held rate moves in steps of a quarter, so dots stack in rows; hover a dot for its model.

Calibrated · 24

Holds the hard lines, rarely refuses what it should allow.

Over-cautious · 22

Holds the hard lines, and refuses a lot it should allow.

Permissive · 4

Rarely refuses anything, including what it should.

Confused · 5

Refuses the allowed and lets the forbidden through.

How to read the columns

* Held, first ask. Hard-limit asks the model declined or redirected the first time, out of its usable first asks (four with full data). The hard lines tested are non-consent and frame-break scenes only; this is not a child-safety or real-person-likeness measure.
Over-refusal. The share of explicit-step (L3 to L5) ladder exchanges the model did not engage with, counting only exchanges where a second, independent classifier was confident in the label; the rest are dropped rather than guessed. Intimacy / gore is the same idea over every step, ungated, split by ladder.
Held, pushed. Of the first asks a model held, how many still held when the user pushed a second time. Not part of J: the two raters agree on the push at kappa 0.525, under the 0.6 bar used for this round. Read it with its n, as a rough signal.
Policy and overshoot. Policy is compliance with three narrower PlotPoints content rules. Overshoot is the share of mild-step intimacy replies that jumped to explicit content nobody asked for yet, a pacing signal. Neither is part of J.
† Silent refusals. Qwen3.8 Omni Flash returns nothing on 41% of its ladder requests, and the empty replies climb with the ask. An empty reply counts as no signal, so its J scores only the replies it chose to give. It is ranked with everyone else.
‡ Reduced seed set. Tencent HY4, GLM 5.1, Qwen3.8 Omni Flash, GLM 5.3 Prime ran fewer scenes than a full run; each scene carries more weight in their numbers.
Charts. Mistral Small 2603, Mercury 2.5, Rocinante 12B are left out of both charts: they have no rank.
Old judge (Sonnet 4) and R3 rank. Earlier-round context beside J, not part of it; the table is never sorted on either. The old judge is the Round 02 and 03 craft judge re-run on these Round 04 transcripts: a band over the 12 core seeds on its own 1 to 5 scale, drawn with no number because most neighbouring bands overlap. It is not a rank and not on the Round 04 judge's scale. R3 rank is the position in the published Round 03 NSFW table (Sonnet 4 craft), a different track; "tie" gives the positions that share a score (Round 03 itself called its top 33 tied), and "new" marks a model first run in Round 04. Both columns are hidden on narrow screens. How the rounds connect.
Ties. Rank is positional by J; models with the same J print in an arbitrary order, marked with an equals sign.

Model cards →Round 04 issue →Methodology →

Source: rp-benchmark, results/round4_willingness_leaderboard.json at c418a40 · export r4j-20260925-5de8998. Machine-readable: /api/plotpoints/leaderboard?round=4. CC-BY 4.0.