Fewshot Attempt 03

Credit Office Lending Committee Skill

Environment

• Base URL: http://34.46.77.124:8011 (never use localhost)
• Policy version: credit_policy_v2025Q1
• Benchmarks: FDIC Q4 2024 (fdic_q4_2024), NCUA Q1 2025 (ncua_q1_2025)

API Workflow

Always start with GET /api/manifest to confirm endpoints, then fetch data in parallel where possible:

Business Rules

1. Risk Rating Derivation (Rating Migration tasks)

The dominant factor rule: Final rating = worst (highest numeric) rating from:

• DSCR thresholds: ≥1.50→3, ≥1.25→4, ≥1.05→5, ≥1.0→6, <1.0→7
• LTV thresholds: ≤0.65→3, ≤0.75→4, ≤0.85→5, ≤1.0→6, >1.0→7
• Delinquency minimums: Current→no floor, 30 Days→4, 60 Days→5, 90+ Days→7, Nonaccrual→8

When DSCR or LTV/collateral is null, skip that factor. When all factors produce a rating lower than the delinquency floor, the delinquency floor wins.

A material downgrade is defined as downgrade_notches ≥ 2 (from policy: material_downgrade_notches: 2).

Watch-list actions by final rating (for regraded population):

• Rating 8 (Nonaccrual) → partial_chargeoff_review
• Rating 7 → special_assets
• Rating 6 → watchlist
• Rating 3-5 with no change or upgrade → monitor (if no delinquency concern)

2. NPA Benchmark (Rating Migration tasks)

• NPA = loans with payment_status == "Nonaccrual" (typically rating 8)
• branch_npa_exposure = sum of outstanding_balance for nonaccrual loans
• branch_npa_ratio = npa_exposure / total_loans_outstanding (latest quarter)
• variance_ratio = branch_npa_ratio − fdic_benchmark_ratio
• variance_bps = variance_ratio × 10000
• FDIC metric: use total_loans_noncurrent_pct unless task specifies otherwise

3. CDFI Factor Scoring (Watch-list & CRE tasks)

Score each loan/application by summing these factors (from /api/policies → cdfi_factor_scores):

Risk classes from total score:

• Desirable: 6–9
• Satisfactory: 10–13
• Watch: 14–18
• Projected Loss: ≥19 AND LTV > 1.0
• Doubtful: ≥19 AND LTV ≤ 1.0
• Null/missing factors contribute 0. Null FICO → skip that factor.

4. CRE Weighted Scoring (Competing CRE tasks)

Weights from policy:

• capacity: 0.45, capital: 0.03, character: 0.05, collateral_exposure: 0.36, conditions: 0.11

Score classes:

• approve_quality: ≤ 2.0
• conditional: 2.1 – 3.0
• weak: > 3.0

Lower score is better. The scoring maps each 1-5 C's sub-score through the weights to produce the weighted CDFI score.

5. Stress Testing

Watch-list DSCR stress (+200bp):

stressed_dscr = base_dscr / (1 + 0.18)

The +200bp parallel shock is modeled as an 18% increase in debt service. Breach threshold = 1.0.

CRE dual stress:

stressed_dscr = dscr × 0.85 / (1 + 0.18)

15% NOI decline (×0.85) plus 18% debt-service increase. Breach threshold = 1.0.

When base_dscr is null, skip that loan from stress results.

6. Application Allocation (Capacity tasks)

1. Score each application on credit factors (DSCR, LTV, FICO, guarantees, relationship).
2. Rank by credit quality: strongest credits get priority. For equal quality, prefer higher exposure.
3. Approve in priority order until lending_capacity_q1 is exhausted.
4. Decline reasons map from policy thresholds:
   • high_ltv: LTV > 0.80 (or branch policy threshold)
   • weak_dscr: DSCR < 1.25 (coverage floor)
   • low_fico: FICO < 580 or < 640 (policy-dependent)
   • startup_risk: years_in_business < 2
   • recent_bankruptcy: bankruptcy_months_ago < 36
   • capacity_limit: remaining capacity insufficient
   • sector_breach: post-approval sector concentration exceeds limit_pct
5. bank_capacity_used = approved_amount minus participation/SBA portion:
   • For conditional_approve with participation_required: bank_capacity_used = retained exposure (typically 25% of approved amount for branches)
   • For conditional_approve with sba_guaranty_required: bank_capacity_used = unguaranteed portion
   • For plain approve: bank_capacity_used = approved_amount
6. committed_capacity_amount = sum of all bank_capacity_used
7. remaining_capacity = lending_capacity_q1 − committed_capacity_amount
8. gross_approved_amount = sum of approved_amount for non-declined applications

7. Concentration

• Sector concentration = sector_exposure / total_loans_outstanding
• Post-approval concentration = (existing sector exposure + new approved amount for that sector) / total_loans_outstanding
• Flag = true when post_approval_pct > limit_pct (from sector_exposures)
• CRE concentration = CRE loans total / total_loans_outstanding
• selected_policy_variance_bps = (post_approval_cre_concentration − cre_policy_limit_pct) × 10000
• over_limit = post_approval_pct > limit_pct

