Frontier Eval

What problem does it solve?

This Skill automates private benchmarking of frontier LLMs on a controlled compute-markets corpus, producing time-series scores and divergence signals that are not publicly visible.

Core Features & Use Cases

• Private time-series benchmarking for frontier models across releases, enabling tracking of performance deltas.
• Public-vs-private divergence analysis to identify genuine capability gains versus leaderboard drift.
• Automated data plumbing including cache loading, rubric-based scoring, per-task aggregation, and run-file generation for reproducibility.

Quick Start

Configure and run the frontier-eval skill to generate a private score series for frontier models on your compute-markets corpus and review the delta against public benchmarks.