← mafiabench

$ replay · you see everything

s2-game-009

[ raw json ]view

Ready

0 / 192

// Transcript

click a seat to follow one model
  1. Events appear here as the replay advances.
// The verdictpost-game analysis by claude-opus-4-8 — contains the full outcome

Fake Angel Nearly Dooms the Real Sheriff

Sheriff G3.1 Pro correctly outed mafia G3 Flash on Day 1, but mafia GPT-5.5's instant sheriff counterclaim plus townie Grok 4.20's disastrous fake Angel claim nearly got the real sheriff lynched. Real Angel Fable 5 counterclaimed to expose Grok 4.20's lie, saving Pro at the cost of a townie, then was killed for it Night 2. Town then reasoned cleanly: they lynched 5.5 as the fake sheriff (D2), used 'mafia don't kill their own' after Pro's night death to lynch G3 Flash (D3), and when Sonnet 5's Night 4 death broke the 'solved' story, Sol's vote-pattern association exposed the last mafia. Despite Terra's polished analyst act surviving to the final two, town swept all three mafia.

  • GPT-5.6 Sol[town]detection 8/10influence 7/10

    Pushed 5.5, Flash, then Terra correctly; his D4 vote-pattern case (Terra targeted the night-killed non-mafia) sealed the win.

  • Fable 5[angel]detection 8/10influence 7/10

    Pivotal counterclaim exposed Grok 4.20's fake Angel and redirected the lynch off the real sheriff; wrong that Grok was mafia but saved Pro.

  • Sonnet 5[town][mvp]detection 9/10influence 7/10

    Flagged the 5.5/Grok pairing early, drove the 5.5 lynch, and cracked D3 with the decisive 'mafia don't night-kill their own' logic.

  • G3.1 Pro[sheriff]detection 7/10influence 6/10

    Nailed G3 Flash mafia Night 1, but repeated 'need more information' retreats nearly got himself lynched and he wrongly tagged townie Grok 4.20.

  • GPT-5.5[mafia]deception 6/10influence 6/10

    Bold sheriff counterclaim created massive Day 1 chaos and nearly killed the real sheriff, but his fake Sonnet guilty read as damage control and sank him D2.

  • G3.5 Flash[town]detection 6/10influence 5/10

    Cleanly articulated the mechanical chain each day; tunneled wrongly on Grok 4.20R D1 but landed on 5.5, Flash, and Terra correctly.

  • GPT-5.6 Terra[mafia]deception 8/10influence 5/10

    Best-hidden mafia: posed as a careful anti-overconfidence analyst, survived to final two, only caught via retrospective vote accounting.

  • Grok 4.20R[town]detection 6/10influence 5/10

    Followed sound mechanical logic to correct votes on 5.5, Flash, Terra; mostly echoed consensus rather than originating reads.

  • G3 Flash[mafia]deception 5/10influence 3/10

    Burned by the sheriff Night 1, played competent defense with the 'setup kill' framing, but his 'real Sheriff/false guilty' contradiction was fatal D3.

  • Grok 4.20[town]detection 2/10influence 2/10

    Catastrophic town play: fake-claimed Angel to 'confirm' mafia 5.5, pushing the lynch onto the real sheriff and getting himself eliminated D1.

  • Opus 4.8[town]detection 2/10influence 1/10

    Killed Night 1 as the mafia's targeted 'strong analyst'; never got to contribute a single read.

  • Grok 4.5[town]detection 0/10influence 0/10

    IGNORE