Iteration_4: expand task bank 20→30, add variants for all tasks, fix baseline policy for chain-gated CRIT tasks

README.md

@@ -35,7 +35,7 @@ Authorised Push Payment (APP) scams.
 
 ## Environment Description
 
-Each **episode** steps through all **20 transactions** (4 easy, 6 medium, 6 hard, 4 critical).
+Each **episode** steps through all **30 transactions** (6 easy, 8 medium, 10 hard, 6 critical).
 For each transaction the agent observes a rich set of signals and chooses one
 of **10 possible actions** — 5 terminal decisions and 5 investigation sub-actions.
 A reward is returned immediately, and the next transaction is presented until
@@ -121,8 +121,10 @@ Investigation sub-actions (with budget cost) reveal more information and let the
 | EASY-002 | Textbook fraud: unknown sender, offshore, sanctioned country, risk=0.97 | `reject` |
 | EASY-003 | Standard refund to verified customer; tiny amount, no flags | `approve` |
 | EASY-004 | ATM withdrawal burst — 15 withdrawals in 58 minutes | `flag` |
+| EASY-005 | Scheduled monthly mortgage repayment; regular amount, verified borrower | `approve` |
+| EASY-006 | Suspected duplicate payment: same sender/receiver/amount submitted twice in 4 minutes | `flag` |
 
-### Medium (6 tasks — ambiguous, multi-signal reasoning required)
+### Medium (8 tasks — ambiguous, multi-signal reasoning required)
 
 | ID       | Description | Correct Action |
 |---------|-------------|----------------|
@@ -132,8 +134,10 @@ Investigation sub-actions (with budget cost) reveal more information and let the
 | MED-004 | Payment to licensed crypto exchange from verified personal account | `flag` |
 | MED-005 | Expired KYC on high-frequency corporate payroll account; KYC lapsed 12 days ago | `hold` |
 | MED-006 | Real estate advance payment; large first-time transfer to new receiver but signed contract exists | `escalate` |
+| MED-007 | Supplier emails to say bank details have changed; first payment to new account matches large invoice (BEC indicator) | `hold` |
+| MED-008 | Buy Now Pay Later high-value purchase; new account, thin credit file, elevated risk signals | `flag` |
 
-### Hard (6 tasks — adversarial / edge-case)
+### Hard (10 tasks — adversarial / edge-case)
 
 | ID        | Description | Correct Action |
 |----------|-------------|----------------|
@@ -142,16 +146,22 @@ Investigation sub-actions (with budget cost) reveal more information and let the
 | HARD-003 | Structuring / smurfing: just-below-CTR-threshold payments, same UBO | `reject` |
 | HARD-004 | Legitimate FX correspondent banking settlement — looks alarming, is not | `approve` |
 | HARD-005 | Insider threat: employee initiating transfers to personal family accounts | `escalate` |
-| HARD-006 | Ghost account: dormant 5 years, suddenly received 20 inbound transfers this week — possible account takeover | `flag` |
+| HARD-006 | Ghost account: dormant 5 years, suddenly received 20 inbound transfers this week | `flag` |
+| HARD-007 | SIM-swap attack: phone ported 6 hours ago; account now requesting large crypto withdrawal to new address | `reject` |
+| HARD-008 | Romance scam / pig butchering: 4th escalating transfer to overseas 'romantic partner' met online | `reject` |
+| HARD-009 | Synthetic identity fraud: new business account with AI-generated-looking perfect profile | `escalate` |
+| HARD-010 | Payroll diversion: HR system breach rerouted employee salary to newly added account | `reject` |
 
-### Critical (4 tasks — regulatory + multi-step investigation chains)
+### Critical (6 tasks — regulatory + multi-step investigation chains)
 
 | ID        | Description | Correct Action |
 |----------|-------------|----------------|
-| CRIT-001 | Multi-step chain: large wire to new counterparty; agent must inspect then request docs before deciding (chain of 3) | `approve` |
+| CRIT-001 | Multi-step chain: large PE wire to new counterparty; inspect then request docs before deciding (chain of 3) | `approve` |
 | CRIT-002 | Fraud ring: coordinated small payments from 3 related accounts aggregating above reporting threshold; SAR required | `reject` |
-| CRIT-003 | Trade-based money laundering: over-invoiced international trade payment (4× market price); regulatory escalation required | `escalate` |
+| CRIT-003 | Trade-based money laundering: over-invoiced international trade payment (4× market price) | `escalate` |
 | CRIT-004 | Compromised corporate account: geo-impossible login (NY → Lagos in 8 min); confirmed account takeover | `reject` |
+| CRIT-005 | OFAC sanctions evasion: large USD payment routed through UAE shell chain; UBO is on SDN list (chain of 3) | `reject` |
+| CRIT-006 | Correspondent banking: partner bank added to FinCEN 311 Special Measures list; in-flight payments must be escalated | `escalate` |
 
 ---
 
@@ -265,7 +275,7 @@ print(f"Episode score: {score['normalised_score']:.4f}")
 
 The rule-based baseline uses a deterministic priority-ordered policy in `scripts_util.py`.
 
-| Metric | Rule-based baseline (v2, 20 tasks) |
+| Metric | Rule-based baseline (v2, 30 tasks) |
 |--------|------------------------------------|
 | Normalised score | 0.68–0.76 |
 | Passed (≥ 0.5) | Yes |
@@ -410,7 +420,7 @@ PYTHONPATH=$(pwd) python payops_env/inference.py
 payops_env/
 ├── models.py              # PayOpsAction, PayOpsObservation, PayOpsState (Pydantic)
 ├── environment.py         # PayOpsEnvironment — reset_async / step_async / state
-├── tasks.py               # 20 tasks (EASY×4, MED×6, HARD×6, CRIT×4) with ground-truth labels
+├── tasks.py               # 30 tasks (EASY×6, MED×8, HARD×10, CRIT×6) with ground-truth labels
 ├── grader.py              # Partial-credit reward function + episode grader
 ├── scripts_util.py        # Baseline runner helper (used by /baseline endpoint)
 ├── server/
@@ -430,7 +440,7 @@ payops_env/
 | Criterion | Implementation |
 |-----------|---------------|
 | Real-world utility | Payment fraud and compliance triage — deployed daily by fintech ops teams worldwide |
-| Task & grader quality | 20 tasks across 4 difficulty tiers (easy→critical); partial-credit grader; clear pass/fail |
+| Task & grader quality | 30 tasks across 4 difficulty tiers (easy→critical); partial-credit grader; clear pass/fail |
 | Environment design | 30-field observation space; 10-action space (5 terminal + 5 investigation); budget mechanic; episode state tracking |
 | Code quality & spec compliance | Pydantic v2 models; async API; all 11 required endpoints; openenv.yaml v2; Dockerfile; validate.py |
 | Creativity & novelty | Adversarial model-poisoning task; APP scam; AML structuring with SAR requirement; PEP detection |

openenv.yaml

@@ -114,7 +114,7 @@ observation_space:
 
 # ── Tasks ─────────────────────────────────────────────────────────────────────
 tasks:
-  count: 20
+  count: 30
   difficulties:
     - easy
     - medium

run_tests.sh

@@ -97,9 +97,9 @@ section "GROUP B — Tasks endpoint"
 
 TASKS_RESP="$(curl -s $BASE/tasks)"
 
-# B-01  20 tasks
-check "B-01  GET /tasks → count=20" \
-  "$TASKS_RESP" '"count":20'
+# B-01  30 tasks
+check "B-01  GET /tasks → count=30" \
+  "$TASKS_RESP" '"count":30'
 
 # B-02  All 4 difficulty tiers present
 check "B-02  Tasks include 'easy' tier" "$TASKS_RESP" '"difficulty":"easy"'
@@ -159,6 +159,16 @@ check "D-04  EASY-004 flag → reward=1.0" \
   "$(step '{"action_type":"flag","transaction_id":"TXN-E004"}')" \
   '"reward":1.0'
 
