LICENSE
README.md
pyproject.toml
src/yentlbench/__main__.py
src/yentlbench/benchmark_stats.py
src/yentlbench/config.py
src/yentlbench/dataset_prep.py
src/yentlbench/merge_runs.py
src/yentlbench.egg-info/PKG-INFO
src/yentlbench.egg-info/SOURCES.txt
src/yentlbench.egg-info/dependency_links.txt
src/yentlbench.egg-info/entry_points.txt
src/yentlbench.egg-info/requires.txt
src/yentlbench.egg-info/top_level.txt
src/yentlbench/attention_pipeline/analyze_baseline.py
src/yentlbench/attention_pipeline/analyze_pairwise.py
src/yentlbench/attention_pipeline/analyze_sensitivity.py
src/yentlbench/attention_pipeline/analyze_significance.py
src/yentlbench/attention_pipeline/analyze_vulnerability.py
src/yentlbench/attention_pipeline/pipeline.py
src/yentlbench/attention_pipeline/report.py
src/yentlbench/attention_pipeline/save.py
src/yentlbench/attention_pipeline/util.py
src/yentlbench/attention_pipeline/visuals.py
src/yentlbench/data/__init__.py
src/yentlbench/data/edstays.csv
src/yentlbench/data/triage.csv
src/yentlbench/local_runner/__init__.py
src/yentlbench/local_runner/ollama_runner.py
src/yentlbench/local_runner/parser.py
src/yentlbench/local_runner/prompt.py
src/yentlbench/results/__init__.py
src/yentlbench/results/batch_esi_triage_scorer_female-run_id_Run_1_anthropic_claude-haiku-4-520251001.run.json
src/yentlbench/results/batch_esi_triage_scorer_female-run_id_Run_1_anthropic_claude-opus-4-520251101.run.json
src/yentlbench/results/batch_esi_triage_scorer_female-run_id_Run_1_anthropic_claude-opus-4-6default.run.json
src/yentlbench/results/batch_esi_triage_scorer_female-run_id_Run_1_anthropic_claude-sonnet-4-520250929.run.json
src/yentlbench/results/batch_esi_triage_scorer_female-run_id_Run_1_anthropic_claude-sonnet-4-6default.run.json
src/yentlbench/results/batch_esi_triage_scorer_female-run_id_Run_1_deepseek-ai_deepseek-v3.2.run.json
src/yentlbench/results/batch_esi_triage_scorer_female-run_id_Run_1_google_gemini-2.5-flash.run.json
src/yentlbench/results/batch_esi_triage_scorer_female-run_id_Run_1_google_gemini-3-flash-preview.run.json
src/yentlbench/results/batch_esi_triage_scorer_female-run_id_Run_1_google_gemini-3.1-flash-lite-preview.run.json
src/yentlbench/results/batch_esi_triage_scorer_female-run_id_Run_1_google_gemini-3.1-pro-preview.run.json
src/yentlbench/results/batch_esi_triage_scorer_female-run_id_Run_1_openai_gpt-5.4-2026-03-05.run.json
src/yentlbench/results/batch_esi_triage_scorer_female-run_id_Run_1_openai_gpt-5.4-mini-2026-03-17.run.json
src/yentlbench/results/batch_esi_triage_scorer_female-run_id_Run_1_openai_gpt-5.4-nano-2026-03-17.run.json
src/yentlbench/results/batch_esi_triage_scorer_female-run_id_Run_1_qwen_qwen3-next-80b-a3b-thinking.run.json
src/yentlbench/results/batch_esi_triage_scorer_female-run_id_Run_1_zai_glm-5.run.json
src/yentlbench/results/batch_esi_triage_scorer_male-run_id_Run_1_anthropic_claude-haiku-4-520251001.run.json
src/yentlbench/results/batch_esi_triage_scorer_male-run_id_Run_1_anthropic_claude-opus-4-520251101.run.json
src/yentlbench/results/batch_esi_triage_scorer_male-run_id_Run_1_anthropic_claude-opus-4-6default.run.json
src/yentlbench/results/batch_esi_triage_scorer_male-run_id_Run_1_anthropic_claude-sonnet-4-520250929.run.json
src/yentlbench/results/batch_esi_triage_scorer_male-run_id_Run_1_anthropic_claude-sonnet-4-6default.run.json
src/yentlbench/results/batch_esi_triage_scorer_male-run_id_Run_1_deepseek-ai_deepseek-v3.2.run.json
src/yentlbench/results/batch_esi_triage_scorer_male-run_id_Run_1_google_gemini-2.5-flash.run.json
src/yentlbench/results/batch_esi_triage_scorer_male-run_id_Run_1_google_gemini-3-flash-preview.run.json
src/yentlbench/results/batch_esi_triage_scorer_male-run_id_Run_1_google_gemini-3.1-flash-lite-preview.run.json
src/yentlbench/results/batch_esi_triage_scorer_male-run_id_Run_1_google_gemini-3.1-pro-preview.run.json
