Fewshot Attempt 02

Bank Branch Credit-Risk Lending Committee — SKILL

Environment

Base URL: http://34.46.77.124:8011. Never use localhost/127.0.0.1 or run env/setup.sh.

API Workflow

1. Discover: GET /api/manifest → GET /api/health
2. Branch data: GET /api/branches, GET /api/branches/{id}, GET /api/branches/{id}/metrics, GET /api/branches/{id}/loans, GET /api/branches/{id}/sector-exposures
3. Applications: GET /api/applications (filter by branch)
4. Policies & benchmarks: GET /api/policies, GET /api/benchmarks (or benchmark-specific endpoints from manifest)
5. Segment: GET /api/segments/{segment_id} and related NCUA benchmark endpoints

Numeric Precision (Universal)

Sorting Rules (Universal)

• loan_id / application_id: ascending string (lexicographic)
• final_rating / current_rating: ascending integer
• sector, action, reason_codes, conditions: ascending alphabetically
• trigger_id: ascending (e.g. ET001 < ET002)
• state_code / peer_states: ascending
• Workout queue: descending exposure, then ascending loan_id
• Priority ranking: highest priority first (order determined by scoring)

Common Enum Sets

Payment Status

Current, 30 Days Past Due, 60 Days Past Due, 90+ Days Past Due, Nonaccrual

Risk Classes (CDFI-style, lower score = better)

Prime, Desirable, Satisfactory, Watch, Doubtful, Projected Loss

CDFI factor scoring uses objective loan attributes (payment status, DSCR, LTV, FICO/credit, collateral coverage, time in business, etc.). Each factor contributes points; higher total = worse risk. Map total factor_score to risk class using standard CDFI bands.

Recommended Actions

monitor, watchlist, special_assets, workout, partial_chargeoff_review, legal_referral

Escalation by severity: monitor (least severe), watchlist, special_assets, workout, partial_chargeoff_review, legal_referral (most severe).

Decision Types

approve, conditional_approve, decline, defer, participation_required

Decline Reason Codes

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

Application Conditions

participation_required, reduced_amount, board_exception, sba_guaranty_required, startup_monitoring, none

Benchmark Versions

fdic_q4_2024, ncua_q1_2025

Benchmark Metrics

total_loans_noncurrent_pct, total_real_estate_noncurrent_pct, construction_development_noncurrent_pct, total_real_estate_30_89_pct

Monitoring Cadence

monthly, quarterly, semiannual

Posture

continue_approving, continue_with_tighter_conditions, temporarily_pause

Score Classes (CRE)

approve_quality, conditional, weak

Direction (peer comparison)

higher, lower, equal

Interpretation Enums

• 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

Task-Specific Business Rules

1. Rating Migration Review

Population: loans where current_rating >= target_current_rating_min (typically 3).

Re-rating: Derive a new final_rating for each target loan using available financials (payment status, DSCR, LTV, collateral, days past due). A nonaccrual loan maps to rating 8. Severely delinquent (90+) maps to 7. Rating 6 for 30/60 DPD or weak DSCR. Rating 3–5 for performing with adequate coverage.

final_rating_exposure_totals: Group target loans by final_rating. Sum loan_count and exposure. Sort ascending by final_rating.

migration_from_current_rating_3: Subset of target loans originally rated 3 whose final_rating differs. Group by final_rating. Sort ascending. Include loan_ids sorted ascending.

Material downgrades: Loans where final_rating > current_rating AND downgrade_notches >= 2. downgrade_notches = final_rating - current_rating. Sort ascending by loan_id.

Watch-list action coverage: Assign action by final_rating band:

• Rating 6 → watchlist
• Rating 7 → special_assets
• Rating 8 → partial_chargeoff_review

Only include loans that received a new final_rating (the regraded population). Group by action, sort ascending by action string. covered_loan_count/covered_exposure = totals for all regraded loans assigned an action.

NPA benchmark:

• branch_npa_exposure = sum of exposure for loans with payment_status == "Nonaccrual"
• branch_total_loans = total exposure across ALL branch loans
• branch_npa_ratio = branch_npa_exposure / branch_total_loans (4dp)
• fdic_benchmark_ratio = lookup from FDIC benchmark for total_loans_noncurrent_pct
• variance_ratio = branch_npa_ratio − fdic_benchmark_ratio (4dp)
• variance_bps = variance_ratio × 10000 (2dp)

Top problem credit: The loan with the highest final_rating, then highest exposure as tiebreaker. Include borrower_name, exposure, current_rating, final_rating, payment_status, recommended_action.

2. Allocation Package

Capacity: lending_capacity_q1 comes from branch metrics (e.g., tier 1 capital × leverage multiplier, or stated lending limit). Read from branch metrics endpoint.

Priority ranking: Score each application on financial strength (DSCR, LTV, FICO, collateral, borrower history). Rank descending by score. Include only approved and conditional_approve applications in priority_ranking.

Allocation: Process applications in priority order. For each:

• approve: full requested amount, bank_capacity_used = approved_amount
• conditional_approve: approved_amount = requested, bank_capacity_used = bank's retained portion (requested minus guaranteed/participated)
• decline: approved_amount = 0, bank_capacity_used = 0

Stop approving once remaining capacity is exhausted; decline remaining for capacity_limit.

Key formulas:

• gross_approved_amount = sum of approved_amount for all approved + conditional_approve
• committed_capacity_amount = sum of bank_capacity_used for all approved + conditional_approve
• remaining_capacity = lending_capacity_q1 − committed_capacity_amount

Concentration flags: For each application being approved, check if post_approval_pct > limit_pct for its sector. flag: true and handling set accordingly. Sort by sector then application_id.

Decline reasons: Object mapping declined application_id → array of reason codes. Sort codes alphabetically. Only include declined applications.

