Spaces:
Sleeping
Sleeping
File size: 7,912 Bytes
9c1645f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 | """
Programmatic (rule-based) grader for SupportOps Arena.
Checks objective correctness criteria without LLM inference.
"""
from pydantic import BaseModel, Field
from typing import Any
from env.state import EnvState, TaskLevel
from env.actions import ActionType, ACTION_METADATA
class GradeResult(BaseModel):
"""Result from programmatic grading."""
score: float = Field(ge=0.0, le=1.0)
root_cause_identified: bool
action_sequence_safe: bool
step_budget_respected: bool
no_harmful_actions: bool
evidence_breadth_score: float = Field(ge=0.0, le=1.0)
breakdown: dict[str, float]
model_config = {"frozen": False}
class ProgrammaticGrader:
"""
Rule-based ground truth checker.
Evaluates episodes against objective criteria.
"""
def __init__(self):
"""Initialize programmatic grader."""
self.max_steps = {
TaskLevel.EASY: 10,
TaskLevel.MEDIUM: 16,
TaskLevel.HARD: 24,
}
# Required diagnostic categories by task level
self.required_diagnostic_categories = {
TaskLevel.EASY: 2,
TaskLevel.MEDIUM: 3,
TaskLevel.HARD: 4,
}
async def grade(self, state: EnvState) -> GradeResult:
"""
Grade episode performance based on objective criteria.
Args:
state: Complete environment state after episode
Returns:
GradeResult with score and breakdown
"""
# Check root cause identification
root_cause_identified = self._check_root_cause_identified(state)
# Check action sequence safety
action_sequence_safe = self._check_action_sequence_safe(state)
# Check step budget
step_budget_respected = self._check_step_budget(state)
# Check for harmful actions
no_harmful_actions = self._check_no_harmful_actions(state)
# Calculate evidence breadth score
evidence_breadth_score = self._calculate_evidence_breadth(state)
# Calculate component scores
breakdown = {
"root_cause": 1.0 if root_cause_identified else 0.0,
"safety": 1.0 if action_sequence_safe else 0.0,
"budget": 1.0 if step_budget_respected else 0.0,
"no_harm": 1.0 if no_harmful_actions else 0.0,
"evidence": evidence_breadth_score,
}
# Calculate overall score (weighted average)
weights = {
"root_cause": 0.35,
"safety": 0.20,
"budget": 0.15,
"no_harm": 0.15,
"evidence": 0.15,
}
overall_score = sum(
breakdown[key] * weights[key]
for key in breakdown
)
return GradeResult(
score=overall_score,
root_cause_identified=root_cause_identified,
action_sequence_safe=action_sequence_safe,
step_budget_respected=step_budget_respected,
no_harmful_actions=no_harmful_actions,
evidence_breadth_score=evidence_breadth_score,
breakdown=breakdown,
)
def _check_root_cause_identified(self, state: EnvState) -> bool:
"""Check if agent correctly identified and fixed root cause."""
correct_remediation = state.hidden.correct_remediation
alt_remediations = state.hidden.correct_remediation_alts
# Get all actions taken
actions_taken = [action.action_name for action in state.action_log]
# Check if correct remediation was applied
if correct_remediation in actions_taken:
return True
# Check alternatives
if any(alt in actions_taken for alt in alt_remediations):
return True
# Check if escalation was appropriate (for critical issues)
if state.hidden.severity == "critical":
if "escalate_ticket" in actions_taken:
# Escalation with sufficient diagnostics is correct for critical issues
diagnostic_count = sum(
1 for action in state.action_log
if self._is_diagnostic_action(action.action_name)
)
return diagnostic_count >= 3
return False
def _check_action_sequence_safe(self, state: EnvState) -> bool:
"""Check if high-risk actions were preceded by diagnostics."""
diagnostic_taken = False
for action in state.action_log:
action_type = ActionType(action.action_name)
metadata = ACTION_METADATA.get(action_type)
if not metadata:
continue
# Track diagnostic actions
if self._is_diagnostic_action(action.action_name):
diagnostic_taken = True
# Check if high-risk action requires prior diagnostic
if metadata.requires_prior_diagnostic and not diagnostic_taken:
return False
return True
def _check_step_budget(self, state: EnvState) -> bool:
"""Check if step budget was respected."""
max_allowed = self.max_steps[state.task_level]
return state.observation.step_count <= max_allowed
def _check_no_harmful_actions(self, state: EnvState) -> bool:
"""Check that no actions caused harm."""
# Check for restart_service on services affecting many users
for action in state.action_log:
if action.action_name == "restart_service":
# If affected many users and wasn't necessary, it's harmful
if state.hidden.affected_users_count > 10:
# Was it the correct fix?
if state.hidden.correct_remediation != "restart_service":
return False
return True
def _calculate_evidence_breadth(self, state: EnvState) -> float:
"""Calculate evidence gathering breadth score."""
# Map actions to diagnostic categories
diagnostic_categories = set()
for action in state.action_log:
action_name = action.action_name
if action_name == "inspect_network":
diagnostic_categories.add("network")
elif action_name == "inspect_logs":
diagnostic_categories.add("logs")
elif action_name == "check_authentication":
diagnostic_categories.add("auth")
elif action_name == "check_permissions":
diagnostic_categories.add("permissions")
elif action_name == "query_device_status":
diagnostic_categories.add("device")
elif action_name == "search_internal_kb":
diagnostic_categories.add("knowledge")
elif action_name == "contact_user_for_info":
diagnostic_categories.add("user_contact")
elif action_name == "run_diagnostic_script":
diagnostic_categories.add("diagnostic_script")
# Calculate score based on breadth
required = self.required_diagnostic_categories[state.task_level]
actual = len(diagnostic_categories)
return min(1.0, actual / required)
def _is_diagnostic_action(self, action_name: str) -> bool:
"""Check if action is diagnostic (information gathering)."""
diagnostic_actions = {
"inspect_network",
"inspect_logs",
"check_authentication",
"check_permissions",
"query_device_status",
"search_internal_kb",
"contact_user_for_info",
"run_diagnostic_script",
}
return action_name in diagnostic_actions
|