Reflect 3 Attempt 01

Bank Branch Credit-Risk Lending-Committee Skill

API Workflow

1. Start with GET /api/manifest to discover endpoints, benchmark versions, and record counts.
2. GET /api/policies for business rules: risk-rating thresholds, CDFI factor scores, CRE scoring weights, stress formulas, concentration policies.
3. Fetch target-branch data: GET /api/branches/{id} (metadata, limits), GET /api/branches/{id}/metrics (financials — use most recent quarter), GET /api/branches/{id}/loans (portfolio), GET /api/branches/{id}/sector-exposures (concentrations), GET /api/branches/{id}/applications (pending apps).
4. Benchmarks: GET /api/benchmarks/fdic/q4-2024 and GET /api/benchmarks/ncua/q1-2025.
5. Credit-union segments: GET /api/credit-union-segments/{id}.
6. Use the base URL from environment_access.md; never hardcode localhost.

Risk-Rating Rules

Dominant-factor rule: Final re-derived rating = worst (maximum numeric) rating from available DSCR, LTV/collateral, and delinquency factors. Skip null factors.

DSCR thresholds: ≥1.50→3, ≥1.25→4, ≥1.05→5, ≥1.00→6, <1.00→7

LTV thresholds: ≤0.65→3, ≤0.75→4, ≤0.85→5, ≤1.00→6, >1.00→7

Delinquency minimums (floor): 30 Days Past Due→4, 60 Days Past Due→5, 90+ Days Past Due→7, Nonaccrual→8, Current→no floor

Material downgrade: ≥2 notches worse (final_rating − current_rating ≥ 2).

CDFI Factor Scoring (for Risk Classes)

• Null factors contribute 0 to the total.
• Classes: 0–5→Prime, 6–9→Desirable, 10–13→Satisfactory, 14–18→Watch, ≥19→Doubtful, ≥19 and ltv>1.0→Projected Loss.

CRE Weighted Score

Weights: capacity=0.45, capital=0.03, character=0.05, collateral_exposure=0.36, conditions=0.11. Lower score is better.

Score each dimension 1–5 (best to worst), then compute weighted sum.

Classes: ≤2.0→approve_quality, ≤3.0→conditional, >3.0→weak.

Stress-Test Conventions

Watch-list (+200bp): stressed_dscr = dscr / 1.18, breach threshold = 1.00.

CRE dual-stress: stressed_dscr = dscr × 0.85 / 1.18, breach threshold = 1.00.

Shock label: "+200bp" (watch-list) or the compact formula string (CRE).

Concentration Rules

• Sector ceiling: branches.sector_ceiling_pct (default) or per-sector limit_pct from sector-exposures.
• CRE limit: branches.cre_policy_limit_pct.
• Denominator: total_loans_outstanding from the most recent quarterly metrics (not including new approvals).
• Post-approval pct: (existing sector/CRE exposure + approved amounts) / total_loans_outstanding.
• Over-limit mitigation: board_exception, participation_required, or reduced_amount. Existing over-ceiling may be grandfathered; new approvals may not worsen that sector without mitigation.
• NPA ratio: nonperforming_loans / total_loans_outstanding.
• Variance ratio: branch_ratio − benchmark_ratio.
• Variance bps: variance_ratio × 10,000.

Numeric Precision

Sorting Conventions

• final_rating_exposure_totals: ascending by final_rating.
• migration lists: ascending by destination final_rating.
• loan_ids within any list: ascending.
• material_downgrades: ascending by loan_id.
• by_action / conditions: ascending alphabetically by action/condition name.
• decisions: ascending by application_id.
• concentration_flags: ascending by sector, then application_id.
• post_approval_concentrations: ascending by sector.
• decline_reason_codes per app: ascending alphabetically.
• stress results: ascending by loan_id / application_id.
• workout_queue: descending by exposure, then ascending by loan_id.
• severe_bucket_counts: ascending by current_rating, then payment_status (Current < 30DPD < 60DPD < 90+DPD < Nonaccrual).
• peer_states: ascending state_code.
• escalation_triggers: ascending trigger_id.
• priority_ranking: highest quality first (strongest credit).

Key Enum Values

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

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

reason_codes (decline): 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

unselected_reason_codes: sector_breach, weak_dscr, high_ltv, fdic_adverse_variance

posture: continue_approving, continue_with_tighter_conditions, temporarily_pause

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

monitoring_cadence: monthly, quarterly, semiannual

NPA benchmark_metric: total_loans_noncurrent_pct (general), total_real_estate_noncurrent_pct, construction_development_noncurrent_pct

CRE benchmark_metric: total_real_estate_30_89_pct

Common Pitfalls

1. Delinquency is a floor, not standalone: for 30DPD, rating must be ≥4; other factors can push it worse.
2. Null factors are unavailable, not zero-rated: skip them from the worst-of computation. When ALL rating factors (DSCR, LTV, delinquency) are null/unavailable, keep current_rating unchanged.
3. Use the most recent metrics quarter (typically Q1 2025) for total_loans_outstanding, nonperforming_loans, and delinquency rates.
4. SBA guaranty does not reduce bank_capacity_used: committed capacity = full approved_amount, not the bank's retained portion.
5. Match benchmark metric to context: NPA loans → total_loans_noncurrent_pct; CRE delinquency → total_real_estate_30_89_pct.
6. Material downgrade is ≥2 notches, not ≥1.
7. Denominator for concentrations is current total_loans_outstanding, not post-approval total.
8. Sector limits: branch sector_ceiling_pct is the default; individual sectors may have different limit_pct values in sector-exposures (e.g., Healthcare at 0.19 vs default 0.21).
9. Conditions field is an array of strings even when single-valued (e.g., ["none"] or ["board_exception"]).
10. Priority ranking includes only approved and conditional_approve applications, ordered by credit quality (strongest first).
11. Weighted CDFI score is lower-is-better (1=best, 5=worst per dimension).
12. FDIC benchmark values vary by metric — use the field matching the analysis context, not an arbitrary one.
13. NCUA peer median is the median of the specified peer_states, not the average.
14. Outstanding balance is exposure for loans; requested_amount is exposure for applications.
15. Credit union segment data comes from /api/credit-union-segments/{id}, not the branches endpoint.