LICENSE
README.md
THIRD_PARTY_NOTICES.md
pyproject.toml
explainbench_cli.egg-info/PKG-INFO
explainbench_cli.egg-info/SOURCES.txt
explainbench_cli.egg-info/dependency_links.txt
explainbench_cli.egg-info/entry_points.txt
explainbench_cli.egg-info/requires.txt
explainbench_cli.egg-info/top_level.txt
src/core/dataset/__init__.py
src/core/dataset/extract_ground_truths/__init__.py
src/core/dataset/extract_ground_truths/effect/__init__.py
src/core/dataset/extract_ground_truths/effect/audit_files.py
src/core/dataset/extract_ground_truths/effect/build_step1.py
src/core/dataset/extract_ground_truths/effect/build_step2.py
src/core/dataset/extract_ground_truths/effect/build_step3.py
src/core/dataset/extract_ground_truths/effect/build_step4.py
src/core/dataset/extract_ground_truths/effect/build_step5.py
src/core/dataset/extract_ground_truths/effect/get_divergent_lines.py
src/core/dataset/extract_ground_truths/effect/infer_expression.py
src/core/dataset/extract_ground_truths/effect/paid_inference.py
src/core/dataset/extract_ground_truths/effect/postprocessing_util.py
src/core/dataset/extract_ground_truths/effect/process_agent_patch.py
src/core/dataset/extract_ground_truths/effect/source_util.py
src/core/dataset/extract_ground_truths/effect/trace_step1_generate_qualname_whitelist.py
src/core/dataset/extract_ground_truths/effect/trace_step2_generate_call_stack_whitelist.py
src/core/dataset/extract_ground_truths/effect/trace_util.py
src/core/dataset/extract_ground_truths/effect/prompts/template.txt
src/core/execution/__init__.py
src/core/execution/allowed_functions.json
src/core/execution/allowed_qualnames.json
src/core/execution/inspect.py
src/core/execution/trace.py
src/core/execution/track.py
src/core/execution/util.py
src/core/execution/monkey_patch/__init__.py
src/core/execution/monkey_patch/inspect.py
src/core/execution/monkey_patch/trace.py
src/core/execution/monkey_patch/track.py
src/core/tracer/__init__.py
src/core/tracer/inspector.py
src/core/tracer/protocol.py
src/core/tracer/tracer.py
src/core/tracer/util.py
src/core/tracer/serializer/__init__.py
src/core/tracer/serializer/_serializer.py
src/core/tracer/serializer/util.py
src/core/tracer/serializer/ext/__init__.py
src/core/tracer/serializer/ext/astropy.py
src/core/tracer/serializer/ext/common.py
src/core/tracer/serializer/ext/django.py
src/core/tracer/serializer/ext/matplotlib.py
src/core/tracer/serializer/ext/numpy.py
src/core/tracer/serializer/ext/numpy_ext.py
src/core/tracer/serializer/ext/pandas.py
src/core/tracer/serializer/ext/pylint.py
src/core/tracer/serializer/ext/pytest.py
src/core/tracer/serializer/ext/sklearn.py
src/core/tracer/serializer/ext/sphinx.py
src/core/tracer/serializer/ext/stdlib.py
src/core/tracer/serializer/ext/sympy.py
src/core/tracer/serializer/ext/xarray.py
src/core/tracer_plugin/__init__.py
src/core/tracer_plugin/django_plugin.py
src/core/tracer_plugin/pytest_plugin.py
src/core/tracer_plugin/sympy_plugin.py
src/explainbench/__init__.py
src/explainbench/__main__.py
src/explainbench/checker.py
src/explainbench/cli.py
src/explainbench/resources.py
src/explainbench/schemas.py
src/explainbench/submission.py
src/explainbench/data/benchmark_instance_ids.json
src/explainbench/data/context/e2e_intent.json
src/explainbench/data/context/local_intent.json
src/explainbench/data/ground_truths/e2e_intent.json
src/explainbench/data/ground_truths/local_intent.json
src/explainbench/evaluation/__init__.py
src/explainbench/evaluation/artifacts.py
src/explainbench/evaluation/checkpoints.py
src/explainbench/evaluation/choices.py
src/explainbench/evaluation/config.py
src/explainbench/evaluation/inference.py
src/explainbench/evaluation/predictions.py
src/explainbench/evaluation/preparation.py
src/explainbench/evaluation/registry.py
src/explainbench/evaluation/results.py
src/explainbench/evaluation/runner.py
src/explainbench/evaluation/schemas.py
src/explainbench/evaluation/scoring.py
src/explainbench/evaluation/service.py
src/explainbench/evaluation/tasks.py
src/explainbench/question_builders/__init__.py
src/explainbench/question_builders/common/__init__.py
src/explainbench/question_builders/common/artifacts.py
src/explainbench/question_builders/common/atomic_files.py
src/explainbench/question_builders/common/fingerprints.py
src/explainbench/question_builders/common/locking.py
src/explainbench/question_builders/common/orchestration.py
src/explainbench/question_builders/common/status.py
src/explainbench/question_builders/common/subprocess_runner.py
src/explainbench/question_builders/local/__init__.py
src/explainbench/question_builders/local/config.py
src/explainbench/question_builders/local/publication.py
src/explainbench/question_builders/local/registry.py
src/explainbench/question_builders/local/runners.py
src/explainbench/question_builders/local/service.py
src/explainbench/question_builders/local/submission_adapter.py
src/explainbench/question_builders/local/workspace.py
tests/test_checker.py
tests/test_core_source_checksums.py
tests/test_evaluate_cli.py
tests/test_evaluation_artifacts.py
tests/test_evaluation_config.py
tests/test_evaluation_inference.py
tests/test_evaluation_preparation.py
tests/test_evaluation_registry.py
tests/test_evaluation_results.py
tests/test_evaluation_runner.py
tests/test_evaluation_tasks.py
tests/test_examples.py
tests/test_execution_tracer_payload.py
tests/test_local_effect_cli_interfaces.py
tests/test_paid_inference_persistence.py
tests/test_question_builder_canonical_commands.py
tests/test_question_builder_cli.py
tests/test_question_builder_orchestration.py
tests/test_resources.py
tests/test_submission.py