Help & Documentation

Tracks three values per entity: rating (skill estimate), deviation (uncertainty), and volatility (consistency). When an entity is inactive, deviation grows, meaning the system becomes less sure about their true skill. This is the Bayesian-correct approach: "we're forgetting, not penalizing."

Use when items go through periods of activity and inactivity (athletes between seasons, products between updates).

A pure pairwise preference model. Estimates the probability that item A beats item B as P(A>B) = exp(s_A) / (exp(s_A) + exp(s_B)). No decay, no volatility, just clean probabilities from comparisons. This is what LMSYS Chatbot Arena uses under the hood.

Use when you want calibrated win probabilities ("A has a 72% chance of being preferred over B") and don't need time decay.

The classic chess rating system. Simple formula: after each comparison, the winner gains K×(1-expected) points and the loser loses the same. Converges fast but doesn't track uncertainty. Every rating looks equally confident.

Use for casual community voting where simplicity matters more than statistical rigor.

Developed by Microsoft for Xbox matchmaking. Tracks rating (μ) and uncertainty (σ) like Glicko-2, but natively handles games with more than 2 players. A 4-player board game session updates all ratings in one pass.

Use when entities compete in groups (board game nights, multiplayer tournaments, team sports).

Open-source alternative to TrueSkill with closed-form updates (no iterative solver). Comparable accuracy, faster computation. Good for real-time leaderboards.

Use as a drop-in replacement for TrueSkill when you need faster updates.

Bayesian Elo with a logistic likelihood. Monotonic (rating always increases for wins) and robust to outlier results. Designed for competitive programming contests.

Use for formal tournament settings where you need robustness against surprise results.

Chen et al. 2013. Joint inference of entity scores AND annotator reliability η ∈ [0, 1]. Surfaces noisy raters, anti-correlated raters, and trolls explicitly. Run as a periodic audit on accumulated votes, not a live engine.

Use as a methodology audit when you need to filter low-quality annotators before RLHF export. Available via /api/v1/analytics/rankings/{id}/crowd-bt.

Plackett 1975 / Luce 1959. Generalises Bradley-Terry from pairs to ranked lists of K items (top-3 of 8, top-5 of 12). Collapses to BT on pairwise data. Connects to DeepSeek's GRPO recipe and HuggingFace TRL group preference learning.

Use when voters give ranked lists instead of pairs, OR as a second-opinion aggregator on pairwise data. Available via /api/v1/analytics/rankings/{id}/plackett-luce.

Option 1: Claude Desktop (stdio)

Add to your Claude Desktop config file:

{
  "mcpServers": {
    "aimrank": {
      "command": "python",
      "args": ["-m", "api.mcp.server"],
      "env": {
        "PYTHONPATH": "/path/to/aimrank-platform/backend",
        "DATABASE_URL": "postgresql://user:pass@host/aimrank"
      }
    }
  }
}

Option 2: HTTP (Claude Code, n8n, hosted agents)

Point your agent at the HTTP endpoint:

{
  "mcpServers": {
    "aimrank": {
      "type": "streamable-http",
      "url": "https://your-domain.com/mcp/mcp"
    }
  }
}

56 Available Tools

A sample of the 56 tools available to agents.