pytest_output.txt

��============================= test session starts =============================
platform win32 -- Python 3.13.6, pytest-9.0.3, pluggy-1.6.0 -- C:\Users\Ryan\Desktop\scaler\.venv\Scripts\python.exe
cachedir: .pytest_cache
rootdir: C:\Users\Ryan\Desktop\scaler
configfile: pyproject.toml
plugins: anyio-4.13.0, cov-7.1.0
collecting ... collected 71 items

tests/test_tasks.py::TestEasyTaskGrader::test_critical_alert_correct PASSED [  1%]
tests/test_tasks.py::TestEasyTaskGrader::test_critical_alert_incorrect PASSED [  2%]
tests/test_tasks.py::TestEasyTaskGrader::test_false_positive_correct PASSED [  4%]
tests/test_tasks.py::TestEasyTaskGrader::test_episode_score_calculation PASSED [  5%]
tests/test_tasks.py::TestEasyTaskGrader::test_metrics_breakdown PASSED   [  7%]
tests/test_tasks.py::TestMediumTaskGrader::test_productive_investigation FAILED [  8%]
tests/test_tasks.py::TestMediumTaskGrader::test_wasteful_investigation FAILED [  9%]
tests/test_tasks.py::TestMediumTaskGrader::test_resource_efficiency_calculation PASSED [ 11%]
tests/test_tasks.py::TestMediumTaskGrader::test_episode_score_with_efficiency PASSED [ 12%]
tests/test_tasks.py::TestMediumTaskGrader::test_critical_missed_penalty FAILED [ 14%]
tests/test_tasks.py::TestHardTaskGrader::test_correlation_detection FAILED [ 15%]
tests/test_tasks.py::TestHardTaskGrader::test_failure_prevention_bonus FAILED [ 16%]
tests/test_tasks.py::TestHardTaskGrader::test_system_failure_penalty FAILED [ 18%]
tests/test_tasks.py::TestHardTaskGrader::test_missed_correlated_alert_penalty FAILED [ 19%]
tests/test_tasks.py::TestHardTaskGrader::test_correlation_detection_rate FAILED [ 21%]
tests/test_tasks.py::TestHardTaskGrader::test_stability_score_perfect FAILED [ 22%]
tests/test_tasks.py::TestHardTaskGrader::test_stability_score_degraded PASSED [ 23%]
tests/test_tasks.py::test_grader_reset PASSED                            [ 25%]
test_openenv_compliance.py::TestPydanticModels::test_observation_is_pydantic_model PASSED [ 26%]
test_openenv_compliance.py::TestPydanticModels::test_action_is_pydantic_model PASSED [ 28%]
test_openenv_compliance.py::TestPydanticModels::test_reward_is_pydantic_model PASSED [ 29%]
test_openenv_compliance.py::TestPydanticModels::test_episode_state_is_pydantic_model PASSED [ 30%]
test_openenv_compliance.py::TestPydanticModels::test_alert_is_pydantic_model PASSED [ 32%]
test_openenv_compliance.py::TestPydanticModels::test_observation_has_required_fields PASSED [ 33%]
test_openenv_compliance.py::TestPydanticModels::test_action_has_required_fields PASSED [ 35%]
test_openenv_compliance.py::TestPydanticModels::test_reward_has_required_fields PASSED [ 36%]
test_openenv_compliance.py::TestPydanticModels::test_action_type_is_literal PASSED [ 38%]
test_openenv_compliance.py::TestPydanticModels::test_alert_type_is_literal PASSED [ 39%]
test_openenv_compliance.py::TestPydanticModels::test_observation_serialization PASSED [ 40%]
test_openenv_compliance.py::TestPydanticModels::test_action_serialization PASSED [ 42%]
test_openenv_compliance.py::TestPydanticModels::test_reward_serialization PASSED [ 43%]
test_openenv_compliance.py::TestStepInterface::test_step_exists PASSED   [ 45%]
test_openenv_compliance.py::TestStepInterface::test_step_accepts_action PASSED [ 46%]
test_openenv_compliance.py::TestStepInterface::test_step_returns_tuple PASSED [ 47%]
test_openenv_compliance.py::TestStepInterface::test_step_returns_observation PASSED [ 49%]
test_openenv_compliance.py::TestStepInterface::test_step_returns_reward PASSED [ 50%]
test_openenv_compliance.py::TestStepInterface::test_step_returns_done PASSED [ 52%]
test_openenv_compliance.py::TestStepInterface::test_step_returns_info PASSED [ 53%]
test_openenv_compliance.py::TestStepInterface::test_info_contains_processed_alerts PASSED [ 54%]
test_openenv_compliance.py::TestStepInterface::test_info_contains_correlation_groups PASSED [ 56%]
test_openenv_compliance.py::TestStepInterface::test_info_contains_system_failure PASSED [ 57%]
test_openenv_compliance.py::TestStepInterface::test_reward_has_value PASSED [ 59%]