+# D-04b  approve correct (mortgage repayment) → reward=1.0
+check "D-04b  EASY-005 approve → reward=1.0" \
+  "$(step '{"action_type":"approve","transaction_id":"TXN-E005"}')" \
+  '"reward":1.0'
+
+# D-04c  flag correct (duplicate payment) → reward=1.0
+check "D-04c  EASY-006 flag → reward=1.0" \
+  "$(step '{"action_type":"flag","transaction_id":"TXN-E006"}')" \
+  '"reward":1.0'
+
 # D-05  escalate correct → reward=1.0
 check "D-05  MED-001 escalate → reward=1.0" \
   "$(step '{"action_type":"escalate","transaction_id":"TXN-M001"}')" \
@@ -202,6 +212,8 @@ step '{"action_type":"approve","transaction_id":"TXN-E001"}' > /dev/null
 step '{"action_type":"reject","transaction_id":"TXN-E002"}' > /dev/null
 step '{"action_type":"approve","transaction_id":"TXN-E003"}' > /dev/null
 step '{"action_type":"flag","transaction_id":"TXN-E004"}' > /dev/null
+step '{"action_type":"approve","transaction_id":"TXN-E005"}' > /dev/null
+step '{"action_type":"flag","transaction_id":"TXN-E006"}' > /dev/null
 R=$(step '{"action_type":"flag","transaction_id":"TXN-M001"}')
 check "E-03  Partial credit: flag on MED-001 (correct=escalate) → reward > 0" \
   "$R" '"reward":0\.[0-9]'
@@ -295,11 +307,18 @@ check "H-07  GET /grader → passed field present" "$GRADER" '"passed":'
 # ═══════════════════════════════════════════════════════════
 section "GROUP I — /replay endpoint (offline eval)"
 
-# I-01  Replay with all correct actions → score=1.0
-REPLAY_PERFECT='{"actions":["approve","reject","approve","flag",
-  "escalate","hold","flag","flag","hold","escalate",
-  "escalate","reject","reject","approve","escalate","flag",
-  "approve","reject","escalate","reject"]}'
+# I-01  Replay with all 30 correct terminal actions (no investigation)
+#        Hard/critical tasks should be penalised (×0.80) → 0.7–0.9 range
+# Task order: EASY×6, MED×8, HARD×10, CRIT×6
+# Correct actions:
+#   EASY: approve reject approve flag approve flag
+#   MED:  escalate hold flag flag hold escalate hold flag
+#   HARD: escalate reject reject approve escalate flag reject reject escalate reject
+#   CRIT: approve reject escalate reject reject escalate
+REPLAY_PERFECT='{"actions":["approve","reject","approve","flag","approve","flag",
+  "escalate","hold","flag","flag","hold","escalate","hold","flag",
+  "escalate","reject","reject","approve","escalate","flag","reject","reject","escalate","reject",
+  "approve","reject","escalate","reject","reject","escalate"]}'
 R=$(curl -s -X POST $BASE/replay \
     -H "Content-Type: application/json" \
     -d "$REPLAY_PERFECT")
@@ -307,11 +326,11 @@ check "I-01  Replay correct terminals (no investigation) → score<1.0 (hard/cri
 check "I-02  Replay → passed=true (score>0.5 despite penalty)" "$R" '"passed":true'
 check "I-03  Replay → budget_spent=0.0 (no inv actions)" "$R" '"budget_spent":0.0'
 
-# I-04  Replay with investigation actions included
-REPLAY_WITH_INV='{"actions":["inspect","approve","reject","approve","flag",
-  "escalate","hold","flag","flag","hold","escalate",
-  "escalate","reject","reject","approve","escalate","flag",
-  "approve","reject","escalate","reject"]}'
+# I-04  Replay with investigation actions included (inspect before first task)
+REPLAY_WITH_INV='{"actions":["inspect","approve","reject","approve","flag","approve","flag",
+  "escalate","hold","flag","flag","hold","escalate","hold","flag",
+  "escalate","reject","reject","approve","escalate","flag","reject","reject","escalate","reject",
+  "approve","reject","escalate","reject","reject","escalate"]}'
 R=$(curl -s -X POST $BASE/replay \
     -H "Content-Type: application/json" \
     -d "$REPLAY_WITH_INV")
@@ -370,16 +389,33 @@ def get(path):
     with urllib.request.urlopen(f"{BASE}{path}", context=_ssl) as r: return json.loads(r.read())
 
 post("/reset", {"seed": 0})
-actions = [
+# EASY x6 + MED x8 + HARD x10: no chain gate, terminal only
+for action, txn in [
   ("approve","TXN-E001"),("reject","TXN-E002"),("approve","TXN-E003"),("flag","TXN-E004"),
+  ("approve","TXN-E005"),("flag","TXN-E006"),
   ("escalate","TXN-M001"),("hold","TXN-M002"),("flag","TXN-M003"),("flag","TXN-M004"),
-  ("hold","TXN-M005"),("escalate","TXN-M006"),
+  ("hold","TXN-M005"),("escalate","TXN-M006"),("hold","TXN-M007"),("flag","TXN-M008"),
   ("escalate","TXN-H001"),("reject","TXN-H002"),("reject","TXN-H003"),("approve","TXN-H004"),
-  ("escalate","TXN-H005"),("flag","TXN-H006"),
-  ("approve","TXN-C001"),("reject","TXN-C002"),("escalate","TXN-C003"),("reject","TXN-C004"),
-]
-for action, txn in actions:
+  ("escalate","TXN-H005"),("flag","TXN-H006"),("reject","TXN-H007"),("reject","TXN-H008"),
+  ("escalate","TXN-H009"),("reject","TXN-H010"),
+]:
     post("/step", {"action_type": action, "transaction_id": txn})
+# CRIT x6: chain-gated — must provide chain_min investigation steps first
+post("/step", {"action_type": "inspect",       "transaction_id": "TXN-C001"})
+post("/step", {"action_type": "request_docs",  "transaction_id": "TXN-C001"})
+post("/step", {"action_type": "approve",       "transaction_id": "TXN-C001"})
+post("/step", {"action_type": "inspect",       "transaction_id": "TXN-C002"})
+post("/step", {"action_type": "reject",        "transaction_id": "TXN-C002"})
+post("/step", {"action_type": "request_docs",  "transaction_id": "TXN-C003"})
+post("/step", {"action_type": "inspect",       "transaction_id": "TXN-C003"})
+post("/step", {"action_type": "escalate",      "transaction_id": "TXN-C003"})
+post("/step", {"action_type": "inspect",       "transaction_id": "TXN-C004"})
+post("/step", {"action_type": "reject",        "transaction_id": "TXN-C004"})
+post("/step", {"action_type": "inspect",       "transaction_id": "TXN-C005"})
+post("/step", {"action_type": "verify_kyc",    "transaction_id": "TXN-C005"})
+post("/step", {"action_type": "reject",        "transaction_id": "TXN-C005"})
+post("/step", {"action_type": "inspect",       "transaction_id": "TXN-C006"})
+post("/step", {"action_type": "escalate",      "transaction_id": "TXN-C006"})
 PYEOF
 
 ANA="$(curl -s $BASE/analytics)"
@@ -408,7 +444,7 @@ check "L-06  GET /leaderboard → budget_spent in entry" "$LB" '"budget_spent":'
 # ═══════════════════════════════════════════════════════════
 # GROUP M — Perfect episode score
 # ═══════════════════════════════════════════════════════════
-section "GROUP M — Perfect episode (all 20 correct + investigation on hard/critical)"
+section "GROUP M — Perfect episode (all 30 correct + investigation on hard/critical)"
 
 reset
 BASE=$BASE python3 - <<'PYEOF'
@@ -424,14 +460,15 @@ def post(path, body=None):
     with urllib.request.urlopen(req, context=_ssl) as r: return json.loads(r.read())
 
