Fewshot Attempt 01

Credit Office Lending Committee Skill

Environment

• Base URL: http://34.46.77.124:8011
• Start with GET /api/manifest for endpoint inventory and GET /api/health for record counts.
• Benchmark versions: FDIC fdic_q4_2024, NCUA ncua_q1_2025. Policy version: credit_policy_v2025Q1.

API Endpoints

Risk Rating Derivation (Dominant-Factor Rule)

Final re-derived rating = worst (highest numeric) rating across these independent factors:

DSCR Thresholds

LTV Thresholds

Delinquency Minimums

Re-rating scope: Only re-derive ratings for loans whose current_rating meets the task's stated threshold (e.g., "≥ 3" or "≥ 6"). Loans outside scope keep their current_rating as-is.

CDFI Factor Scoring

Score each loan across these factors, then sum. Factors with null values contribute 0.

Risk Class Assignment

Use Projected Loss when BOTH conditions (score ≥ 19 AND ltv > 1.0) are met; Doubtful when score ≥ 19 but ltv ≤ 1.0 or ltv is null.

Stress Testing

Watch-List (+200bp) Stress

stressed_dscr = dscr / 1.18

Apply only to loans where DSCR is available (non-null). Breach threshold: 1.0 (stressed_dscr < 1.0 → breaches_threshold = true).

CRE Dual Stress

stressed_dscr = dscr * 0.85 / 1.18

Breach threshold: 1.0. Applied to CRE applications being compared.

CRE Weighted Score (Competing CRE Decisions)

Weights from policy:

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

Score classes (lower is better):

For factor computation: use the application's dscr, ltv, fico, sba_guaranty_pct, documentation_complete, existing_relationship_years, prior_delinquencies_12m, co_guarantor_strength, years_in_business, and borrower financials to derive each dimension score (0=best, higher=worse), then compute the weighted sum.

Weighted Score Factor Mapping (inferred from training pattern)

Capacity (weight 0.45): Based on DSCR. ≥1.50→0, ≥1.25→1, ≥1.05→2, ≥1.00→3, <1.00→4. Null→0. **Capital** (weight 0.03): Based on net_income / total_assets or similar leverage metric. Low leverage→0. **Character** (weight 0.05): Based on FICO band. >720→0, 680-720→1, 580-679→3, <580→5. Collateral/Exposure (weight 0.36): Based on LTV. <0.40→0, 0.40-0.60→2, 0.60-0.80→4, >0.80→6. Conditions (weight 0.11): Based on documentation_complete (1→0, 0→1), sba_guaranty_pct (null→1, >0→0), prior_delinquencies (>0→1, 0→0), years_in_business (<3→1, ≥3→0), co_guarantor_strength (none→2, limited→1, standard→0, strong→0).

Weighted score = Σ(weight × factor_score). Round to 1 decimal.

Concentration Calculations

Sector Concentration

concentration_pct = sector_exposure / branch_total_loans_outstanding

Use the most recent quarter's total_loans_outstanding from branch metrics.

Post-Approval Concentration

post_approval_pct = (sector_exposure + approved_amount) / (total_loans + approved_amount)

For CRE-specific: sum only loans where loan_type == "CRE" plus the approved CRE amount, divided by total_loans + approved_amount.

Flag Logic

• over_limit = true when post_approval_pct > limit_pct
• flag = true when post-approval pct exceeds or is very close to the sector ceiling
• Variance bps from policy limit: (concentration_pct - limit_pct) * 10000

NPA Benchmark Analysis

branch_npa_ratio = branch_npa_exposure / branch_total_loans
variance_ratio = branch_npa_ratio - fdic_benchmark_ratio
variance_bps = variance_ratio * 10000

branch_npa_exposure = sum of outstanding_balance for loans with payment_status == "Nonaccrual". Use the most recent quarter's total_loans_outstanding for branch_total_loans.

Metric selection: use total_loans_noncurrent_pct from FDIC Q4 2024 unless the task specifically calls for a real-estate or construction sub-metric. If the task involves CRE, use total_real_estate_noncurrent_pct or total_real_estate_30_89_pct as directed by the template enum.

Watch-List Action Assignment

Map final_rating + payment_status → recommended_action:

General principle: actions escalate with rating severity. partial_chargeoff_review for rating-8 nonaccrual; special_assets for rating-7; watchlist for rating-6; lower ratings with severe delinquency get special_assets or watchlist. Loans that are downgraded but remain in moderate territory get monitor.

Application Decision Rules

Capacity

Branch lending_capacity_q1 is the total pool. Subtract each approved loan's bank_capacity_used in priority order. remaining_capacity = lending_capacity_q1 - committed_capacity_amount.