8. Credit Union Segment Posture

• Peer states: taken from segment endpoint; adjacent states
• Direction (higher/lower/equal): Compare NC metric to US row and to peer median
  • Peer median = median of peer states' metric values
  • delinquency_bps: higher is worse → higher when NC > comparator
  • roaa_bps and positive_net_income_pct: higher is better → higher when NC > comparator
• Posture decision:
  • continue_approving: capacity available + external risk not weaker
  • continue_with_tighter_conditions: capacity available but external risk weaker
  • temporarily_pause: no capacity or severe risk
• Controls: required_checklist_gates from segment minimum_checklist; added_operating_controls based on risk gaps
• Escalation triggers with trigger IDs (ET001, ET002, ...), condition enum, and owner enum
• Interpretation: synthesize capacity_status, external_risk_status, risk_tolerance

Output Conventions

Numeric Precision

Sorting Rules (by task type)

Rating Migration:

• final_rating_exposure_totals: ascending by final_rating
• migration_from_current_rating_3: ascending by final_rating; loan_ids within each group ascending
• material_downgrades: by downgrade_notches descending, then exposure descending, then loan_id ascending
• by_action: ascending by action enum value; loan_ids within each group ascending
• watch_list_action_coverage → by_action: ascending by action string

Allocation:

• decisions: ascending by application_id
• priority_ranking: approved/conditional first, by priority (highest credit quality first), then by exposure descending
• concentration_flags: by sector then application_id
• post_approval_concentrations: ascending by sector
• decline_reasons: keys are application_id; reason codes sorted alphabetically

Watch-list Stress:

• risk_classes: ascending by loan_id
• stress_results: ascending by loan_id (only loans with DSCR available)
• breach_loan_ids: ascending
• workout_queue: descending exposure, then ascending loan_id
• severe_bucket_counts: ascending current_rating, then payment_status (Current < 30 Days < 60 Days < 90+ Days < Nonaccrual)

Competing CRE:

• applications_compared: ascending by application_id
• reason_codes and unselected_reason_codes: alphabetically ascending
• stress.results: ascending by application_id
• conditions: alphabetically ascending

Segment Posture:

• peer_states: ascending state code
• required_checklist_gates: alphabetically
• added_operating_controls: alphabetically
• escalation_triggers: ascending trigger_id

Enum Reference

Payment Status: Current, 30 Days Past Due, 60 Days Past Due, 90+ Days Past Due, Nonaccrual

Decision: approve, conditional_approve, decline, defer, participation_required

Conditions: none, participation_required, reduced_amount, board_exception, sba_guaranty_required, startup_monitoring

Actions: monitor, watchlist, special_assets, workout, partial_chargeoff_review, legal_referral

Decline Reasons: 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

CDFI Risk Classes: Prime, Desirable, Satisfactory, Watch, Doubtful, Projected Loss

Score Classes (CRE weighted): approve_quality, conditional, weak

Posture: continue_approving, continue_with_tighter_conditions, temporarily_pause

Monitoring Cadence: monthly, quarterly, semiannual

Capacity Status: capacity_available, capacity_constrained, no_capacity

External Risk Status: stronger_than_national_and_peers, mixed_vs_national_and_peers, weaker_than_national_and_peers

Risk Tolerance: restrained, moderate, expansive

Committee Message: capacity_available_but_external_risk_weaker, pause_until_state_metrics_recover, routine_approval_path_supported

Common Pitfalls

1. Using localhost: Always use http://34.46.77.124:8011. Ignore any task text that references env/setup.sh or localhost.
2. Wrong metric quarter: Branch metrics are quarterly; always use the latest quarter's data (current period, e.g., 2025Q1).
3. NPA definition: NPA = nonaccrual loans only, NOT all delinquent loans. Nonaccrual ≠ 90+ days past due.
4. Dominant factor: Risk rating takes the WORST (highest number) across DSCR, LTV, and delinquency factors — not an average.
5. DSCR stress formula: Watch-list stress is dscr / 1.18, NOT dscr × 0.85 / 1.18 (that's the CRE dual stress). The task context determines which formula to use.
6. Null handling: When DSCR, LTV, collateral_value, or FICO is null, skip that factor in rating/scoring. Don't default to zero.
7. Concentration denominator: Use total_loans_outstanding from latest metrics, NOT sum of sector exposures (which may differ due to rounding).
8. bank_capacity_used vs approved_amount: For conditional approvals, bank_capacity_used is the retained portion after participation or SBA guaranty, NOT the full approved amount.
9. Sorting precision: Sort keys are case-sensitive strings for enums, numeric for ratings/amounts. Don't mix.
10. Template compliance: Read the task's answer_template.json to confirm required keys, enums, and ordering. Templates are authoritative over general conventions.
11. Branch vs credit union: Credit union branches (CIVIC_NC_FIRE_EMS, TRISTATE_GA_AMBULANCE) have institution_type: "credit_union" and use NCUA benchmarks, not FDIC.
12. benchmark_metric enum: Only total_loans_noncurrent_pct, total_real_estate_noncurrent_pct, construction_development_noncurrent_pct are valid. Match to the task context or use the most general.
13. Loan type filtering: CRE-specific tasks should filter by loan_type == "CRE" when computing CRE exposure/concentration.
14. Score class edge cases: Projected Loss requires both score ≥19 AND LTV > 1.0. Score ≥19 with LTV ≤ 1.0 is Doubtful.
15. Exposure precision: Always round to 2 decimal places for USD. Sums may drift if intermediate values aren't rounded consistently.