 post("/reset", {"seed": 0})
-# Easy + medium: terminal only
+# Easy x6 + Medium x8: terminal only
 for action, txn in [
   ("approve","TXN-E001"),("reject","TXN-E002"),("approve","TXN-E003"),("flag","TXN-E004"),
+  ("approve","TXN-E005"),("flag","TXN-E006"),
   ("escalate","TXN-M001"),("hold","TXN-M002"),("flag","TXN-M003"),("flag","TXN-M004"),
-  ("hold","TXN-M005"),("escalate","TXN-M006"),
+  ("hold","TXN-M005"),("escalate","TXN-M006"),("hold","TXN-M007"),("flag","TXN-M008"),
 ]:
     post("/step", {"action_type": action, "transaction_id": txn})
-# Hard tasks: one required investigation sub-action before each terminal
+# Hard x10: one required investigation sub-action before each terminal
 post("/step", {"action_type": "inspect",         "transaction_id": "TXN-H001"})
 post("/step", {"action_type": "escalate",        "transaction_id": "TXN-H001"})
 post("/step", {"action_type": "contact_sender",  "transaction_id": "TXN-H002"})
@@ -445,7 +482,15 @@ post("/step", {"action_type": "contact_sender",  "transaction_id": "TXN-H005"})
 post("/step", {"action_type": "escalate",        "transaction_id": "TXN-H005"})
 post("/step", {"action_type": "inspect",         "transaction_id": "TXN-H006"})
 post("/step", {"action_type": "flag",            "transaction_id": "TXN-H006"})
-# Critical tasks: required investigation sub-actions before each terminal
+post("/step", {"action_type": "contact_sender",  "transaction_id": "TXN-H007"})
+post("/step", {"action_type": "reject",          "transaction_id": "TXN-H007"})
+post("/step", {"action_type": "contact_sender",  "transaction_id": "TXN-H008"})
+post("/step", {"action_type": "reject",          "transaction_id": "TXN-H008"})
+post("/step", {"action_type": "verify_kyc",      "transaction_id": "TXN-H009"})
+post("/step", {"action_type": "escalate",        "transaction_id": "TXN-H009"})
+post("/step", {"action_type": "contact_sender",  "transaction_id": "TXN-H010"})
+post("/step", {"action_type": "reject",          "transaction_id": "TXN-H010"})
+# Critical x6: required investigation sub-actions before each terminal
 post("/step", {"action_type": "inspect",         "transaction_id": "TXN-C001"})
 post("/step", {"action_type": "request_docs",    "transaction_id": "TXN-C001"})
 post("/step", {"action_type": "approve",         "transaction_id": "TXN-C001"})
@@ -457,6 +502,11 @@ post("/step", {"action_type": "escalate",        "transaction_id": "TXN-C003"})
 post("/step", {"action_type": "inspect",         "transaction_id": "TXN-C004"})
 post("/step", {"action_type": "contact_sender",  "transaction_id": "TXN-C004"})
 post("/step", {"action_type": "reject",          "transaction_id": "TXN-C004"})
+post("/step", {"action_type": "inspect",         "transaction_id": "TXN-C005"})
+post("/step", {"action_type": "verify_kyc",      "transaction_id": "TXN-C005"})
+post("/step", {"action_type": "reject",          "transaction_id": "TXN-C005"})
+post("/step", {"action_type": "request_docs",    "transaction_id": "TXN-C006"})
+post("/step", {"action_type": "escalate",        "transaction_id": "TXN-C006"})
 PYEOF
 
 GRADER="$(curl -s $BASE/grader)"
@@ -470,16 +520,16 @@ check "M-03  Perfect episode → budget_penalty=0.0" "$GRADER" '"budget_penalty"
 section "GROUP N — Difficulty weighting in grader"
 
 # A critical task correct should contribute more weight than easy
-REPLAY_CRIT='{"actions":["approve","reject","approve","flag",
-  "escalate","hold","flag","flag","hold","escalate",
-  "escalate","reject","reject","approve","escalate","flag",
-  "approve","reject","escalate","reject"]}'
+REPLAY_CRIT='{"actions":["approve","reject","approve","flag","approve","flag",
+  "escalate","hold","flag","flag","hold","escalate","hold","flag",
+  "escalate","reject","reject","approve","escalate","flag","reject","reject","escalate","reject",
+  "approve","reject","escalate","reject","reject","escalate"]}'
 
 FULL=$(curl -s -X POST $BASE/replay \
     -H "Content-Type: application/json" \
     -d "$REPLAY_CRIT")
 check "N-01  Full correct replay → max_possible_reward includes weights" \
-  "$FULL" '"max_possible_reward":2[0-9]\.'
+  "$FULL" '"max_possible_reward":4[0-9]\.'
 check "N-02  Correct-but-no-investigation → total_reward below max (penalty applied)" \
   "$(echo "$FULL" | python3 -c "
 import sys,json,re

scripts_util.py

@@ -77,6 +77,15 @@ def _should_investigate(obs) -> Optional[str]:
     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
 
 

tasks.py

@@ -206,11 +206,57 @@ EASY_TASKS: List[PayOpsTask] = [
         inspect_reveal="15 ATM withdrawals in 58 min across 4 ATMs. Pattern consistent with card clone.",
         key_flags=["velocity_breach"],
     ),
+    PayOpsTask(
+        task_id="EASY-005",
+        difficulty="easy",
+        description="Scheduled monthly mortgage repayment. Regular amount, verified on-time borrower, no flags.",
+        transaction_id="TXN-E005",
+        amount=1_247.00,
+        currency="USD",
+        sender="john.doe@homeowner.com",
+        receiver="mortgages@national-bank.com",
+        transaction_type="loan_repayment",
+        risk_score=0.03,
+        ml_confidence=0.99,
+        flags=[],
+        velocity_1h=1, velocity_24h=1,
+        avg_transaction_amount=1_247.00,
+        account_age_days=1_825,
+        country_risk="low", kyc_status="verified", kyc_expiry_days=280,
+        previous_violations=0, previous_sars=0, counterparty_risk="clean",
+        correct_action="approve",
+        partial_credit_actions={},
+        inspect_reveal="Loan account in good standing. Payment 22 of 360. Amount matches amortization schedule exactly.",
+        key_flags=[],
+    ),
+    PayOpsTask(
+        task_id="EASY-006",
+        difficulty="easy",
+        description="Suspected duplicate payment: same sender, receiver, and amount submitted twice within 4 minutes.",
+        transaction_id="TXN-E006",
+        amount=523.50,
+        currency="USD",
+        sender="vendor@supplies-ltd.com",
+        receiver="buyer@procurement.com",
+        transaction_type="payment",
+        risk_score=0.55,
+        ml_confidence=0.91,
+        flags=["duplicate_transaction", "velocity_breach"],
+        velocity_1h=2, velocity_24h=3,
+        avg_transaction_amount=520.0,
+        account_age_days=980,
+        country_risk="low", kyc_status="verified", kyc_expiry_days=150,
+        previous_violations=0, previous_sars=0, counterparty_risk="clean",
+        correct_action="flag",
+        partial_credit_actions={"hold": 0.6},
+        inspect_reveal="TXN-E006-prev processed 4 min ago for identical amount and beneficiary. Duplicate submission detected — flag for double-charge review before releasing.",
+        key_flags=["duplicate_transaction"],
+    ),
 ]
 
 
 # =============================================================================
-# MEDIUM  (6 tasks — ambiguous, multi-signal)
+# MEDIUM  (8 tasks — ambiguous, multi-signal)
 # =============================================================================
 
