docs/data-boundaries.md

Data Boundaries

The repository is designed to be publishable without raw benchmark logs.

Public

data/public may contain:

• model-level aggregate scores
• task-level success/failure aggregates
• token and timing aggregates
• public external leaderboard references
• model registry metadata and public Hugging Face URLs
• task selection lists

Internal Only

data/internal_only is ignored by Git except for its README. Use it for IBM Bob comparisons, private model probes, or non-distributable aggregate tables.

Raw Data

data/raw is ignored by Git except for its README. Use it for local Terminal-Bench run directories, trial logs, transcripts, trajectories, recordings, debug payloads, and machine-specific runner state.

Publication Guard

make validate scans public text files for known internal markers and verifies that no internal/probe model identifiers are present in public aggregate data.