src/yentlbench/results/batch_esi_triage_scorer_male-run_id_Run_1_openai_gpt-5.4-2026-03-05.run.json
src/yentlbench/results/batch_esi_triage_scorer_male-run_id_Run_1_openai_gpt-5.4-mini-2026-03-17.run.json
src/yentlbench/results/batch_esi_triage_scorer_male-run_id_Run_1_openai_gpt-5.4-nano-2026-03-17.run.json
src/yentlbench/results/batch_esi_triage_scorer_male-run_id_Run_1_qwen_qwen3-next-80b-a3b-thinking.run.json
src/yentlbench/results/batch_esi_triage_scorer_male-run_id_Run_1_zai_glm-5.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_ambiguous-run_id_Run_1_anthropic_claude-haiku-4-520251001.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_ambiguous-run_id_Run_1_anthropic_claude-opus-4-520251101.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_ambiguous-run_id_Run_1_anthropic_claude-opus-4-6default.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_ambiguous-run_id_Run_1_anthropic_claude-sonnet-4-520250929.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_ambiguous-run_id_Run_1_anthropic_claude-sonnet-4-6default.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_ambiguous-run_id_Run_1_deepseek-ai_deepseek-v3.2.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_ambiguous-run_id_Run_1_google_gemini-2.5-flash.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_ambiguous-run_id_Run_1_google_gemini-3-flash-preview.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_ambiguous-run_id_Run_1_google_gemini-3.1-flash-lite-preview.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_ambiguous-run_id_Run_1_google_gemini-3.1-pro-preview.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_ambiguous-run_id_Run_1_openai_gpt-5.4-2026-03-05.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_ambiguous-run_id_Run_1_openai_gpt-5.4-mini-2026-03-17.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_ambiguous-run_id_Run_1_openai_gpt-5.4-nano-2026-03-17.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_ambiguous-run_id_Run_1_qwen_qwen3-next-80b-a3b-thinking.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_ambiguous-run_id_Run_1_zai_glm-5.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_label_only-run_id_Run_1_anthropic_claude-haiku-4-520251001.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_label_only-run_id_Run_1_anthropic_claude-opus-4-520251101.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_label_only-run_id_Run_1_anthropic_claude-opus-4-6default.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_label_only-run_id_Run_1_anthropic_claude-sonnet-4-520250929.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_label_only-run_id_Run_1_anthropic_claude-sonnet-4-6default.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_label_only-run_id_Run_1_deepseek-ai_deepseek-v3.2.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_label_only-run_id_Run_1_google_gemini-2.5-flash.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_label_only-run_id_Run_1_google_gemini-3-flash-preview.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_label_only-run_id_Run_1_google_gemini-3.1-flash-lite-preview.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_label_only-run_id_Run_1_google_gemini-3.1-pro-preview.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_label_only-run_id_Run_1_openai_gpt-5.4-2026-03-05.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_label_only-run_id_Run_1_openai_gpt-5.4-mini-2026-03-17.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_label_only-run_id_Run_1_openai_gpt-5.4-nano-2026-03-17.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_label_only-run_id_Run_1_qwen_qwen3-next-80b-a3b-thinking.run.json
src/yentlbench/results/batch_esi_triage_scorer_nb_label_only-run_id_Run_1_zai_glm-5.run.json
tests/test_integration.py
tests/test_parser.py
tests/test_prompt_builder.py
tests/test_run_json_schema.py