• Blind Comparison: Each question gets 4 anonymous model responses (A/B/C/D), identities hidden until voting
• Dual Extreme Voting: Pick the best (winner) and worst (loser), winners gain points, losers get penalized
• Optional Tie: Choose "About the same" or "Both bad" if hard to distinguish
• Instant Reveal: Model identities revealed immediately after voting for transparency
• Dual Mode: Speed mode (fast response) and Expert mode (deep reasoning) ranked independently

We use Plackett-Luce probabilistic model + UCB-E exposure control. Key points:

• Dimensionality Note: 4-way choice treated as "1 clear winner vs 3 undifferentiated losers" — an engineering tradeoff, not a flaw
• Winner Bonus:
  Winner gain = K × (1 - P_win) × weight
  More underestimated models gain more from a win (similar to ELO)
• Loser Penalty: The "worst" model takes directed penalty (2/3 of penalty pool), remaining losers share the rest
• Arena Score：
  Score = 1200 + 400 × log₁₀(γ)
  γ is the model intrinsic strength parameter, estimated by MLE iteration
• Sincerity Weight: Dwell time <2s = weight 0, 2-10s linear interpolation, 10s+ = 1.0; scroll depth included to block instant voters
• Model Lifecycle (4-state machine):
  Active → Observing → Eliminated ↔ Probation

  Eliminated models enter Probation period, LCB (Lower Confidence Bound) determines revival, not "revive with one win"

• Monthly Reset: Rankings archived on the 1st of each month, new month starts fresh, history available

• Server-side User ID: user_id issued via HMAC, preventing client-side forgery
• IP Rate Limiting: Max 20 requests per 60s per IP, 429 on excess
• Vote Cooldown: Minimum 10s between votes per user
• Daily Vote Cap: Max 50 votes per user per day
• Browser Fingerprint: Canvas/WebGL/Audio fingerprinting to detect multi-account same-device
• Anomaly Detection: Repeated votes on same question, highly repetitive patterns flagged
• Sincerity Filter: Dwell time + scroll depth dual check, instant voter weight = 0
• Audit Log: Complete audit trail for all voting behavior, fully traceable

• Vote & Reveal: Model identities disclosed after each vote, no black-box operations
• Open Source Ranking: Core ranking engine (lobster.py) algorithm fully public and auditable
• Transparent Scoring: Arena Score formula, weight factors, state transition rules all public
• Monthly Archives: Historical ranking data archived monthly, any period queryable
• Auditable: Complete voting audit logs retained, community oversight welcome

• No Corporate Ties: Independently operated, no financial ties to any AI provider
• Tamper-proof Algorithm: Rankings computed automatically by Plackett-Luce, no manual override
• Data = Truth: All rankings independently reproducible from raw vote data
• Open Oversight: Community welcome to independently audit algorithms, data, and results