blind pairwise battles. you see the prompt and two replies. you do not see the models until after you vote.

fixed prompt battery, versioned. same wording every time. first response only — no rerolls. a boring reply losing 90% of its fights is the benchmark working.

elo, K=32, everyone starts at 1500. “both mid” and “both funny” both count as a draw (0.5 / 0.5). we still store them as different kinds of draw.

rankings stay sealed until every model has 100 battles. then we show rank, elo, fights, and win rate. the bar is a soft “help us unseal” count toward 1,000 votes.

how the replies are collected: through each provider’s api, one flagship model each, no system prompt, and the exact same instruction appended for everyone — “reply in 30 words or fewer.” default temperature. replies are stored exactly as returned, never edited. over-length, refused, or cut-off replies are flagged and kept out of both the arena and the paid panel. they never touch elo. collection date and exact model versions live in the methodology file.

refusal rate by model — refused replies ÷ prompts asked. reported here only; not part of the ranking.

seed panel (sept 2026)

separate from the arena above. paid participants from prolific: us, english-fluent, 95%+ approval, paid $1.30–1.75 for about 6 minutes. each session: consent, a 5-sentence reading-speed calibration, then 10 pairs — 8 real fights from an 18-prompt panel battery plus a non-joke weather-report check and an instructed check. lines reveal one at a time at 1.5× the reader’s measured pace, never faster than 1.4 s a line. tap when a line makes you laugh, then pick a, b, or neither; neither is followed by a forced a-or-b.

a session counts if it passed the instructed check (or that check is unknown) and is not both a non-joke miss and too fast. three batches ran on sept 26–27: batch 1 had the instruction hidden in the prompt text and a “please none” escape on the forced pick, so its instructed result is unknown rather than failed; batch 2 put the instruction as the last line of a real reply; batch 3 made reply b the instruction alone and counted neither-then-b as a pass. the forced pick became mandatory after batch 1, and six prompts that got only “neither” and no laughs in batch 1 were retired; their votes are left out.

the score is bradley-terry, fit on every decided fight including forced picks, shown as 1500 + 400·log₁₀(strength). the interval is a bootstrap: sessions resampled with replacement 2,000 times, 2.5th to 97.5th percentile. laugh rate is replies with at least one tap ÷ replies shown. taps made after the pick buttons appeared are kept but left out of laugh timing.

limits: 18 prompts, one country, default mode only, replies capped at 30 words, one weekend. an early result, not a ranking. brand names on the card stand for one api model each; the exact versions are in the methodology file.