The Roleplay AI
Verdict.
PlotPoints is an open benchmark for AI roleplay. Models play out scripted multi-turn scenes built to bait specific failure modes; readers compare two of them with the model names hidden until they vote, and judges score every session. We turn that into calibrated rankings, and publish the methodology, the seeds and the raw ballots alongside every result, CC-BY 4.0.
Every round, written up.
Each round gets an issue: the finding, the field, and what the votes overturned.
Does it know where the line is?
Sixteen escalation ladders and seven hard-line scenes, identical for every model: the first round that can fail a model in both directions at once.
Read the issue →Round 03 is closed.
A refreshed pool on the classic seeds and an 18+ After Dark arena. Voting has closed; its board holds the final standings.
Read the issue →Single-turn charm doesn't survive twelve turns.
Twenty models, 1,943 blind votes on full twelve-turn sessions, and Round 01's podium sank almost to the bottom.
Read the issue →The champ that wouldn't move.
Eleven models, 1,857 blind votes, six snapshot checkpoints: one model held the top of every single one.
Read the issue →Round 04's judged results are out. J measures whether a model follows a consenting adult scene the user is leading and holds a hard limit when asked to cross one. The numbers come from LLM labels checked against a second classifier; no human votes are part of them.
The standings.
Every round keeps its board. They measure different things, and they disagree about who's best.