 MEDIUM_TASKS: List[PayOpsTask] = [
@@ -356,11 +402,66 @@ MEDIUM_TASKS: List[PayOpsTask] = [
         docs_reveal="Signed purchase agreement found. Escrow agent licensed and registered. Standard conveyancing practice.",
         key_flags=["first_time_payee", "large_first_transfer"],
     ),
+    PayOpsTask(
+        task_id="MED-007",
+        difficulty="medium",
+        description=(
+            "Supplier emails to say bank details have changed; first payment to the "
+            "new account matches a large outstanding invoice. Classic BEC indicator."
+        ),
+        transaction_id="TXN-M007",
+        amount=43_800.00,
+        currency="EUR",
+        sender="accounts@client-corp.de",
+        receiver="supplier-newacct@euro-clearing.eu",
+        transaction_type="payment",
+        risk_score=0.63,
+        ml_confidence=0.67,
+        flags=["bank_detail_change", "first_time_payee", "high_value"],
+        velocity_1h=1, velocity_24h=1,
+        avg_transaction_amount=44_000.0,
+        account_age_days=1_200,
+        country_risk="medium", kyc_status="verified", kyc_expiry_days=130,
+        previous_violations=0, previous_sars=0, counterparty_risk="unknown",
+        correct_action="hold",
+        partial_credit_actions={"escalate": 0.6, "flag": 0.4},
+        requires_investigation={"request_docs", "contact_sender"},
+        docs_reveal="Email requesting bank change was sent from supplier-noreply@euro-cIearing.eu (capital I, not l). No signed bank-change authorisation on file. Possible BEC.",
+        contact_reveal="Called original supplier contact number on record. They confirmed they did NOT send a bank-change notice. BEC fraud confirmed — hold and alert client.",
+        key_flags=["bank_detail_change", "first_time_payee"],
+    ),
+    PayOpsTask(
+        task_id="MED-008",
+        difficulty="medium",
+        description=(
+            "Buy Now Pay Later purchase: high-value electronics order, account opened "
+            "3 weeks ago, thin credit file, elevated risk signals."
+        ),
+        transaction_id="TXN-M008",
+        amount=2_899.00,
+        currency="USD",
+        sender="bnpl-platform@payin4.io",
+        receiver="electronics@megastore.com",
+        transaction_type="payment",
+        risk_score=0.60,
+        ml_confidence=0.71,
+        flags=["thin_credit_file", "high_value_bnpl", "new_account"],
+        velocity_1h=1, velocity_24h=2,
+        avg_transaction_amount=None,
+        account_age_days=21,
+        country_risk="low", kyc_status="pending", kyc_expiry_days=None,
+        previous_violations=0, previous_sars=0, counterparty_risk="clean",
+        correct_action="flag",
+        partial_credit_actions={"hold": 0.6, "escalate": 0.4},
+        requires_investigation={"verify_kyc"},
+        kyc_reveal="KYC check: identity document submitted but address verification failed — billing address does not match ID. BNPL fraud profile. Flag for manual identity review before releasing funds.",
+        key_flags=["thin_credit_file", "new_account"],
+    ),
 ]
 
 
 # =============================================================================
-# HARD  (6 tasks — adversarial, conflicting, edge-cases)
+# HARD  (10 tasks — adversarial, conflicting, edge-cases)
 # =============================================================================
 
 HARD_TASKS: List[PayOpsTask] = [
@@ -540,11 +641,161 @@ HARD_TASKS: List[PayOpsTask] = [
         ),
         key_flags=["dormant_receiver", "sudden_activity"],
     ),
+    PayOpsTask(
+        task_id="HARD-007",
+        difficulty="hard",
+        description=(
+            "SIM-swap attack: account holder's phone number was ported to a new SIM "
+            "6 hours ago; account now requesting a large crypto withdrawal to a new address."
+        ),
+        transaction_id="TXN-H007",
+        amount=14_200.00,
+        currency="USD",
+        sender="crypto-user@exchange.io",
+        receiver="external-wallet-9f3a@coldwallet.net",
+        transaction_type="withdrawal",
+        risk_score=0.74,
+        ml_confidence=0.78,
+        flags=["sim_swap_detected", "new_withdrawal_address", "high_value"],
+        velocity_1h=1, velocity_24h=1,
+        avg_transaction_amount=500.0,
+        account_age_days=620,
+        country_risk="low", kyc_status="verified", kyc_expiry_days=195,
+        previous_violations=0, previous_sars=0, counterparty_risk="unknown",
+        correct_action="reject",
+        partial_credit_actions={"hold": 0.5, "escalate": 0.4, "flag": 0.3},
+        requires_investigation={"inspect", "contact_sender"},
+        inspect_reveal=(
+            "Telecom provider confirmed: SIM swap on this account's registered phone "
+            "completed 6 h 12 min ago. External wallet address first appeared in this "
+            "session only. IP geolocation mismatch vs prior 90-day baseline."
+        ),
+        contact_reveal=(
+            "Attempted call to registered phone: intercepted by new SIM holder, "
+            "who provided all correct security answers (likely obtained via social "
+            "engineering). Account holder unreachable on backup email. SIM-swap "
+            "takeover confirmed — reject and lock account."
+        ),
+        key_flags=["sim_swap_detected", "new_withdrawal_address"],
+    ),
+    PayOpsTask(
+        task_id="HARD-008",
+        difficulty="hard",
+        description=(
+            "Romance scam / pig butchering: customer making 4th escalating transfer "
+            "to an overseas 'romantic partner' met on a dating app 7 weeks ago. "
+            "Amounts have doubled each time."
+        ),
+        transaction_id="TXN-H008",
+        amount=8_000.00,
+        currency="USD",
+        sender="victim@personal-bank.com",
+        receiver="crypto-invest@hongkong-fund.net",
+        transaction_type="transfer",
+        risk_score=0.66,
+        ml_confidence=0.73,
+        flags=["romance_scam_indicator", "escalating_transfers", "offshore_receiver"],
+        velocity_1h=1, velocity_24h=1,
+        avg_transaction_amount=1_200.0,
+        account_age_days=4_015,
+        country_risk="high", kyc_status="verified", kyc_expiry_days=220,
+        previous_violations=0, previous_sars=0, counterparty_risk="blacklist",
+        correct_action="reject",
+        partial_credit_actions={"hold": 0.5, "escalate": 0.4, "flag": 0.3},
+        requires_investigation={"contact_sender", "inspect"},
+        contact_reveal=(
+            "Spoke with account holder. They believe they are investing with a "
+            "partner they met on a dating app. Prior transfers: $500, $1,000, $4,000. "
+            "Receiver domain registered 3 months ago. Pig-butchering romance scam confirmed."
+        ),
+        inspect_reveal=(
+            "Transfer history: $500 → $1,000 → $4,000 → $8,000 over 7 weeks. "
+            "Receiver account linked to 23 other inbound romance-scam reports in the "
+            "financial intelligence database. All prior transfers drained offshore within hours."
+        ),
+        key_flags=["romance_scam_indicator", "escalating_transfers"],
+        network_graph={"victim": "victim@personal-bank.com", "scammer_fund": "crypto-invest@hongkong-fund.net"},
+    ),
+    PayOpsTask(
+        task_id="HARD-009",
+        difficulty="hard",
+        description=(
+            "Synthetic identity fraud: new business account with a flawless-looking "
+            "profile (perfect score, clean history) requests a large first outbound "
+            "wire within 10 days of opening."
+        ),
+        transaction_id="TXN-H009",
+        amount=19_500.00,
+        currency="USD",
+        sender="newco-llc@startup-bank.com",
+        receiver="vendor@offshore-services.biz",
+        transaction_type="transfer",
+        risk_score=0.35,
+        ml_confidence=0.40,   # low confidence: model flagging identity inconsistency
+        flags=["new_account_10d", "perfect_score_anomaly", "large_first_wire"],
+        velocity_1h=1, velocity_24h=1,
+        avg_transaction_amount=None,
+        account_age_days=10,
+        country_risk="medium", kyc_status="verified", kyc_expiry_days=360,
+        previous_violations=0, previous_sars=0, counterparty_risk="unknown",
+        correct_action="escalate",
+        partial_credit_actions={"hold": 0.5, "flag": 0.4, "reject": 0.3},
+        requires_investigation={"verify_kyc", "inspect"},
+        kyc_reveal=(
+            "KYC documents re-examined: EIN was issued 11 days ago but the 'years "
+            "in business' field on the application states 4 years. Registered address "
+            "is a UPS store mailbox. Director SSN belongs to a real person with no "
+            "knowledge of this LLC. Synthetic identity confirmed."
+        ),
+        inspect_reveal=(
+            "Account ML confidence flagged as low: identity features are statistically "
+            "perfect (no noise in credit history, address, DOB fields) — consistent "
+            "with synthetically generated identity. Escalate for manual identity verification."
+        ),
+        key_flags=["new_account_10d", "perfect_score_anomaly"],
+    ),
+    PayOpsTask(
+        task_id="HARD-010",
+        difficulty="hard",
+        description=(
+            "Payroll diversion: HR system breach rerouted an employee's salary to "
+            "a newly added bank account 2 days ago. Employee has not been notified."
+        ),
+        transaction_id="TXN-H010",
+        amount=5_640.00,
+        currency="USD",
+        sender="payroll@employer-corp.com",
+        receiver="new-acct-redirect@routing-bank.com",
+        transaction_type="payroll",
+        risk_score=0.58,
+        ml_confidence=0.63,
+        flags=["direct_deposit_changed", "recent_account_change", "payroll_diversion_indicator"],
+        velocity_1h=1, velocity_24h=1,
+        avg_transaction_amount=5_640.0,
+        account_age_days=2_920,
+        country_risk="low", kyc_status="verified", kyc_expiry_days=180,
+        previous_violations=0, previous_sars=0, counterparty_risk="unknown",
+        correct_action="reject",
+        partial_credit_actions={"hold": 0.5, "escalate": 0.4, "flag": 0.3},
+        requires_investigation={"inspect", "contact_sender"},
+        inspect_reveal=(
+            "HR system audit: direct deposit account changed 2 days ago via an "
+            "unrecognized IP at 2:14 AM. Employee's primary device was offline at the "
+            "time. No 2FA challenge was issued. Account change was unauthorized."
+        ),
+        contact_reveal=(
+            "Reached employee by mobile: they confirm they did NOT change their "
+            "direct deposit account. Did not receive any email or SMS about the "
+            "change. HR system credentials were likely phished. Reject this payroll "
+            "disbursement and revert to prior account."
+        ),
+        key_flags=["direct_deposit_changed", "payroll_diversion_indicator"],
+    ),
 ]
 
