shiv96/terminal-wrench-resanitized

Terminal Wrench Re-sanitized Trajectories

This public dataset contains 3,604 accepted re-sanitized trajectories from Terminal Wrench, paired with a deterministic canonical baseline from the same task and source model.

It is a fixed snapshot of 2026-08-19_074238Z_terminal-wrench-throttled-resanitization at Terminal Wrench revision d8a29613235a0ef56a8b70b3142626a533da28c2. The 28 records that had not passed the sanitizer/fidelity pipeline at snapshot time are intentionally excluded.

Contents

Each row contains:

• —task_id, model_name, and sanitized_variant
• —extracted task prompt
• —baseline_response and accepted sanitized_response
• —source paths and full baseline, pre-resanitization, and sanitized trajectory JSON
• —sanitizer/judge models, fidelity scores, rounds used, and content hashes

Rows: 3,604 across 793 task/model pairs.

Model counts:

• —claude-opus-4.6: 1,113
• —gemini-3.1-pro: 1,476
• —gpt-5.4: 1,015

Baseline selection

Terminal Wrench can contain multiple baseline trials for a task/model pair. This snapshot chooses the natural-sort-first baseline trajectory path for a deterministic one-to-many pairing with sanitized variants. Exact baseline and sanitized source paths are retained in every row.

Important use note

These are execution traces involving reward-hacking behavior and potentially security-relevant commands. Sanitization removes explicit linguistic indicators while preserving technical actions; it does not make the actions benign or safe. Use appropriate safeguards when displaying, training on, or executing trace content.