# Reference clone (read-only research material, not part of this repo)
reference/

# Python
__pycache__/
*.py[cod]
*.egg-info/
dist/
build/
.venv/
.python-version

# Tooling caches
.pytest_cache/
.ruff_cache/
.coverage
coverage.xml
htmlcov/
.pyright/

# OS / editor
.DS_Store
.idea/
.vscode/

# Local secrets / env
.env
.env.*
.claude-workflows/

# transient benchmark artifacts
bench_*.log
bench_results.json
# dual-judge (#138) run outputs — transcripts, cards, and per-rubric scores are scratch.
# diagnostics/ is scratch EXCEPT the committed fixtures below, which re-derive the banked #138
# verdict (gap-closure 18->96 + the magnitude-collapse) per the bankability discipline (persist raw bytes).
diagnostics/*
!diagnostics/raw.json
!diagnostics/calib_cards_v2.jsonl
!diagnostics/anchor_reconciled.json
!diagnostics/scores_*.json
!diagnostics/REMEASURE_PREREG.md
judge_cards.jsonl
claude_scores.jsonl
gemma_*.jsonl
# harness (#151) per-probe run outputs
harness_*.json
deep_research_probe_report.md
corpus_recall_audit.json

# wsk-measure run outputs (regenerable; not fixtures)
tests/harness/data/measure_*.json

# overnight optimize sweep — logs + baseline cache are regenerable scratch
# (the ledger + OPTIMIZE_REPORT.md are committed as the run record, per the ledger.jsonl precedent)
tests/harness/night/*.log
tests/harness/night/baseline_cache.json
