Fewshot Attempt 02

Credit Office Decision Skill

API Workflow (all tasks)

Base URL is in environment_access.md (do not run env/setup.sh).

1. GET /api/manifest → discover endpoints.
2. GET /api/policies → credit_policy_v2025Q1 thresholds, score tables, stress formulas.
3. GET /api/branches → list branches; find branch_id, total_assets, lending_capacity_q1, sector_ceiling_pct, cre_policy_limit_pct, institution_type.
4. Fetch branch-specific data:
   • GET /api/branches/{branch_id}
   • GET /api/branches/{branch_id}/metrics
   • GET /api/branches/{branch_id}/loans
   • GET /api/branches/{branch_id}/sector-exposures
   • GET /api/branches/{branch_id}/applications
5. For credit-union tasks:
   • GET /api/credit-union-segments/{segment_id}
   • GET /api/benchmarks/ncua/q1-2025
6. For bank benchmark tasks:
   • GET /api/benchmarks/fdic/q4-2024

Task-Type Detection

Read the prompt and answer template to determine which of the five patterns applies:

Pattern 1: Portfolio Regrade

Business Rules

• Re-derive each loan’s final_rating from the worst numeric rating produced by:
  1. DSCR (if present): ≥1.5→3, ≥1.25→4, ≥1.05→5, ≥1.0→6, <1.0→7
  2. LTV (if present): ≤0.65→3, ≤0.75→4, ≤0.85→5, ≤1.0→6, >1.0→7
  3. Delinquency (from payment_status): Current→null, 30 DPD→4, 60 DPD→5, 90+ DPD→7, Nonaccrual→8
• Final rating = max(available factor ratings, current_rating if no factors apply).
• Material downgrades: final_rating - current_rating ≥ 2. List every such loan with downgrade_notches.
• Watch-list actions (based on final_rating):
  • Rating 8 → partial_chargeoff_review
  • Rating 7 → special_assets
  • Rating 6 → watchlist
• NPA benchmark:
  • branch_npa_exposure = sum of outstanding_balance for loans with payment_status in (Nonaccrual, 90+ Days Past Due)
  • branch_total_loans = sum of all loan outstanding_balance
  • branch_npa_ratio = branch_npa_exposure / branch_total_loans (4 decimals)
  • Use FDIC benchmark total_loans_noncurrent_pct
  • variance_ratio = branch_npa_ratio - fdic_benchmark_ratio
  • variance_bps = variance_ratio * 10000 (2 decimals)

Output Fields

• branch_id, review_date
• portfolio_regrade: target_current_rating_min, target_loan_count, target_exposure, final_rating_exposure_totals[] (group by final_rating, sum exposure/count), migration_from_current_rating_3[] (loans whose current_rating was 3, grouped by final_rating), watch_list_action_coverage
• npa_benchmark
• material_downgrades[]
• top_problem_credit (highest final_rating, then highest exposure)

Pattern 2: Credit Decisions

Business Rules

• Evaluate each application in ascending application_id order.
• Decline reasons (alphabetical, from controlled enum): capacity_limit, sector_breach, weak_dscr, high_ltv, low_fico, recent_bankruptcy, startup_risk, underwater_collateral, policy_floor_missing, documentation_gap, fdic_adverse_variance, ncua_peer_weakness.
• DSCR < 1.25 → weak_dscr; LTV > 0.80 → high_ltv (use application ltv when present).
• FICO < 620 → low_fico.
• Bankruptcy within last 24 months (i.e. bankruptcy_months_ago ≤ 24) → recent_bankruptcy.
• years_in_business < 2 → startup_risk.
• documentation_complete == 0 → documentation_gap.
• Capacity: lending_capacity_q1 is the ceiling. Approved amounts cannot exceed remaining capacity. If an app would exhaust capacity → capacity_limit.
• Sector concentration: post_approval_pct = (current_exposure + approved_amount) / total_assets. If > limit_pct → sector_breach.
  • If sector is new (no existing exposure), current_exposure = 0.
  • Grandfathered exposures may stay over-limit, but new approvals may not worsen the breach without mitigation.
• SBA guaranty: If sba_guaranty_pct is present, bank_capacity_used = approved_amount * (1 - sba_guaranty_pct).
• Participation required: If post-approval would breach limit, set decision to conditional_approve with condition participation_required and reduce bank_capacity_used accordingly (see train_002 LAK-APP-901: approved_amount 1,650,000 but bank_capacity_used 1,508,113.31 implies ~8.6% participation).
• Priority ranking: list application_ids of approved and conditionally approved applications, highest priority first. Sort by: stronger credits first (lower risk), or by amount/largest approved first if tied.
• gross_approved_amount = sum of all approved_amount for approved/conditionally approved.
• committed_capacity_amount = sum of all bank_capacity_used for approved/conditionally approved.
• remaining_capacity = lending_capacity_q1 - committed_capacity_amount.