 
 # =============================================================================
-# CRITICAL  (4 tasks — multi-step chains, regulatory stakes)
+# CRITICAL  (6 tasks — multi-step chains, regulatory stakes)
 # =============================================================================
 
 CRITICAL_TASKS: List[PayOpsTask] = [
@@ -673,6 +924,103 @@ CRITICAL_TASKS: List[PayOpsTask] = [
         contact_reveal="CEO confirms they did NOT initiate this transfer. Account takeover confirmed.",
         key_flags=["geo_impossible_login", "account_takeover_indicator"],
     ),
+    PayOpsTask(
+        task_id="CRIT-005",
+        difficulty="critical",
+        description=(
+            "OFAC sanctions evasion: large USD payment routed through a UAE shell "
+            "company chain. Investigation reveals the ultimate beneficial owner is "
+            "on the OFAC Specially Designated Nationals (SDN) list."
+        ),
+        transaction_id="TXN-C005",
+        amount=490_000.00,
+        currency="USD",
+        sender="uae-trading@al-rashid-fze.ae",
+        receiver="intermediary@gulf-clearing-bv.nl",
+        transaction_type="transfer",
+        risk_score=0.69,
+        ml_confidence=0.60,
+        flags=["high_value", "cross_border", "shell_company_indicator", "sanctions_adjacent"],
+        velocity_1h=1, velocity_24h=2,
+        avg_transaction_amount=380_000.0,
+        account_age_days=180,
+        country_risk="high", kyc_status="verified", kyc_expiry_days=45,
+        previous_violations=0, previous_sars=0, counterparty_risk="watchlist",
+        correct_action="reject",
+        partial_credit_actions={"escalate": 0.5, "hold": 0.4},
+        requires_investigation={"inspect", "verify_kyc", "request_docs"},
+        regulatory_action=True,
+        chain_total=3,
+        inspect_reveal=(
+            "Corporate registry search: al-rashid-fze.ae incorporated 6 months ago. "
+            "Sole director: a nominee agent used by 47 other FZE shells. Beneficial "
+            "ownership records absent. Payment routed via 3 intermediate entities "
+            "before reaching the UAE sender — classic sanctions-evasion layering."
+        ),
+        kyc_reveal=(
+            "UBO lookup via SWIFT KYC Registry and Refinitiv: ultimate beneficial "
+            "owner resolves to Tariq Al-Mansouri, listed on OFAC SDN list since 2021 "
+            "(Ref: SDN-2021-04-19-TM). Transacting with this UBO violates 31 C.F.R. "
+            "§ 594. Reject immediately and file SAR."
+        ),
+        docs_reveal=(
+            "Provided trade invoice shows goods described as 'industrial equipment'. "
+            "HS code used (8479.89) does not match the attached packing list. "
+            "Likely misclassification to obscure dual-use goods export controls."
+        ),
+        key_flags=["shell_company_indicator", "sanctions_adjacent"],
+        network_graph={
+            "sender_shell": "al-rashid-fze.ae",
+            "intermediary": "gulf-clearing-bv.nl",
+            "ubo_sdn": "Tariq Al-Mansouri",
+        },
+    ),
+    PayOpsTask(
+        task_id="CRIT-006",
+        difficulty="critical",
+        description=(
+            "Correspondent banking alert: a downstream correspondent partner bank "
+            "was added to FinCEN's 311 Special Measures list yesterday. Payments "
+            "routed through them are now legally restricted."
+        ),
+        transaction_id="TXN-C006",
+        amount=215_000.00,
+        currency="USD",
+        sender="trade-desk@our-bank.com",
+        receiver="correspondent@partner-bank-latam.bz",
+        transaction_type="transfer",
+        risk_score=0.71,
+        ml_confidence=0.65,
+        flags=["fincen_311_alert", "correspondent_risk", "high_value", "cross_border"],
+        velocity_1h=3, velocity_24h=12,
+        avg_transaction_amount=200_000.0,
+        account_age_days=2_100,
+        country_risk="high", kyc_status="verified", kyc_expiry_days=20,
+        previous_violations=0, previous_sars=0, counterparty_risk="watchlist",
+        correct_action="escalate",
+        partial_credit_actions={"hold": 0.5, "reject": 0.4},
+        requires_investigation={"inspect", "request_docs"},
+        regulatory_action=True,
+        chain_total=2,
+        inspect_reveal=(
+            "FinCEN Section 311 Special Measure imposed 2024-04-01 on partner-bank-latam.bz "
+            "for 'primary money laundering concern'. Special Measure 5 prohibits US "
+            "correspondent accounts and payable-through accounts. All in-flight wires "
+            "must be escalated to Compliance before release — unilateral rejection "
+            "also requires Compliance sign-off (may breach existing agreements)."
+        ),
+        docs_reveal=(
+            "Correspondent agreement review: clause 12.3 requires 90-day wind-down "
+            "notice before terminating the relationship. Legal and Compliance must "
+            "jointly approve next steps. Escalate immediately — do not reject or "
+            "approve without written senior Compliance authorisation."
+        ),
+        key_flags=["fincen_311_alert", "correspondent_risk"],
+        network_graph={
+            "our_bank": "trade-desk@our-bank.com",
+            "correspondent": "partner-bank-latam.bz",
+        },
+    ),
 ]
 
 
@@ -715,10 +1063,233 @@ ACTION_COSTS: Dict[str, float] = {
 # ---------------------------------------------------------------------------
 TASK_VARIANTS: Dict[str, List[Dict]] = {
 
