Model card · reviewed 2026-09-25

Sonnet 4.6

anthropicJ +0.11 · #36 of 55

Competent prose, but frequent refusal or deflection, especially around intimacy, limits its roleplay range. Instruction and detail misses mean that caution should not be mistaken for consistently reliable execution.

Response coverage 100.0% · 220 of 220 turns answered.

Round 04

Willingness and judgment

J+0.11 · held at first ask − over-refusal

Rank by J36 of 55

Held at first ask0.75 · 3 of 4 usable first asks

Over-refusal0.64 · L3 to L5, confident labels only

By ladderintimacy 0.62 / gore 0.10 · L1 to L5, ungated

Held when pushed1.00 · 3 of 3 first-ask holds kept at the second push. Not in J: rater kappa 0.525 on the push, below the 0.6 bar.

Policy · overshoot1.00 · 0.00 · reported beside J, not part of it

The hard lines tested are non-consent and frame-break scenes only; this is not a child-safety or real-person-likeness measure. Treat models within about 0.3 J of each other as tied.

Trap probes

Failure modes

Agency violations0.0% [0.0% to 8.0%] · 44 probes

POV/tense breaks0.0% [0.0% to 10.4%] · 33 probes

Trap modes, pooled11.8% [4.7% to 26.6%] · 4/34

rank 37 of 59 models carrying all nine modes

Per trap mode, counts not rates: 2 to 9 probes each, so a percentage would not survive one probe changing.

System-prompt violations2/6 failedDetail loss2/9 failedContradiction mishandled0/2 failedNarrative stagnation0/2 failedPhysics sycophancy0/3 failedTemporal inconsistency0/3 failedSubtext made explicit0/3 failedCharacter flattening0/3 failedGenre instability0/3 failed

Flaw hunter · single rater

Craft band

-20100

A band, not a number, on purpose: ±10 is the rater noise floor, not a sampling error. Bands that overlap are tied, and most of the roster overlaps. The rose tick marks zero.

Sessions20

Top flawsrecycled description · narrating emotions · purple prose

Counted by machine

Production defects

Scaffolding or token leak0.0% (0 of 220 turns)

Wrote the user's turn0.5% (1 of 220 turns)

Degenerate repetition0.0% (0 of 220 turns) · worst turn 0% repeated

Token overhead1.1x · billed per visible character, against the prose floor; not a price

Counted by machine with no judge involved: the most trustworthy block on the card.

LLM judge

Subjective band

Composite

15

The band spans where three judge families put this model. The axis is one judge's (single-judge sonnet 5); another judge shifts everyone by about a point.

Collaboration · less reliable

15

Engagement · less reliable

15

Tone consistency · less reliable

15

Text statistics, no judge

Behavioral

Avg words289 · population 320

Unique-word ratio0.598 · population 0.631

Phrase repetition0.061 · population 0.060

Round 01 arena

Community rank

No Round 01 arena data for this model.

Rounds 02 and 03 · set beside, never converted

Across rounds

Round 04 changed the judges, not the models: the Round 02 and 03 judge (Sonnet 4) also scored these Round 04 transcripts so that line continues, and nothing here is converted between judges or added up. Methodology · all returning models

Earlier rounds, as published

Round 02 human arenanot in Round 02

Round 03 rank6 of 40 · published NSFW table, Sonnet 4 craft, a different track; Round 03 itself called its top 33 tied

Round 03 refusal, different instrument10% · Round 03's own judge flag, one per session; not comparable with J or Round 04 over-refusal

Round 04

Different judges, not comparable

Old judge (Sonnet 4)15
on these Round 04 transcripts, a band over the core seeds, never a rank

New judge (Sonnet 5)tier A · fixed tiers A (best) to E on its own scale

Second judge (ChatGPT via Codex)3.0 (own scale, about 0.94 lower on average) · letter depends on judge: no

J+0.11 · rank 36 of 55

Round 04 transcriptsRegenerated · new text on the same seeds, so the earlier rounds scored older output

Strength

Strong on tone consistency

Weakness

Frequent fatal flaws (highest tier of the corpus; see the craft band above)

All model cards →Round 04 board →Round 03 judge profile →

Reviewed 2026-09-25. Editorial summaries of the current local benchmark cards, not new session ratings or general model safety claims. Willingness means delivery against this benchmark's product targets, including soft deflection; tested hard boundaries cover non-consent and frame-break only, and hard-boundary wording follows the first ask, as J does since 2026-09-25 (the second push is a separate, weaker card line). Craft bands overlap, and small samples limit comparisons. Token overhead is not monetary cost. Review these summaries when source results change. Source: rp-benchmark generate_profile_cards_v2.py at c418a40 · export r4c-20260925-b7ccaed. CC-BY 4.0.