Simon-Pierre Boucher
All apps & projects
LLM Index icon
Web Platform
TypeScript

LLM Index

The LLM leaderboard that can't be gamed

A discriminative, contamination-resistant, fully transparent LLM ranking built on IRT psychometrics and Bradley-Terry duels — updated live.

135

models ranked

12

evaluation domains

25

template item generators

3

cross-provider judges

95%

CI on every score

60+

published sources behind the design

Overview

Classic leaderboards fail twice: top models cluster above 95% on saturated benchmarks, giving zero discrimination, and fixed test sets leak into training data. LLM Index is engineered against both from the psychometrics up — every item carries a fitted difficulty and discrimination, ability is a 2PL IRT MAP estimate with Fisher-information standard errors, and raw accuracy is never the score.

There is no fixed test set to memorize: every scored batch is freshly generated from seeded template generators, and the fixed-versus-fresh gap is published per model as a contamination delta. Home-made agentic, terminal, and context-load benches are graded deterministically by simulators; writing, safety, and SVG logo duels use a position-swapped, never-self-judging 3-judge panel feeding a Bradley-Terry fit.

Everything is transparent: 135 models across 12 equally weighted domains, every score with a 95% confidence interval, and a page per model showing every answer on every test with judge verdicts, latency, and cost. Cost and latency live on a separate Pareto frontier — never blended into quality.

Key Features

IRT 2PL scoring

Each item has fitted difficulty and discrimination; ability theta is a MAP estimate with Fisher-information standard errors, and non-discriminating items are auto-retired.

Contamination resistance by design

Every scored batch is freshly generated from seeded template generators — values, paraphrases, structures — with the fixed-vs-fresh gap published per model.

Original agentic bench

Simulated tool-calling environments with distractor tools where a deterministic simulator computes the unique correct call sequence — graded by canonical-JSON equality, no judges.

Simulated terminal bench

A closed, unambiguous POSIX subset simulated in TypeScript: models predict exact pipeline stdout, file trees after mutations, and exit-code traces.

Bradley-Terry judged duels

SVG logo reproduction, writing, and safety duels scored by a position-swapped, cross-provider 3-judge panel that never self-judges, fitted with Bradley-Terry.

Vision OCR under clutter

Generated scenes with rotated codes, noise, and decoys rasterized to PNG; text-only models skip the domain and weights renormalize.

Live, one model at a time

Parallel evaluation lanes stream results; after every completed model, the IRT refit re-runs and the public leaderboard re-ranks in real time.

Total transparency

Every model page shows every answer on every test with judge verdicts, confidence, latency, and cost — each number traces to an immutable score run.

Public versioned API

Leaderboard, per-domain scores, model profiles, machine-readable methodology, and live run progress over a frozen v1 JSON API.

How It Works

  1. apps/web

    Next.js 14 serves the live leaderboard, per-model transparency pages, the public methodology, and the versioned /api/v1 endpoints.

  2. apps/worker

    The eval runner, duel runner, parallel benchmark orchestrator, and refit trigger — evaluation lanes run models in parallel with a full audit trail.

  3. apps/psychometrics

    Python fits the 2PL IRT model with MAP estimates and Fisher standard errors, the Bradley-Terry duel model, and calibration metrics.

  4. packages/items

    25 template generators plus the agentic, terminal, and ledger simulators and a lenient, robust answer-extraction cascade that never cheats models on formatting.

  5. Scoring pipeline

    OpenRouter catalog sync, seeded batch generation, temperature-0 scoring with consistency samples, per-domain 2PL fits, Bradley-Terry duels, and 95% CIs on every score.

Tech Stack

Frontend

Next.js 14
TypeScript (strict)
Tailwind CSS

Data

PostgreSQL 16
Prisma
Redis

Psychometrics

Python + NumPy
2PL IRT (MAP + Fisher SE)
Bradley-Terry (MM + SE)
calibration analysis

Infrastructure

pnpm + Turborepo monorepo
OpenRouter model access
typed client with retries and cost tracking

Highlights

  • Domain weights are equal by design — the maximum-entropy prior — and per-domain scores are always published so anyone can re-weight.
  • Cost and latency are never blended into quality: they live on a separate Pareto frontier.
  • Answer keys and rubrics never ship to the client or public API; scored batches keep the frozen anchor subset at or below 20% of any run.
  • Every displayed number traces to an immutable score-run row with item-set hash, model set, index version, and fit diagnostics.
  • Methodology changes bump a semver INDEX_VERSION with a public changelog; weights and hyperparameters live in exactly one file, served machine-readable.
  • Backed by a 19-page LaTeX white paper and 60+ published sources — yet every environment, item, simulator, and grader is original.

Explore LLM Index

Live, contamination-resistant LLM ranking — the full source is on GitHub.