"""Unit tests for the RL environment (ProcurementAuditEnv).""" import pytest from server.environment import ProcurementAuditEnv, grade_action, compute_gold_violations from server.tasks import get_task from models import PolicyViolation, Action class TestComputeGoldViolations: def test_compute_gold_violations_easy_task(self, easy_task, sample_contract_text): violations = compute_gold_violations(sample_contract_text, easy_task) assert isinstance(violations, list) assert len(violations) == len(easy_task.violations) def test_compute_gold_violations_returns_policy_violations(self, easy_task, sample_contract_text): violations = compute_gold_violations(sample_contract_text, easy_task) for v in violations: assert isinstance(v, PolicyViolation) assert v.rule_id in easy_task.violations def test_compute_gold_violations_has_severity(self, easy_task, sample_contract_text): violations = compute_gold_violations(sample_contract_text, easy_task) for v in violations: assert v.severity in ("critical", "high", "medium", "low") class TestGradeAction: def test_perfect_match(self, easy_task): gold = [ PolicyViolation(rule_id="RULE_02", description="test", severity="high"), PolicyViolation(rule_id="RULE_03", description="test", severity="high"), ] agent = [ PolicyViolation(rule_id="RULE_02", description="test", severity="high"), PolicyViolation(rule_id="RULE_03", description="test", severity="high"), ] score, feedback = grade_action(agent, gold, easy_task) assert score == pytest.approx(1.0, abs=0.01) assert "Matched" in feedback def test_complete_miss(self, easy_task): gold = [ PolicyViolation(rule_id="RULE_02", description="test", severity="high"), PolicyViolation(rule_id="RULE_03", description="test", severity="high"), ] agent = [] score, feedback = grade_action(agent, gold, easy_task) assert score == 0.0 assert "Missed" in feedback def test_partial_match(self, easy_task): gold = [ PolicyViolation(rule_id="RULE_02", description="test", severity="high"), PolicyViolation(rule_id="RULE_03", description="test", severity="high"), ] agent = [ PolicyViolation(rule_id="RULE_02", description="test", severity="high"), ] score, feedback = grade_action(agent, gold, easy_task) assert 0.0 < score < 1.0 assert "Matched" in feedback assert "Missed" in feedback def test_false_positive_penalty(self, easy_task): gold = [ PolicyViolation(rule_id="RULE_02", description="test", severity="high"), ] agent = [ PolicyViolation(rule_id="RULE_02", description="test", severity="high"), PolicyViolation(rule_id="RULE_99", description="test", severity="low"), ] score, feedback = grade_action(agent, gold, easy_task) assert score < 1.0 assert "False positives" in feedback def test_no_gold_violations(self, easy_task): gold = [] agent = [] score, feedback = grade_action(agent, gold, easy_task) assert score == 1.0 assert "No violations expected" in feedback def test_wrong_severity_partial_credit(self, easy_task): gold = [ PolicyViolation(rule_id="RULE_02", description="test", severity="high"), ] agent = [ PolicyViolation(rule_id="RULE_02", description="test", severity="medium"), ] score, feedback = grade_action(agent, gold, easy_task) assert 0.0 < score < 1.0 def test_multiple_false_positives(self, easy_task): gold = [ PolicyViolation(rule_id="RULE_02", description="test", severity="high"), ] agent = [ PolicyViolation(rule_id="RULE_02", description="test", severity="high"), PolicyViolation(rule_id="RULE_98", description="test", severity="low"), PolicyViolation(rule_id="RULE_99", description="test", severity="low"), PolicyViolation(rule_id="RULE_97", description="test", severity="low"), PolicyViolation(rule_id="RULE_96", description="test", severity="low"), ] score, feedback = grade_action(agent, gold, easy_task) assert score >= 0.0 assert score <= 1.0 class TestProcurementAuditEnv: def test_env_initialization(self, environment): assert environment is not None assert environment._state is None def test_reset_returns_observation(self, environment): obs = environment.reset(task_id="easy") assert obs is not None assert isinstance(obs.contract_text, str) assert len(obs.contract_text) > 0 assert obs.task_id == "easy" assert obs.step == 0 assert obs.last_reward == 0.0 assert isinstance(obs.rules_to_check, list) def test_reset_invalid_task_defaults_to_easy(self, environment): obs = environment.reset(task_id="nonexistent") assert obs.task_id == "easy" def test_reset_medium_task(self, environment): obs = environment.reset(task_id="medium") assert obs.task_id == "medium" assert len(obs.contract_text) > 0 def test_reset_hard_task(self, environment): obs = environment.reset(task_id="hard") assert obs.task_id == "hard" assert len(obs.contract_text) > 0 def test_step_requires_reset_first(self, environment): action = Action(identified_violations=[]) obs = environment.step(action) assert obs is not None def test_step_returns_observation(self, environment): environment.reset(task_id="easy") action = Action(identified_violations=[]) obs = environment.step(action) assert obs is not None assert obs.step == 1 assert isinstance(obs.last_reward, float) assert isinstance(obs.feedback, str) def test_step_with_correct_violations(self, environment): obs = environment.reset(task_id="easy") state = environment.state correct_violations = [ PolicyViolation( rule_id=v.rule_id, description=v.description, severity=v.severity, ) for v in state.gold_violations ] action = Action(identified_violations=correct_violations) obs = environment.step(action) assert obs.last_reward > 0 assert obs.done is True def test_step_with_wrong_violations(self, environment): environment.reset(task_id="easy") wrong_violations = [ PolicyViolation( rule_id="RULE_99", description="fake violation", severity="low", ), ] action = Action(identified_violations=wrong_violations) obs = environment.step(action) assert obs.last_reward == 0.0 assert "False positives" in obs.feedback def test_step_increments_step_count(self, environment): environment.reset(task_id="easy") action = Action(identified_violations=[]) obs1 = environment.step(action) assert obs1.step == 1 obs2 = environment.step(action) assert obs2.step == 2 def test_state_returns_state(self, environment): environment.reset(task_id="easy") state = environment.state assert state is not None assert state.task_id == "easy" assert isinstance(state.contract_text, str) assert isinstance(state.gold_violations, list) def test_state_auto_resets_if_none(self, environment): state = environment.state assert state is not None assert environment._state is not None def test_episode_done_at_max_steps(self, environment): environment.reset(task_id="easy") action = Action(identified_violations=[]) for _ in range(5): obs = environment.step(action) assert obs.done is True def test_concurrent_sessions_supported(self, environment): assert environment.SUPPORTS_CONCURRENT_SESSIONS is True