Bank Capacity Used

• Full approve: bank_capacity_used = approved_amount (= requested_amount)
• Participation required: bank_capacity_used = requested_amount * (1 - participation_pct). Participation pct is derived from sector overage: the portion that exceeds the ceiling.
• SBA guaranty: bank_capacity_used = requested_amount * (1 - sba_guaranty_pct)
• Decline: bank_capacity_used = 0.0

Decision Criteria

• approve: DSCR ≥ 1.25, LTV ≤ 0.80, FICO ≥ 680, no red flags, sector within limit, capacity available
• conditional_approve: Approvable credit but needs mitigation — participation_required when sector is near/at ceiling; sba_guaranty_required + startup_monitoring for startups (<3 years) with SBA; reduced_amount when capacity is tight; board_exception for policy edge cases
• decline: DSCR < 1.0, LTV > 1.0, FICO < 580, recent bankruptcy, documentation gaps, capacity exhausted, sector already over limit with no grandfathering room
• defer: Application needs more information or conditions to be met before decision
• participation_required: Good credit but bank cannot hold full exposure

Decline Reason Codes

Order reason codes alphabetically within each declined application.

Material Downgrades

A downgrade is material when final_rating - current_rating ≥ 2 (i.e., ≥ 2 notches worse). Include all such downgrades in material_downgrades, sorted ascending by loan_id.

Top Problem Credit

Select from the regraded population by: highest final_rating (worst), then highest exposure as tiebreaker. The top problem credit drives the most severe recommended_action.

CU Segment Posture

State Metrics

From NCUA benchmark table: look up the segment's state_code, copy delinquency_bps, loan_to_share_pct, roaa_bps, positive_net_income_pct exactly as integers.

Peer States

From the segment endpoint's peer_states field. Sort ascending.

Direction Comparisons (NC vs US / NC vs Peer Median)

For each metric:

• delinquency_bps: higher is worse
• loan_to_share_pct: higher is worse (more leveraged)
• roaa_bps: lower is worse
• positive_net_income_pct: lower is worse

Compare the state value to US row or peer median; assign "higher", "lower", or "equal".

Peer Median Calculation

For each metric across peer_states rows, take the median of values. If even number of peers, use the lower-middle value (or average — follow the integer convention from benchmarks).

Posture Decision

• continue_approving: capacity available, external risk stronger than or equal to national/peers
• continue_with_tighter_conditions: capacity available but external risk is mixed or weaker
• temporarily_pause: no capacity or external risk significantly weaker

Controls

required_checklist_gates: from segment's minimum_checklist field. added_operating_controls: derive from segment's internal_context and notes. Always include monitoring controls and the controls mentioned in context (insurance binder, lien perfection, senior underwriter review, etc.).

Escalation Triggers

Derive from segment's internal_context and risk profile. Common triggers:

• segment_recent_delinquency_ge_90_bps: when recent_delinquency_bps ≥ 90
• missing_insurance_or_lien_exception: when control_issue mentions insurance/lien gaps
• quarterly_capacity_exceeded_or_exception_requested: general capacity trigger
• state_delinquency_gap_widens_25_bps: when state delinquency exceeds national by 25+ bps

Assign trigger_ids as ET001, ET002, ET003 ascending. Owners: credit_risk_manager (delinquency), operations_control_manager (documentation/controls), lending_committee_chair (capacity).

Sorting Conventions

Numeric Precision Rules

Critical: compute at full precision, round each output field independently at write time. Do not chain rounded values into downstream calculations — use the unrounded intermediates.

Example: variance_bps is computed as (npa_ratio_full - benchmark) * 10000 and then rounded to 2dp — it is NOT derived from the already-rounded variance_ratio field, which would lose precision.

Common Pitfalls

1. Dominant factor is worst/highest, not average. Take the max rating across DSCR, LTV, and delinquency factors, not a blend.
2. Delinquency floors are minimums. If DSCR suggests rating 3 but loan is 60 DPD, the floor is 5 — the final rating is max(3, 5, ltv_rating) = 5.
3. NPA = Nonaccrual only. Don't include 90+ DPD in NPA exposure unless the loan is specifically marked Nonaccrual.
4. Benchmark version must match. Use fdic_q4_2024 for FDIC, ncua_q1_2025 for NCUA — do not mix versions.
5. Workout queue sorts descending by exposure. Almost everything else is ascending. This is the standout exception.
6. Watch-list stress only on loans with DSCR. Skip loans where dscr is null in stress_results.
7. Projected Loss requires BOTH score ≥ 19 AND ltv > 1.0. If ltv is null or ≤ 1.0, it's Doubtful at score ≥ 19.
8. Participation bank_capacity_used ≠ approved_amount. For participation, the bank only commits its retained share. approved_amount is the full loan amount; bank_capacity_used is the bank's retained portion.
9. Variance bps = variance_ratio × 10000, not × 100.
10. CRE concentration post-approval uses only CRE-typed exposure plus the new CRE loan, over total_loans + new amount.
11. Flags in concentration_flags only fire when limit is breached or near-breached. Not every application gets a flag — only those where post_approval_pct approaches or exceeds limit_pct.
12. Declined loans have approved_amount = 0.0 and bank_capacity_used = 0.0, with conditions: ["none"].
13. Priority ranking includes approved AND conditionally approved only, in descending priority order (best first).
14. Top problem credit picks the highest (worst) final_rating, breaking ties by highest exposure.