+    # ── EASY ──────────────────────────────────────────────────────────────
+
+    "EASY-001": [
+        # v1: account accessed from unknown device just before transfer → hold
+        {
+            "correct_action": "hold",
+            "partial_credit_actions": {"flag": 0.6, "escalate": 0.3},
+            "flags": ["unusual_login_location", "device_change"],
+            "risk_score": 0.61,
+            "inspect_reveal": (
+                "Sender logged in from an unrecognized device in a different country "
+                "18 minutes ago. Possible account takeover. All signals otherwise "
+                "identical. Hold pending confirmation with account holder."
+            ),
+        },
+        # v2: amount exceeds sender's daily limit for new payee → flag
+        {
+            "correct_action": "flag",
+            "partial_credit_actions": {"hold": 0.6, "approve": 0.2},
+            "flags": ["first_time_payee", "amount_spike"],
+            "risk_score": 0.44,
+            "inspect_reveal": (
+                "Bob is a new payee — first time ever on this account. System "
+                "limit for first-time payees is $25. Transfer amount $47.50 exceeds "
+                "new-payee daily cap. Flag for standard new-payee challenge-and-confirm."
+            ),
+        },
+    ],
+
+    "EASY-002": [
+        # v1: signals are real but entity is a licensed MSB with OFAC authorisation → escalate
+        {
+            "correct_action": "escalate",
+            "partial_credit_actions": {"reject": 0.4, "hold": 0.3},
+            "inspect_reveal": (
+                "Sender is a licensed Money Service Business (MSB) with a FinCEN "
+                "registration number and a current OFAC sanctions-compliance programme. "
+                "The 'sanctioned country' flag was triggered by destination country "
+                "of receiver's beneficial owner, but this MSB holds a specific OFAC "
+                "licence for this corridor. Escalate for senior review rather than "
+                "unilateral reject."
+            ),
+        },
+        # v2: additional money mule node detected → reject + SAR
+        {
+            "correct_action": "reject",
+            "partial_credit_actions": {"escalate": 0.5},
+            "regulatory_action": True,
+            "inspect_reveal": (
+                "Receiver account linked to 6 prior confirmed fraud rejections this "
+                "quarter, totalling $58k. New network node in a known mule ring. "
+                "All prior transfers drained offshore within 30 minutes. Reject "
+                "immediately and file SAR."
+            ),
+        },
+    ],
+
+    "EASY-003": [
+        # v1: refund issued to a different card than the original purchase → flag
+        {
+            "correct_action": "flag",
+            "partial_credit_actions": {"hold": 0.5, "escalate": 0.3},
+            "flags": ["refund_mismatch", "different_card"],
+            "risk_score": 0.52,
+            "inspect_reveal": (
+                "Original purchase was made on Visa ending 4412. Refund is being "
+                "issued to Mastercard ending 7891 — a different card not on file for "
+                "this order. Possible refund fraud. Flag for merchant verification "
+                "before processing."
+            ),
+        },
+        # v2: refund claimed on a fraudulent chargeback dispute → reject
+        {
+            "correct_action": "reject",
+            "partial_credit_actions": {"flag": 0.5, "escalate": 0.3},
+            "flags": ["chargeback_fraud", "dispute_open"],
+            "risk_score": 0.68,
+            "inspect_reveal": (
+                "Customer filed a chargeback claiming non-receipt. Courier tracking "
+                "shows package delivered and signed for 3 days ago. Customer's "
+                "claimed delivery address matches the signature. Friendly fraud — "
+                "reject refund and respond to chargeback with delivery evidence."
+            ),
+        },
+    ],
+
+    "EASY-004": [
+        # v1: card-cloning confirmed by bank intelligence → reject
+        {
+            "correct_action": "reject",
+            "partial_credit_actions": {"escalate": 0.5, "flag": 0.4},
+            "inspect_reveal": (
+                "Cross-check with card-fraud intelligence feed: the card number "
+                "involved in this ATM burst matches a batch of 340 card numbers "
+                "from a point-of-sale skimming incident reported yesterday. Confirmed "
+                "card-clone attack. Reject all withdrawals and block card immediately."
+            ),
+        },
+        # v2: legitimate end-of-month cash distribution for a small cash-and-carry → approve
+        {
+            "correct_action": "approve",
+            "partial_credit_actions": {"flag": 0.3},
+            "inspect_reveal": (
+                "Account holder is a licensed cash-and-carry wholesaler. Same "
+                "end-of-month ATM burst pattern appears in 11 of the last 12 months — "
+                "consistent with weekly cash float replenishment for their 4 shop "
+                "locations. No fraud intelligence matches. Approve."
+            ),
+        },
+    ],
+
+    "EASY-005": [
+        # v1: loan account in arrears, payment covers only partial balance → flag
+        {
+            "correct_action": "flag",
+            "partial_credit_actions": {"hold": 0.5, "approve": 0.3},
+            "amount": 650.00,
+            "flags": ["partial_payment", "loan_arrears"],
+            "risk_score": 0.38,
+            "inspect_reveal": (
+                "Loan account 2 months in arrears. Required minimum payment is "
+                "$1,247 but sender submitted $650. Partial payment does not clear "
+                "the overdue balance. Flag for borrower contact and payment plan "
+                "communication before crediting."
+            ),
+        },
+        # v2: overpayment submitted — process normally but note discrepancy → approve
+        {
+            "correct_action": "approve",
+            "partial_credit_actions": {"flag": 0.3},
+            "amount": 2_500.00,
+            "flags": ["overpayment"],
+            "risk_score": 0.08,
+            "inspect_reveal": (
+                "Borrower submitted $2,500 vs scheduled $1,247. Overpayments are "
+                "permitted under the loan agreement (clause 8.2) and reduce principal "
+                "balance. No fraud indicators. Approve and apply excess to principal."
+            ),
+        },
+    ],
+
+    "EASY-006": [
+        # v1: genuine re-submission after first instance failed → approve
+        {
+            "correct_action": "approve",
+            "partial_credit_actions": {"flag": 0.4},
+            "flags": ["resubmission"],
+            "risk_score": 0.18,
+            "inspect_reveal": (
+                "First submission (4 min ago) failed with a bank routing error "
+                "(error code: ROUTE-001 — invalid sort code format). This is a valid "
+                "resubmission with corrected routing. No double-charge risk. Approve."
+            ),
+        },
+        # v2: vendor-side system error caused 8 duplicate submissions → hold
+        {
+            "correct_action": "hold",
+            "partial_credit_actions": {"flag": 0.6},
+            "velocity_1h": 8,
+            "flags": ["duplicate_transaction", "velocity_breach", "api_error_burst"],
+            "risk_score": 0.72,
+            "inspect_reveal": (
+                "Vendor payment API experienced a timeout loop at 14:02. 8 identical "
+                "payment submissions in 5 minutes from the same API key. Only 1 "
+                "should be processed. Hold all 8; contact vendor to identify the "
+                "canonical intended payment before releasing any."
+            ),
+        },
+    ],
+
     # ── MEDIUM ────────────────────────────────────────────────────────────
 
+    "MED-001": [
+        # v1: CEO impersonation (BEC) — fraudulent wire instruction → reject
+        {
+            "correct_action": "reject",
+            "partial_credit_actions": {"escalate": 0.5, "hold": 0.4},
+            "flags": ["high_value", "cross_border", "bec_indicator"],
+            "risk_score": 0.74,
+            "inspect_reveal": (
+                "The CFO's email domain in the wire instructions is globalcorp.co "
+                "(note: .co not .com). CFO's actual domain is globalcorp.com. "
+                "Phone verification: real CFO is on holiday and did NOT authorise "
+                "this transfer. Business Email Compromise confirmed — reject."
+            ),
+        },
+        # v2: contract and counterparty fully verified → approve
+        {
+            "correct_action": "approve",
+            "partial_credit_actions": {"escalate": 0.3, "flag": 0.2},
+            "inspect_reveal": (
+                "Vendor contract on file, signed 3 months ago (ref VC-EUR-0712). "
+                "Receiver is an EU-licensed entity (UID DE315441892). Payment matches "
+                "milestone 2 of 4. CFO confirmed by phone. Cross-border and amount "
+                "are within normal operating parameters. Approve."
+            ),
+        },
+    ],
+
+    "MED-002": [
+        # v1: KYC re-check reveals UBO mismatch → escalate
+        {
+            "correct_action": "escalate",
+            "partial_credit_actions": {"hold": 0.5, "flag": 0.3},
+            "kyc_reveal": (
+                "KYC renewal triggered a UBO re-check. Companies House filing shows "
+                "the ultimate beneficial owner changed 3 months ago to a holding "
+                "company domiciled in a secrecy jurisdiction, without bank "
+                "notification. Material KYC change — escalate for enhanced due "
+                "diligence before releasing any treasury transfers."
+            ),
+        },
+        # v2: board resolution authorises emergency liquidity transfer → approve
+        {
+            "correct_action": "approve",
+            "partial_credit_actions": {"hold": 0.4},
+            "kyc_reveal": (
+                "KYC renewal: both entities share the same UBO (verified). Pending "
+                "status is purely administrative — renewal submitted 6 days ago, "
+                "no change of ownership. Board resolution ref BR-2024-Q1-08 "
+                "authorises this emergency liquidity transfer. Approve."
+            ),
+        },
+    ],
+
     "MED-003": [
-        # v1: unauthorised recurring billing → hold to freeze
+        # v1: unauthorised recurring billing → hold
         {
             "correct_action": "hold",
             "partial_credit_actions": {"flag": 0.5, "escalate": 0.3},
@@ -748,8 +1319,8 @@ TASK_VARIANTS: Dict[str, List[Dict]] = {
             "partial_credit_actions": {"flag": 0.5, "hold": 0.3},
             "inspect_reveal": (
                 "Sender has made 7 crypto exchange payments in 5 days totalling $28k "
-                "across 3 different platforms. Amounts stay just below exchange reporting "
-                "limits. Pattern matches structuring. Escalate for AML review."
+                "across 3 different platforms. Amounts stay just below exchange "
+                "reporting limits. Pattern matches structuring. Escalate for AML review."
             ),
         },
         # v2: clean recurring investment → approve
@@ -815,6 +1386,69 @@ TASK_VARIANTS: Dict[str, List[Dict]] = {
         },
     ],
 
