K-BENCHMARK Methodology

Evaluation Categories

Cross-Model Debate Adjudication

Unlike traditional benchmarking setups that simply verify raw responses against static tokens, ALPAR AI employs a dynamic cross-model debate. Two models independently evaluate an incident, challenge each other's assessments, and submit rebuttals; a separate referee ('Supreme Court') model then synthesizes the debate into a final Truth Score. This minimizes static scoring bias and provides robust, peer-reviewed evaluation metrics.

Cross-Audit Adjudication Chain

The cross-audit engine resolves discrepancies through a debate between two independent models, and a separate referee ('Supreme Court') model synthesizes the debate into the final Truth Score. The referee chain is defined in the source code (SUPREME_COURT_CHAIN in src/lib/ai/openrouter-gateway.ts) and currently contains these models:

This list is published exactly as coded in src/lib/ai/openrouter-gateway.ts. The adjudication chain currently contains 5 premium models with provider-level failover; when the chain changes, this page is updated to match.

Statistical Rigor

Every score published in K-BENCHMARK is accompanied by a Wilson score interval representing statistical confidence at a 95% confidence level. Re-evaluation is automatically triggered when model providers push updates, ensuring our leaderboard reflects the live capability of active models.