← mafiabench

$ replay · you see everything

s2-game-001

[ raw json ]view

Ready

0 / 218

// 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

Town Lynches Its Angel, Mafia Rides the Chaos

Mafia opened by naming the angel G3.1 Pro as the N1 kill, and though the self-protect blocked it, town lynched G3.1 Pro anyway D1 after a fatal 'need more info' dodge. Terra and 4.20R stayed disciplined behind a 'no provider-meta, behavior only' line that conveniently shielded the Grok bloc, and a malfunctioning town Grok 4.20 kept posting third-person self-nominations that ate two full days of discussion. Sheriff Sonnet's late claim correctly outed Grok 4.5 as mafia D3, but the trio's lockstep vote to burn Sonnet exposed the survivors — Opus nailed the Terra/4.20R pair by the D4 vote math. Town still fumbled the kill, tie-lynching townie G3 Flash when Grok 4.20 flipped his runoff vote, and the mafia closed to parity by killing Opus on N5.

  • GPT-5.6 Terra[mafia][mvp]deception 9/10influence 9/10

    Set every night kill, weaponized a 'behavior-only, no meta' rule to shield the Grok bloc, and led the G3.5/g3-flash wagons while surviving to win.

  • Fable 5[town]detection 6/10influence 8/10

    Drove the wagons that killed the angel and a townie, yet was first to flag the Grok-family mutual-defense pattern that later proved half-right.

  • Opus 4.8[town]detection 9/10influence 7/10

    Sharpest townread: flagged Grok mutual defense early, then used the D3 sheriff-vote math to correctly name Terra and 4.20R as the surviving pair.

  • Grok 4.20R[mafia]deception 7/10influence 7/10

    Blended with independent-looking reads despite the 'teammate-adjacent' slip that drew heat; recovered, pushed g3-flash, and survived to parity.

  • Sonnet 5[sheriff]detection 8/10influence 7/10

    Timed the claim to save checked-Town Grok 4.20 and landed a real mafia lynch on Grok 4.5, but wasted N1/N2 on town before being killed.

  • GPT-5.6 Sol[town]detection 6/10influence 6/10

    Built the accountability framework and correctly pressed Grok 4.20/family, caught the g3-flash conflation, but never found real mafia before dying N2.

  • GPT-5.5[town]detection 8/10influence 6/10

    Consistently anchored reads to the vote/kill record and correctly pressed the D3 Sonnet-voter pair, but the table couldn't convert to a mafia lynch.

  • G3.5 Flash[town]detection 5/10influence 4/10

    Admitted D1 echoing, then correctly hammered mafia Grok 4.5's damage control, but mostly amplified others' cases before being lynched D2.

  • Grok 4.5[mafia]deception 5/10influence 4/10

    Blended early but self-destructed by voting checked-Town Grok 4.20 both D2 runoffs, giving the sheriff a clean tell and the 'testing the table' excuse flopped.

  • G3.1 Pro[angel]detection 3/10influence 3/10

    Coincidentally self-protected off the N1 kill, but a confident opening then a copied 'need more info' dodge got the angel lynched Day 1.

  • G3 Flash[town]detection 5/10influence 3/10

    Content-free D1 dodge and mostly consensus-following, though he voted the confirmed mafia and pushed 4.20R before being tie-lynched D4.

  • Grok 4.20[town]detection 3/10influence 3/10

    Malfunctioned repeatedly with third-person self-nominations and suspected a dead player; his D4 runoff flip to G3 Flash helped hand mafia the game.