test_openenv_compliance.py::TestStepInterface::test_observation_updated_after_step PASSED [ 60%]
test_openenv_compliance.py::TestResetInterface::test_reset_exists PASSED [ 61%]
test_openenv_compliance.py::TestResetInterface::test_reset_returns_observation PASSED [ 63%]
test_openenv_compliance.py::TestResetInterface::test_reset_accepts_seed PASSED [ 64%]
test_openenv_compliance.py::TestResetInterface::test_reset_accepts_options PASSED [ 66%]
test_openenv_compliance.py::TestResetInterface::test_reset_reproducibility PASSED [ 67%]
test_openenv_compliance.py::TestResetInterface::test_reset_clears_episode_state PASSED [ 69%]
test_openenv_compliance.py::TestStateInterface::test_state_exists PASSED [ 70%]
test_openenv_compliance.py::TestStateInterface::test_state_returns_episode_state PASSED [ 71%]
test_openenv_compliance.py::TestStateInterface::test_episode_state_contains_observation PASSED [ 73%]
test_openenv_compliance.py::TestStateInterface::test_episode_state_contains_hidden_state PASSED [ 74%]
test_openenv_compliance.py::TestStateInterface::test_hidden_state_contains_true_severities PASSED [ 76%]
test_openenv_compliance.py::TestStateInterface::test_hidden_state_contains_correlation_groups PASSED [ 77%]
test_openenv_compliance.py::TestStateInterface::test_episode_state_contains_cumulative_reward PASSED [ 78%]
test_openenv_compliance.py::TestStateInterface::test_episode_state_contains_failures_count PASSED [ 80%]
test_openenv_compliance.py::TestStateInterface::test_episode_state_tracks_actions_taken PASSED [ 81%]
test_openenv_compliance.py::TestOpenEnvYAML::test_openenv_yaml_exists PASSED [ 83%]
test_openenv_compliance.py::TestOpenEnvYAML::test_openenv_yaml_is_valid_yaml PASSED [ 84%]
test_openenv_compliance.py::TestOpenEnvYAML::test_openenv_yaml_has_name PASSED [ 85%]
test_openenv_compliance.py::TestOpenEnvYAML::test_openenv_yaml_has_version PASSED [ 87%]
test_openenv_compliance.py::TestOpenEnvYAML::test_openenv_yaml_has_description PASSED [ 88%]
test_openenv_compliance.py::TestOpenEnvYAML::test_openenv_yaml_has_tasks PASSED [ 90%]
test_openenv_compliance.py::TestOpenEnvYAML::test_openenv_yaml_tasks_have_ids PASSED [ 91%]
test_openenv_compliance.py::TestOpenEnvYAML::test_openenv_yaml_has_config PASSED [ 92%]
test_openenv_compliance.py::TestOpenEnvYAML::test_openenv_yaml_config_has_actions PASSED [ 94%]
test_openenv_compliance.py::TestOpenEnvValidation::test_full_episode_workflow PASSED [ 95%]
test_openenv_compliance.py::TestOpenEnvValidation::test_all_task_difficulties PASSED [ 97%]
test_openenv_compliance.py::TestOpenEnvValidation::test_pydantic_validation PASSED [ 98%]
test_openenv_compliance.py::TestOpenEnvValidation::test_serialization_round_trip PASSED [100%]

================================== FAILURES ===================================
_____________ TestMediumTaskGrader.test_productive_investigation ______________
tests\test_tasks.py:142: in test_productive_investigation
    assert grader.investigations_used == 1
           ^^^^^^^^^^^^^^^^^^^^^^^^^^
E   AttributeError: 'MediumTaskGrader' object has no attribute 'investigations_used'
______________ TestMediumTaskGrader.test_wasteful_investigation _______________
tests\test_tasks.py:161: in test_wasteful_investigation
    assert contribution < 0.0, "Wasteful investigation should be penalized"
E   AssertionError: Wasteful investigation should be penalized
E   assert 0.0 < 0.0
______________ TestMediumTaskGrader.test_critical_missed_penalty ______________
tests\test_tasks.py:217: in test_critical_missed_penalty
    assert grader.critical_missed == 1
           ^^^^^^^^^^^^^^^^^^^^^^
E   AttributeError: 'MediumTaskGrader' object has no attribute 'critical_missed'
________________ TestHardTaskGrader.test_correlation_detection ________________
tests\test_tasks.py:246: in test_correlation_detection
    assert contribution > alert.true_severity, "Should get correlation bonus"
E   AssertionError: Should get correlation bonus
E   assert 0.85 > 0.85
E    +  where 0.85 = Alert(id='alert_001', visible_severity=0.8, confidence=0.85, alert_type='CPU', age=1, true_severity=0.85, is_correlated=True, metadata={}).true_severity