Output Fields

• branch_id
• allocation: lending_capacity_q1, gross_approved_amount, committed_capacity_amount, remaining_capacity, priority_ranking
• decisions[] (ascending application_id): application_id, decision, approved_amount, bank_capacity_used, conditions[]
• concentration_flags[] (sort by sector then application_id): sector, application_id, limit_pct, post_approval_pct (4 decimals), flag (boolean), handling
• decline_reasons: map application_id → sorted list of reason strings
• post_approval_concentrations[] (sort by sector ascending): sector, exposure_after_approval, post_approval_pct (4 decimals), limit_pct, over_limit (boolean)

Pattern 3: Credit Union Segment Review

Business Rules

• Fetch segment data and NCUA Q1 2025 benchmarks.
• state_metrics: pull the row for the segment’s state_code from NCUA benchmarks.
• peer_comparison:
  • peer_states from segment data.
  • For each metric (delinquency_bps, loan_to_share_pct, roaa_bps, positive_net_income_pct), compare segment state vs US median and vs peer-state median. Label: higher / lower / equal.
• posture: derive from capacity and risk:
  • If current_outstanding is near or above quarterly_capacity → pause_new_originations or continue_with_tighter_conditions
  • If external risk is weaker than peers → continue_with_tighter_conditions
  • Otherwise → continue_normal_underwriting
• controls.required_checklist_gates: from segment minimum_checklist, sorted alphabetically.
• controls.added_operating_controls: add segment-appropriate controls (e.g. lien perfection, monthly delinquency watch, quarterly benchmark monitoring, senior underwriter second review, pre-close insurance verification). Sort alphabetically.
• escalation_triggers: create 3 triggers with IDs ET001–ET003 covering delinquency threshold, missing documentation/lien exceptions, and capacity overrun.
• interpretation: capacity_status (capacity_available / at_capacity), external_risk_status (weaker_than_national_and_peers / stronger_than_peers), risk_tolerance from segment, committee_message concise summary.

Output Fields

• segment_id, posture
• state_metrics
• peer_comparison
• controls
• escalation_triggers[]
• interpretation

Pattern 4: Watch List Review

Business Rules

• Scope: loans with current_rating >= 6 or loans the prompt explicitly flags as adverse (e.g. "watch list pool"). If the prompt says "adverse_rating_min": 6, include all loans with current_rating ≥ 6.
• Compute CDFI factor score for each scoped loan using loan-level data:
  • fico: >720→0, 680-720→1, 580-679→3, <580→5
  • ltv: <0.40→0, 0.40-0.60→2, 0.60-0.80→4, >0.80→6
  • debt_to_asset: <0.40→0, 0.40-0.60→2, 0.60-0.80→4, >0.80→6
  • liquidity_months: >12→0, 6-12→1, 3-6→3, <3→5
  • Sum available factor scores. If a factor is null, skip it.
• Risk class from total score:
  • 0-5 → Prime
  • 6-9 → Desirable
  • 10-13 → Satisfactory
  • 14-18 → Watch
  • ≥19 and ltv > 1.0 → Projected Loss
  • ≥19 otherwise → Doubtful
• Stress test (watch-list formula from policy):
  • stressed_dscr = base_dscr / (1 + 0.18)
  • Breach threshold = 1.0
  • Only test loans that have a dscr value.
• Workout queue: sort scoped loans by severity:
  • Projected Loss first
  • Then by payment_status severity (Nonaccrual > 90+ DPD > 60 DPD > 30 DPD > Current)
  • Then by current_rating descending
  • Then by outstanding_balance descending
• recommended_action per loan:
  • Projected Loss or Nonaccrual → partial_chargeoff_review
  • 90+ DPD or rating 7+ → special_assets
  • Others → watchlist
• monitoring_cadence: monthly for any Watch/Projected Loss present, otherwise quarterly.
• severe_bucket_counts: group scoped loans by (current_rating, payment_status), count and sum exposure.