Post-approval concentrations: For every sector with exposure, compute:

• exposure_after_approval = existing sector exposure + newly approved exposure in that sector
• post_approval_pct = exposure_after_approval / total_branch_exposure_after_all_approvals
• limit_pct from policy
• over_limit = post_approval_pct > limit_pct

Sort by sector ascending.

3. Watch-List Stress

Adverse population: Loans with current_rating >= adverse_rating_min (typically 6).

CDFI risk classes: Score each adverse loan on objective factors (payment status: current=0, 30=3, 60=5, 90+=8, nonaccrual=10; DSCR: ≥1.25=0, 1.0–1.25=3, <1.0=5; LTV: <60%=0, 60–80%=2, >80%=4; plus other available factors). Sum → factor_score. Map to risk class.

DSCR stress (+200bp):

• stressed_dscr = base_dscr × (1 − 0.15) (approximates 200bp rate shock impact on debt service)
• OR use a rate-shock formula: stressed_dscr = (NOI) / (debt_service × 1.15)
• breach_threshold = 1.0
• breaches_threshold = stressed_dscr < 1.0
• Only include loans with available DSCR
• Sort results ascending by loan_id
• breach_loan_ids: loans where breaches_threshold is true, sorted ascending

Workout queue: All adverse loans. Sort descending by exposure, then ascending by loan_id. Assign recommended_action by severity: Projected Loss + Nonaccrual → partial_chargeoff_review; 90+ DPD → special_assets; Watch risk class → special_assets; Desirable but DSCR breach → watchlist; Desirable no breach → watchlist.

Severe bucket counts: Group adverse loans by (current_rating, payment_status). Sum loan_count and exposure. Sort ascending by current_rating, then payment_status.

4. Competing CRE Decision

Weighted CDFI scoring: Score each application on CRE-relevant factors (DSCR, LTV, debt yield, sponsor strength, market, tenant quality). Weight and sum → weighted_cdfi_score (lower = better, 1dp).

Score class:

• ≤ 2.5: approve_quality
• 2.6–3.5: conditional
• ≥ 3.6: weak (approximate — calibrate from observed data)

CRE dual-stress: formula: "dscr * 0.85 / 1.18" — applies 15% rate shock (×0.85 on numerator NOI proxy) and 18% vacancy/expense stress (÷1.18 on debt service). coverage_breach_threshold: 1.0.

Concentration:

• cre_policy_limit_pct from policies
• existing_cre_exposure from sector-exposures (CRE-related sectors)
• existing_cre_concentration = existing_cre_exposure / branch_total_loans (4dp)
• selected_post_approval_cre_concentration = (existing_cre_exposure + selected_app_amount) / (branch_total_loans + selected_app_amount) (4dp)
• selected_policy_variance_bps = (selected_post_approval_cre_concentration − cre_policy_limit_pct) × 10000 (2dp)
• FDIC benchmark: use total_real_estate_30_89_pct for CRE delinquency comparison
• branch_delinquency_ratio from branch metrics
• fdic_variance_ratio = branch_delinquency_ratio − fdic_benchmark_ratio (4dp)
• fdic_variance_bps = fdic_variance_ratio × 10000 (2dp)

Path selection: Choose the application with lower (better) weighted_cdfi_score. If both breach threshold or have severe issues, the stronger may still get participation_required. The unselected gets defer (if issues are fixable) or decline.

Conditions: Derive from risk profile — CRE concentration breach → committee_cre_exception, bank_retained_exposure_cap; DSCR marginal → minimum_dscr_covenant_1_25; etc. Sort alphabetically.

5. Credit-Union Segment Posture

State metrics: Pull NCUA benchmark data for the target state. Values are integers as reported.

Peer comparison: Select neighboring/regional peer states (2–4 states). Compute median across peers. For each metric, compare NC value to US/peer median → higher/lower/equal.

Posture logic:

• Strong metrics + capacity → continue_approving
• Weak metrics but capacity exists → continue_with_tighter_conditions
• Severe deterioration → temporarily_pause

Controls: Checklist gates are always-required verifications. Added operating controls are risk-mitigating extras triggered by the posture decision.

Escalation triggers: Derived from identified risks. Each gets a trigger_id (ET001, ET002, …), a condition string from the enum, and an owner.

Interpretation: Assess capacity (from segment data), external risk (from benchmark comparison), and risk tolerance. Select committee_message that matches the combination.

Common Pitfalls

1. Precision mismatches: Currency → 2dp, ratios → 4dp, bps → 2dp. Never mix.
2. variance_bps formula: Always variance_ratio × 10000 (not ×100).
3. Sort order violations: loan_ids ascending within groups; sectors/reasons alphabetical; workout queue by descending exposure.
4. Enum drift: Use only the exact enum strings listed above. Don't invent variants.
5. Missing required keys: Every object in the template has a fixed key set — include all required keys even if value is 0 or empty array.
6. Empty vs omitted: "conditions": ["none"] for no-condition applications, not []. breaches_threshold: false not omitted.
7. Localhost trap: Never use localhost. Always http://34.46.77.124:8011.
8. Benchmark lookup: The manifest tells you which benchmark endpoints exist. Don't assume endpoint paths — discover them.
9. Total loan base: When computing ratios, use the full branch loan total (all loans, not just the filtered population) unless the task explicitly scopes otherwise.
10. DSCR stress: The formula varies by task. Read the task prompt for the specific stress specification. For watch-list: +200bp rate shock. For CRE: dual stress dscr * 0.85 / 1.18.
11. Factor scoring: Use only objective, available loan attributes. Don't hallucinate factors that aren't in the API response.
12. Concentration math: post_approval denominator includes the newly approved amounts (total portfolio grows), not just the pre-approval total.