______________ TestHardTaskGrader.test_failure_prevention_bonus _______________
tests\test_tasks.py:269: in test_failure_prevention_bonus
    assert grader.failures_prevented >= 1, "Should register failure prevention"
           ^^^^^^^^^^^^^^^^^^^^^^^^^
E   AttributeError: 'HardTaskGrader' object has no attribute 'failures_prevented'
_______________ TestHardTaskGrader.test_system_failure_penalty ________________
tests\test_tasks.py:278: in test_system_failure_penalty
    assert grader.system_failures == 1
           ^^^^^^^^^^^^^^^^^^^^^^
E   AttributeError: 'HardTaskGrader' object has no attribute 'system_failures'
___________ TestHardTaskGrader.test_missed_correlated_alert_penalty ___________
tests\test_tasks.py:305: in test_missed_correlated_alert_penalty
    assert contribution < -2.0, "Should have extra penalty for correlated miss"
E   AssertionError: Should have extra penalty for correlated miss
E   assert -0.255 < -2.0
_____________ TestHardTaskGrader.test_correlation_detection_rate ______________
tests\test_tasks.py:316: in test_correlation_detection_rate
    grader.chains_handled.add(0)
    ^^^^^^^^^^^^^^^^^^^^^
E   AttributeError: 'HardTaskGrader' object has no attribute 'chains_handled'
_______________ TestHardTaskGrader.test_stability_score_perfect _______________
tests\test_tasks.py:326: in test_stability_score_perfect
    assert stability == 1.0, "Zero failures should give perfect stability"
E   AssertionError: Zero failures should give perfect stability
E   assert 0.99 == 1.0
=============================== tests coverage ================================
_______________ coverage: platform win32, python 3.13.6-final-0 _______________

Name                                    Stmts   Miss  Cover   Missing
---------------------------------------------------------------------
src\adaptive_alert_triage\__init__.py       5      0   100%
src\adaptive_alert_triage\env.py          197     48    76%   138, 269-275, 286-296, 463-464, 485-486, 513-520, 589, 652-684, 693-729, 733
src\adaptive_alert_triage\models.py        49      1    98%   135
src\adaptive_alert_triage\server.py       398    398     0%   19-723
src\adaptive_alert_triage\utils.py         93     21    77%   142, 225, 243, 245, 438-458
src\adaptive_alert_triage\validate.py     235    235     0%   23-417
---------------------------------------------------------------------
TOTAL                                     977    703    28%
=========================== short test summary info ===========================
FAILED tests/test_tasks.py::TestMediumTaskGrader::test_productive_investigation - AttributeError: 'MediumTaskGrader' object has no attribute 'investigations_used'
FAILED tests/test_tasks.py::TestMediumTaskGrader::test_wasteful_investigation - AssertionError: Wasteful investigation should be penalized
assert 0.0 < 0.0
FAILED tests/test_tasks.py::TestMediumTaskGrader::test_critical_missed_penalty - AttributeError: 'MediumTaskGrader' object has no attribute 'critical_missed'
FAILED tests/test_tasks.py::TestHardTaskGrader::test_correlation_detection - AssertionError: Should get correlation bonus
assert 0.85 > 0.85
 +  where 0.85 = Alert(id='alert_001', visible_severity=0.8, confidence=0.85, alert_type='CPU', age=1, true_severity=0.85, is_correlated=True, metadata={}).true_severity
FAILED tests/test_tasks.py::TestHardTaskGrader::test_failure_prevention_bonus - AttributeError: 'HardTaskGrader' object has no attribute 'failures_prevented'
FAILED tests/test_tasks.py::TestHardTaskGrader::test_system_failure_penalty - AttributeError: 'HardTaskGrader' object has no attribute 'system_failures'
FAILED tests/test_tasks.py::TestHardTaskGrader::test_missed_correlated_alert_penalty - AssertionError: Should have extra penalty for correlated miss
assert -0.255 < -2.0
FAILED tests/test_tasks.py::TestHardTaskGrader::test_correlation_detection_rate - AttributeError: 'HardTaskGrader' object has no attribute 'chains_handled'
FAILED tests/test_tasks.py::TestHardTaskGrader::test_stability_score_perfect - AssertionError: Zero failures should give perfect stability
assert 0.99 == 1.0
======================== 9 failed, 62 passed in 1.14s =========================