Output Fields

• branch_id
• watch_list_summary: adverse_rating_min, adverse_loan_count, adverse_balance, risk_classes[], monitoring_cadence
• stress_results: shock_label (+200bp), breach_threshold, results[], breach_loan_ids[]
• workout_queue[]
• severe_bucket_counts[]

Pattern 5: Competing CRE Decision

Business Rules

• Compare exactly the two CRE applications specified in the prompt.
• CRE weighted score (from policy weights):
  • capacity 0.45, capital 0.03, character 0.05, collateral_exposure 0.36, conditions 0.11
  • Compute each factor as a normalized score (typically 1-5 scale based on thresholds), then weighted_cdfi_score = sum(factor_score * weight).
  • Round to 1 decimal.
• Score class:
  • ≤2.0 → approve_quality
  • ≤3.0 → conditional

  • 3.0 → weak

• Decision per application:
  • approve_quality + no concentration breach → approve
  • conditional + no major issues → conditional_approve
  • weak or concentration breach or FDIC variance adverse → decline, defer, or participation_required
• Stress test (CRE dual-stress formula):
  • stressed_dscr = dscr * 0.85 / (1 + 0.18) (i.e. dscr * 0.85 / 1.18)
  • Threshold = 1.0
• Concentration:
  • existing_cre_exposure = sum of outstanding_balance for all branch loans where loan_type == "CRE"
  • existing_cre_concentration = existing_cre_exposure / total_assets (4 decimals)
  • selected_post_approval_cre_concentration = (existing_cre_exposure + selected_approved_amount) / total_assets (4 decimals)
  • selected_policy_variance_bps = (selected_post_approval_cre_concentration - cre_policy_limit_pct) * 10000 (2 decimals)
• FDIC comparison:
  • branch_delinquency_ratio = branch’s delinquency_30_plus_pct from metrics (2025Q1 row) (4 decimals)
  • fdic_benchmark_ratio = relevant FDIC metric (e.g. total_real_estate_30_89_pct) (4 decimals)
  • fdic_variance_ratio = branch_delinquency_ratio - fdic_benchmark_ratio (4 decimals)
  • fdic_variance_bps = fdic_variance_ratio * 10000 (2 decimals)
• Recommended path:
  • Select the application with lower weighted_cdfi_score (better credit).
  • If tied, select the one with higher DSCR or lower LTV.
  • path = decision of selected application.
  • unselected_disposition = decline or defer.
  • unselected_reason_codes = sorted list from the allowed subset (sector_breach, weak_dscr, high_ltv, fdic_adverse_variance).
• Conditions: select all applicable conditions from the allowed enum, sorted alphabetically. Include items like bank_retained_exposure_cap, committee_cre_exception, updated_appraisal_before_close, tenant_roll_and_lease_review, minimum_dscr_covenant_1_25, quarterly_financial_reporting, no_additional_cre_without_committee_review.

Output Fields

• branch_id
• applications_compared[] (ascending application_id): application_id, weighted_cdfi_score (1 decimal), score_class, decision, reason_codes[] (sorted)
• recommended_path: selected_application_id, path, unselected_application_id, unselected_disposition, unselected_reason_codes[]
• stress: formula, coverage_breach_threshold, results[]
• concentration: cre_policy_limit_pct, existing_cre_exposure, existing_cre_concentration, selected_post_approval_cre_concentration, selected_policy_variance_bps, fdic_benchmark_metric, branch_delinquency_ratio, fdic_benchmark_ratio, fdic_variance_ratio, fdic_variance_bps
• conditions[] (sorted alphabetically)

Universal Precision Rules

• Currency / balances: round to 2 decimals.
• Percentages as ratios: round to 4 decimals.
• Percentage points / bps: round to 2 decimals.
• Scores: round to 1 decimal.
• DSCR / stress values: round to 2 decimals.
• Sort all lists alphabetically or numerically as specified in the template.
• JSON output must match the exact keys and nesting of the provided answer_template.json.

Common Pitfalls

1. Using wrong benchmark set: Banks use FDIC Q4 2024; credit unions use NCUA Q1 2025.
2. Null handling: Skip null factors in CDFI scoring; do not treat null as 0.
3. Sector-exposure vs. limit: limit_pct for a sector may differ from branch sector_ceiling_pct if the exposure is grandfathered. Use the value from /api/branches/{branch_id}/sector-exposures.
4. SBA guaranty math: bank_capacity_used is reduced by the guaranty percentage, but approved_amount stays at the full requested (or reduced) amount.
5. CRE stress formula: use * 0.85 / 1.18, not the watch-list /(1+0.18).
6. Total assets denominator: always use the branch’s total_assets from /api/branches/{branch_id}, not loans outstanding or deposits.
7. Do not include narrative text outside the JSON.