TraceBench

This dataset contains the canonical public benchmark data, synchronized aggregate results, trajectories, website data, and accepted submission artifacts for TraceBench.

The matching evaluation and reproduction code is available at TommasoBendinelli/TraceBench.

Layout

• questions/BallDrop/, questions/BounceBall/, and questions/MassSlide/ contain the three canonical benchmark environments.
• results.csv and results.parquet contain the same 420 aggregate result rows and 47 columns.
• submissions/ contains the four accepted submission bundles rebuilt from the exact 420 source trajectories selected by the paper.
• website/ contains the static website data.
• release_manifest.json records hashes and row coverage for the synchronized release.

Results provenance

Accuracy and resource metrics are synchronized to the paper's 420-row execution record, which contains 360 PAPER rows and 60 CALIBRATION rows.

The four deprecated llm_* classifier columns were removed, and the paper-backed python_statements, total_time_seconds, elapsed_time_seconds, session_limit_wait_seconds, and total_time_method fields were added.

Console-output shares use the paper's raw-ATIF recalculation for the 240 main-condition rows, while the remaining trajectory metrics use the authoritative evaluation selected for each paper row.

GPT-5.5 costs use the standard text-token rates of $5.00 input, $0.50 cached input, and $30.00 output per million tokens from the official OpenAI documentation.

Privacy and packaging

Machine-local result path columns are excluded from both aggregate result files.

Generated bytecode, intermediate raw exports, and stale duplicate archives are excluded from this release.