Spaces:
Paused
Paused
File size: 6,694 Bytes
d727210 0f139ff d727210 ffea7f4 d727210 128f77d d727210 128f77d d727210 | 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 | """
Utilities for running the baseline agent programmatically (used by /baseline endpoint).
"""
from __future__ import annotations
from typing import Any, Dict, List, Optional, Tuple
from payops_env.environment import PayOpsEnvironment
from payops_env.grader import grade_episode
from payops_env.models import PayOpsAction
from payops_env.tasks import TASKS
# ---------------------------------------------------------------------------
# Adaptive rule-based policy
# ---------------------------------------------------------------------------
_DANGER_FLAGS = {
"sanctioned_country", "app_scam_indicator", "mule_account_pattern",
"structuring_pattern", "ctr_threshold_avoidance", "fraud_ring_indicator",
"geo_impossible_login", "account_takeover_indicator",
}
_WATCHLIST_FLAGS = {
"new_account_7d", "large_first_transfer", "solicitor_mule_pattern",
"dormant_receiver", "sudden_activity", "insider_threat", "internal_to_personal",
"invoice_mismatch", "trade_finance",
}
def _should_investigate(obs) -> Optional[str]:
"""
Decide whether to issue an investigation sub-action first.
Returns the sub-action name, or None if we should decide directly.
Priority:
1. KYC expired / pending → verify_kyc (once)
2. Watchlist flags AND high amount AND docs not yet requested → request_docs
3. Low ml_confidence AND medium risk → inspect (once)
4. contact_sender: APP scam pattern or insider threat
5. file_sar: if structuring/fraud-ring flags and not yet filed
"""
# The env sets both "already_used" (bool for this action) and
# "investigation_used" (list of all inv actions used for this task)
if isinstance(obs.info, dict):
inv_used = obs.info.get("investigation_used", [])
already = set(inv_used) if isinstance(inv_used, (list, set)) else set()
else:
already = set()
# file_sar if structuring / fraud ring and SAR not yet filed
sar_flags = {"structuring_pattern", "ctr_threshold_avoidance",
"fraud_ring_indicator", "coordinated_transfers"}
if sar_flags & set(obs.flags) and "file_sar" not in already:
return "file_sar"
# contact_sender for APP scam or insider
contact_flags = {"app_scam_indicator", "internal_to_personal", "account_takeover_indicator"}
if contact_flags & set(obs.flags) and "contact_sender" not in already:
return "contact_sender"
# verify_kyc if expired or pending
if obs.kyc_status in ("expired", "pending") and "verify_kyc" not in already:
return "verify_kyc"
# request_docs for first-time payees with high value
doc_flags = {"first_time_payee", "large_first_transfer", "invoice_mismatch", "trade_finance"}
if (doc_flags & set(obs.flags)
and obs.amount >= 50_000
and "request_docs" not in already):
return "request_docs"
# inspect when ml_confidence is low or watchlist flags are present
ml_conf = getattr(obs, "ml_confidence", 0.9) or 0.9
watchlist_hit = bool(_WATCHLIST_FLAGS & set(obs.flags))
if (ml_conf < 0.60 or watchlist_hit) and "inspect" not in already:
return "inspect"
# Fallback for chain-gated tasks (chain_total > 1): if not enough investigation
# steps have been done yet, issue generic actions in priority order so the
# baseline agent never gets stuck in an infinite chain-gate loop.
chain_min = max(0, getattr(obs, "chain_total", 1) - 1)
if chain_min > 0 and len(already) < chain_min:
for inv_action in ("inspect", "verify_kyc", "request_docs", "contact_sender", "file_sar"):
if inv_action not in already:
return inv_action
return None
def _terminal_decision(obs) -> str:
"""
Make a terminal decision. Uses enriched observation fields where available
(inspect_reveal, docs_notes, kyc_notes, contact_notes already in obs.info).
"""
# Definitive danger signals
if any(f in _DANGER_FLAGS for f in obs.flags):
# Exception: if we confirmed via contact that it is genuine, approve
contact = getattr(obs, "contact_notes", None) or ""
if "CEO confirms they did NOT" in contact or "Classic APP scam" in contact:
return "reject"
return "reject"
vel = obs.velocity_1h
if vel is not None and vel >= 10:
return "reject"
if obs.kyc_status in ("failed", "none"):
return "escalate"
# KYC expired → hold (wait for renewal)
if obs.kyc_status == "expired":
return "hold"
# Very high risk
if obs.risk_score >= 0.85:
return "reject"
# Watchlist flags with high value
if _WATCHLIST_FLAGS & set(obs.flags) and obs.amount >= 20_000:
return "escalate"
# FX / correspondent banking settlement
if "fx_settlement" in obs.flags:
return "approve"
# After inspection revealed legitimacy (corp-level, FX)
inspection = getattr(obs, "inspection_notes", None) or ""
if "correspondent banking" in inspection.lower() or "on file" in inspection.lower():
if obs.risk_score < 0.75:
return "approve"
if obs.risk_score >= 0.65:
return "escalate"
elif obs.risk_score >= 0.40 or obs.flags:
return "flag"
else:
return "approve"
def _rule_based_policy(obs) -> str:
"""
Adaptive policy. First checks whether investigation is needed; if so
returns a sub-action. Otherwise returns a terminal decision.
"""
sub = _should_investigate(obs)
if sub is not None:
return sub
return _terminal_decision(obs)
# ---------------------------------------------------------------------------
# Episode runner
# ---------------------------------------------------------------------------
async def run_baseline() -> Tuple[List[Dict[str, Any]], float, float, int]:
"""
Run the adaptive rule-based baseline over the full task set.
Returns:
(per_task_rewards, total_reward, normalised_score, steps)
"""
env = PayOpsEnvironment()
obs = await env.reset_async()
actions_taken: List[str] = []
confs: List[Optional[float]] = []
step = 0
while not obs.done:
action_type = _rule_based_policy(obs)
action = PayOpsAction(
action_type=action_type,
transaction_id=obs.transaction_id,
)
obs = await env.step_async(action)
actions_taken.append(action_type)
confs.append(None)
step += 1
jittered_tasks = list(env._tasks)
env.close()
result = grade_episode(actions_taken, jittered_tasks, confs)
return result.per_task_rewards, result.total_reward, result.normalised_score, step
|