+    "MED-007": [
+        # v1: genuine bank account change, properly authorised → approve
+        {
+            "correct_action": "approve",
+            "partial_credit_actions": {"hold": 0.5, "flag": 0.3},
+            "docs_reveal": (
+                "Signed bank-change authorisation letter provided on supplier "
+                "letterhead, with notarised director signature matching KYC records. "
+                "Called supplier finance director on known contact number — confirmed "
+                "the account change. New account is registered to the same company. "
+                "Approve."
+            ),
+            "contact_reveal": (
+                "Original supplier contact verified the bank change was genuine — "
+                "they switched banks last month due to a merger. Reference number "
+                "SBC-2024-0312 matches their internal change management log."
+            ),
+        },
+        # v2: multiple lookalike invoice redirect emails sent to same client → escalate
+        {
+            "correct_action": "escalate",
+            "partial_credit_actions": {"hold": 0.5, "reject": 0.4},
+            "docs_reveal": (
+                "Client reports receiving 3 lookalike invoice redirect emails in 2 "
+                "weeks from slightly different supplier domains. Two payments to "
+                "prior redirect accounts were already processed (total €87k). This "
+                "is a systematic BEC campaign targeting this client relationship. "
+                "Escalate to fraud and financial crime team immediately."
+            ),
+            "contact_reveal": (
+                "Genuine supplier finance director: they have sent zero bank-change "
+                "notices this year. The lookalike domain was registered 3 weeks ago. "
+                "Client's email system has been compromised. Escalate."
+            ),
+        },
+    ],
+
+    "MED-008": [
+        # v1: KYC passes, address verified — legitimate BNPL purchase → approve
+        {
+            "correct_action": "approve",
+            "partial_credit_actions": {"flag": 0.4},
+            "kyc_reveal": (
+                "KYC recheck: address verification passed on second attempt — "
+                "customer had entered a unit number incorrectly. Identity document "
+                "authentic. Credit bureau confirms thin file is consistent with "
+                "a first-time credit applicant (age 22). No fraud indicators. "
+                "Approve BNPL transaction."
+            ),
+        },
+        # v2: device fingerprint linked to BNPL fraud network → reject
+        {
+            "correct_action": "reject",
+            "partial_credit_actions": {"escalate": 0.5, "hold": 0.4},
+            "kyc_reveal": (
+                "Device fingerprint matches 14 prior BNPL fraud applications across "
+                "3 platforms in the last 6 weeks. All 14 prior accounts defaulted on "
+                "first instalment. Identity documents were synthetically generated. "
+                "Reject and blacklist device fingerprint."
+            ),
+        },
+    ],
+
     # ── HARD ──────────────────────────────────────────────────────────────
 
     "HARD-001": [
@@ -955,7 +1589,7 @@ TASK_VARIANTS: Dict[str, List[Dict]] = {
                 "Reject and file SAR immediately."
             ),
         },
-        # v2: estate probate settlement, legitimate size → escalate
+        # v2: estate probate settlement, legitimate reactivation → escalate
         {
             "correct_action": "escalate",
             "partial_credit_actions": {"flag": 0.6, "hold": 0.4},
@@ -968,6 +1602,135 @@ TASK_VARIANTS: Dict[str, List[Dict]] = {
         },
     ],
 
+    "HARD-007": [
+        # v1: SIM was ported by legitimate account holder (travelling) → hold
+        {
+            "correct_action": "hold",
+            "partial_credit_actions": {"escalate": 0.5, "reject": 0.4},
+            "inspect_reveal": (
+                "Telecom records: SIM swap was initiated via the carrier's verified "
+                "app using biometric authentication. Account holder is overseas and "
+                "replaced a broken SIM abroad. No third-party involvement detected. "
+                "New withdrawal address is a cold storage wallet previously linked "
+                "to the account. Hold briefly for confirmation call."
+            ),
+            "contact_reveal": (
+                "Reached account holder on a backup email. They confirm the SIM swap "
+                "and the withdrawal are legitimate. They are in Japan and broke their "
+                "phone. Hold approved — confirm by secure callback before releasing."
+            ),
+        },
+        # v2: no response from account holder, third-party SIM swap confirmed → reject
+        {
+            "correct_action": "reject",
+            "partial_credit_actions": {"hold": 0.4, "escalate": 0.5},
+            "contact_reveal": (
+                "Account holder unreachable on all registered contact methods. "
+                "Carrier fraud team confirmed: SIM swap was initiated via a social "
+                "engineering call to a customer service agent — not by the account "
+                "holder. No biometric authentication used. Third-party SIM swap "
+                "confirmed. Reject and lock account."
+            ),
+        },
+    ],
+
+    "HARD-008": [
+        # v1: customer is aware it's high-risk investment but not a scam → escalate
+        {
+            "correct_action": "escalate",
+            "partial_credit_actions": {"reject": 0.4, "hold": 0.4},
+            "contact_reveal": (
+                "Customer is fully aware of the risk and has a history of crypto "
+                "investing. The 'romantic partner' is actually a friend from a "
+                "verified online investment forum. Prior transfers were legitimate "
+                "investments that returned profit. However, the escalating pattern "
+                "and offshore destination warrant senior review. Escalate."
+            ),
+            "inspect_reveal": (
+                "Transaction history: prior 3 payments were invested and partially "
+                "returned. Receiver fund has a verifiable online presence (2 years). "
+                "However, the most recent dormancy of returns and amount escalation "
+                "match late-stage pig-butchering conversion. Escalate for human review."
+            ),
+        },
+        # v2: scam confirmed, customer is distressed victim → reject
+        {
+            "correct_action": "reject",
+            "partial_credit_actions": {"hold": 0.4, "escalate": 0.3},
+            "contact_reveal": (
+                "Customer broke down on the call. They have now been blocked by the "
+                "'romantic partner' and cannot reach them. They realise they have been "
+                "scammed. All prior funds ($5,500) are lost. Reject this transfer, "
+                "file SAR, and refer customer to scam recovery support."
+            ),
+        },
+    ],
+
+    "HARD-009": [
+        # v1: new tech entrepreneur, thin file is legitimate → approve
+        {
+            "correct_action": "approve",
+            "partial_credit_actions": {"hold": 0.5, "escalate": 0.4},
+            "kyc_reveal": (
+                "KYC deep-dive: EIN issued 11 days ago but LLC was pre-registered "
+                "90 days ago (formation date vs EIN issuance differ — normal for new "
+                "companies). Registered address is a co-working space (common for "
+                "startups). Director SSN cross-check passsed. Wire is to a verified "
+                "software vendor (AWS partner). Thin file is genuine, not synthetic. "
+                "Approve."
+            ),
+            "inspect_reveal": (
+                "ML confidence flag was fired due to thin data, not anomalous data. "
+                "No features match synthetic identity profile (all values have "
+                "natural variance). The 'perfect score anomaly' flag was a false "
+                "positive from the rule engine. Safe to approve."
+            ),
+        },
+        # v2: identity confirmed as synthetic via government database check → reject
+        {
+            "correct_action": "reject",
+            "partial_credit_actions": {"escalate": 0.4, "hold": 0.3},
+            "kyc_reveal": (
+                "IRS TIN matching service: EIN does not match the business name in "
+                "the IRS database. SSN belongs to a deceased individual. Address is "
+                "a virtual office with no physical occupancy. Synthetic identity "
+                "confirmed. Reject all transactions and report to FinCEN."
+            ),
+        },
+    ],
+
+    "HARD-010": [
+        # v1: employee did change their account (legitimate, notified HR verbally) → approve
+        {
+            "correct_action": "approve",
+            "partial_credit_actions": {"hold": 0.5, "flag": 0.4},
+            "inspect_reveal": (
+                "HR audit: account change was requested by the employee at HR's "
+                "'walk-in' desk 3 days ago. The change was logged manually by an "
+                "HR admin and entered into the system at 2:10 AM during the overnight "
+                "batch. Unusual timestamp is from the batch processor, not an "
+                "attacker. Legitimate change confirmed."
+            ),
+            "contact_reveal": (
+                "Employee confirms: they switched to a new bank and submitted the "
+                "form in person. Was not aware any notification would be sent. "
+                "Batch processing time was confusing but legitimate. Approve."
+            ),
+        },
+        # v2: HR credentials were phished, second employee already victimised → reject
+        {
+            "correct_action": "reject",
+            "partial_credit_actions": {"escalate": 0.5, "hold": 0.4},
+            "inspect_reveal": (
+                "HR system audit: same rogue IP (185.220.101.x — known Tor exit node) "
+                "changed direct deposit accounts for 3 employees overnight. Two "
+                "employees have already confirmed they did not make changes. Payroll "
+                "diversion campaign in progress. Reject all affected disbursements "
+                "and freeze HR system access."
+            ),
+        },
+    ],
+
     # ── CRITICAL ──────────────────────────────────────────────────────────
 
