What problem does it solve?
Submitting benchmark results to the Princeton GAIA leaderboard requires environment validation, cost estimation, benchmark execution, integrity auditing, and signed packaging — a multi-step process that is easy to get wrong. This Skill walks through the entire flow from a clean environment to a signed, HAL-compatible submission package.
Core Features & Use Cases
- End-to-End Benchmark Workflow: Validates prerequisites (API keys, Node.js, CLI build), estimates run cost, executes the GAIA benchmark, and reports progress every 5 questions.
- Integrity-Gated Submission Packaging: Runs a deterministic exploit audit (ADR-167) before signing, refusing to build the leaderboard package on CRITICAL findings such as answer leakage or grader monkey-patching.
- Signed HAL-Compatible Output: Produces results.jsonl, trajectories.jsonl, metadata.json, an audit report, and an Ed25519-signed witness manifest ready for leaderboard upload.
- Use Case: A researcher wants to publish their agent's GAIA Level 1 score. The Skill confirms the environment, estimates the ~$5 run cost, executes 53 questions, audits the run for exploits, and emits a signed submission directory comparable against the leaderboard top-10.
Quick Start
Ask Claude to run the GAIA benchmark at level 1 with the default model and package the results into a signed leaderboard submission.
