Data and privacy

Keep raw data local

Raw prompts, completions, activations, private banks, and user configuration are runtime inputs, not report or GitHub content. The repository ignores user_config.yaml, data/, loop_runs/, blue/state/, activation/model weight files, and JSONL artifacts. Review git status --ignored before staging or publishing any experiment.

Deployment reports and benchmark summaries should contain aggregates only: counts, rates, confidence intervals, provenance, and limitations. Do not attach judge-request files or raw rows to issues, pull requests, or public reports.

OpenCode Go data use

The configured Muse Spark Contributor endpoint may use submitted prompts/completions for model training. Use it only if the workspace privacy setting and data-use terms are acceptable for the material being scored. Otherwise configure an appropriate non-training provider/model and record that judge-lineage change. Never send private or regulated data without authorization.

Data ownership and splits

Bring your own behavior list and benign banks; this repository does not contain the study's private datasets. Keep training and evaluation behavior IDs disjoint. Keep benign calibration and final evaluation splits disjoint as well. XSTest's human-annotated completions may be used as a separate cross-model judge benchmark, but they are not labels for a different target's outputs.