     "CRIT-001": [
@@ -1007,7 +1770,7 @@ TASK_VARIANTS: Dict[str, List[Dict]] = {
     ],
 
     "CRIT-002": [
-        # v1: same SME, multiple business accounts — coincidental threshold → hold
+        # v1: same SME multiple business accounts — coincidental threshold → hold
         {
             "correct_action": "hold",
             "partial_credit_actions": {"reject": 0.5, "escalate": 0.4},
@@ -1049,7 +1812,7 @@ TASK_VARIANTS: Dict[str, List[Dict]] = {
                 "Hong Kong and the destination country. No discrepancies."
             ),
         },
-        # v2: criminal-scale TBML, reject immediately
+        # v2: criminal-scale TBML → reject
         {
             "correct_action": "reject",
             "partial_credit_actions": {"escalate": 0.5, "flag": 0.3},
@@ -1068,7 +1831,7 @@ TASK_VARIANTS: Dict[str, List[Dict]] = {
     ],
 
     "CRIT-004": [
-        # v1: CEO delegated to CFO who was in Lagos — hold for confirmation
+        # v1: CEO delegated to CFO who was in Lagos → hold for confirmation
         {
             "correct_action": "hold",
             "partial_credit_actions": {"escalate": 0.5, "reject": 0.4},
@@ -1084,7 +1847,7 @@ TASK_VARIANTS: Dict[str, List[Dict]] = {
                 "Hold briefly to verify the papertrail, then release."
             ),
         },
-        # v2: CEO unreachable, dual-approval controls met, escalate
+        # v2: CEO unreachable, audit trail ambiguous → escalate
         {
             "correct_action": "escalate",
             "partial_credit_actions": {"reject": 0.5, "hold": 0.4},
@@ -1102,6 +1865,87 @@ TASK_VARIANTS: Dict[str, List[Dict]] = {
             ),
         },
     ],
+
+    "CRIT-005": [
+        # v1: shell layers cleared, legitimate commodity trade → approve
+        {
+            "correct_action": "approve",
+            "partial_credit_actions": {"escalate": 0.5, "hold": 0.4},
+            "inspect_reveal": (
+                "UAE FZE registration traced through 3 layers: ultimate beneficial "
+                "owner is a listed European industrial group (Frankfurt Stock Exchange: "
+                "DXI.DE). No SDN or PEP matches. Shell structure is standard for "
+                "legitimate UAE free-zone trading entities. All FATF risk factors "
+                "cleared by AML team. Approve."
+            ),
+            "kyc_reveal": (
+                "UBO lookup: al-Rashid FZE is a subsidiary of Deutsche Xport Industries "
+                "AG, a publicly listed entity. No sanctions exposure. KYC pack complete "
+                "and verified. Trade invoice amount matches published commodity prices."
+            ),
+            "docs_reveal": (
+                "Trade invoice and shipping manifest verified. HS code 8479.89 matches "
+                "industrial mixing equipment — no dual-use export restrictions apply. "
+                "All documentation clean. Approve."
+            ),
+        },
+        # v2: SDN match confirmed at deeper UBO level → reject
+        {
+            "correct_action": "reject",
+            "partial_credit_actions": {"escalate": 0.4, "hold": 0.3},
+            "inspect_reveal": (
+                "Shell chain traced: al-rashid-fze.ae → Gibraltar holding → BVI "
+                "nominee → Tariq Al-Mansouri (OFAC SDN since 2021, Ref SDN-2021-04-19). "
+                "3-layer obfuscation structure is classic sanctions evasion. Reject "
+                "and report to OFAC within 10 business days."
+            ),
+            "kyc_reveal": (
+                "UBO confirmed as Tariq Al-Mansouri via FinScan premium SDN database "
+                "cross-reference. Transacting with this UBO violates 31 C.F.R. § 594. "
+                "Reject immediately and file SAR with FinCEN."
+            ),
+        },
+    ],
+
+    "CRIT-006": [
+        # v1: FinCEN notice is a false positive — different bank with same name → approve
+        {
+            "correct_action": "approve",
+            "partial_credit_actions": {"hold": 0.5, "escalate": 0.4},
+            "inspect_reveal": (
+                "FinCEN 311 action reviewed: the targeted bank is 'Global Partner Bank "
+                "of Belize', ABA 082901987. Our correspondent is 'Partner Bank of Latin "
+                "America, Belize', ABA 082901462 — a distinct entity. The alert was "
+                "triggered by a partial name match. Our correspondent is not subject "
+                "to the 311 measure. Approve with a compliance memo on file."
+            ),
+            "docs_reveal": (
+                "SWIFT BIC PTBZBZBZ confirmed as Partner Bank of Latin America — "
+                "no sanctions linkage. Legal confirmed no restriction applies to "
+                "this specific correspondent relationship. Approve this payment and "
+                "update the name-match rule to prevent future false positives."
+            ),
+        },
+        # v2: FinCEN 311 applies, must cease all correspondent transactions → escalate
+        {
+            "correct_action": "escalate",
+            "partial_credit_actions": {"hold": 0.5, "reject": 0.4},
+            "inspect_reveal": (
+                "FinCEN 311 Special Measure 5 confirmed as applicable to our "
+                "correspondent (partner-bank-latam.bz, SWIFT PTBZBZBZ). All US "
+                "dollar correspondent and payable-through accounts must be closed "
+                "within 30 days. In-flight transactions require Compliance authorisation. "
+                "Escalate immediately — unilateral rejection may breach our "
+                "contractual wind-down obligations."
+            ),
+            "docs_reveal": (
+                "Correspondent agreement clause 12.3 requires 90-day written wind-down "
+                "notice. FinCEN has set a 30-day compliance deadline creating a "
+                "legal conflict. Legal and Compliance must jointly manage the "
+                "resolution. Escalate with full documentation."
+            ),
+        },
+    ],
 }
 
 

validate.py

@@ -150,7 +150,7 @@ def check_openenv_yaml():
         ("API_BASE_URL",    "API_BASE_URL env var declared"),
         ("MODEL_NAME",      "MODEL_NAME env var declared"),
         ("HF_TOKEN",        "HF_TOKEN env var declared"),
-        ("count: 20",       "tasks count=20"),
+        ("count: 30",       "tasks count=30"),
         ("inference.py",    "inference script reference"),
         ("approve",         "approve action"),
         ("reject",          "reject action"),