Self Attempt 02

Credit Risk Committee JSON Skill

API Workflow

1. Read environment: Get GDPEVO_ENV_BASE_URL from environment_access.md.
2. Discover data:
   • GET /api/manifest — lists all endpoints and record counts.
   • GET /api/branches — resolve the target branch; prompts may use aliases (e.g., "Eastfield" → EASTGATE), so match by application prefix or branch state.
   • GET /api/branches/{branch_id} — branch profile (total_assets, lending_capacity_q1, cre_policy_limit_pct, sector_ceiling_pct, institution_type, state_code).
   • GET /api/branches/{branch_id}/metrics — latest quarter metrics (use 2025Q1); note delinquency_30_plus_pct and total_loans_outstanding.
   • GET /api/branches/{branch_id}/loans — outstanding balances, loan types, sectors, payment status, days past due, DSCR, LTV, FICO, debt-to-asset, liquidity months.
   • GET /api/branches/{branch_id}/sector-exposures — current exposure per sector with limit.
   • GET /api/branches/{branch_id}/applications — filter for the target application IDs.
   • GET /api/policies — scoring rules, thresholds, weights.
   • GET /api/benchmarks/fdic/q4-2024 — benchmark ratios.
   • GET /api/benchmarks/ncua/q1-2025 — for credit unions, match by state_code.
3. Compute scores using the policy definitions below.
4. Emit JSON exactly matching the task answer_template.json; no narrative outside JSON.

Scoring & Business Rules

1. CDFI Factor Scoring (train_001 style)

From /api/policies → cdfi_factor_scores, sum applicable factor scores (0 = best):

• weighted_cdfi_score = sum of all available factor scores (precision 1). Missing fields contribute 0.
• score_class mapping:
  • prime: 0–5
  • desirable: 6–9
  • satisfactory: 10–13
  • watch: 14–18
  • doubtful: ≥19 (use projected_loss only if ≥19 and ltv > 1.0)
• Decision logic: Lower score is better. Rank ascending. The best may still be decline/defer if capacity or sector limits are breached.

2. CRE Weighted Score (train_005 style)

From /api/policies → cre_weighted_score:

• Each component is rated on a scale where lower is better (e.g., 1 = best, 5 = worst).
• score_class:
  • approve_quality: weighted score ≤ 2.0
  • conditional: weighted score ≤ 3.0
  • weak: weighted score > 3.0
• Capacity: use DSCR thresholds from risk rating (1.5+ → best, 1.25+ → next, etc.).
• Collateral_exposure: use LTV thresholds (≤0.65 best, ≤0.75 next, etc.).
• Capital: judge from total_assets vs total_debt (equity position) or debt_to_asset.
• Character: relationship length, prior delinquencies, bankruptcy, guarantor strength.
• Conditions: documentation completeness, market/sector notes, SBA guaranty.

3. Risk Rating (dominant factor rule)

From /api/policies → risk_rating:

• DSCR rating, LTV rating, delinquency rating — final rating is the worst available.
• Delinquency minimums: Current → none; 30 DPD → 4; 60 DPD → 5; 90+ DPD → 7; Nonaccrual → 8.

4. Stress Testing (CRE)

From /api/policies → stress:

• CRE dual-stress formula: stressed_dscr = dscr * 0.85 / (1 + 0.18)
• Coverage breach threshold: 1.00
• breaches_threshold: true if stressed_dscr < 1.00

5. Capacity & Concentration

• Lending capacity: branches.lending_capacity_q1. Requested amount must not exceed residual capacity.
• Sector ceiling: branches.sector_ceiling_pct (default) or sector_exposures.limit_pct (per-sector override). Compute post-approval exposure vs limit.
• CRE concentration: existing_cre_exposure = sum of outstanding_balance for all loans where loan_type == "CRE".
  • existing_cre_concentration = existing_cre_exposure / total_assets
  • selected_post_approval_cre_concentration = (existing_cre_exposure + selected_requested_amount) / total_assets
  • selected_policy_variance_bps = (selected_post_approval_cre_concentration − cre_policy_limit_pct) × 10,000
• FDIC benchmark (bank branches):
  • Metric: total_real_estate_30_89_pct = 0.0051
  • branch_delinquency_ratio = latest delinquency_30_plus_pct
  • fdic_variance_ratio = branch_delinquency_ratio − fdic_benchmark_ratio
  • fdic_variance_bps = fdic_variance_ratio × 10,000
• NCUA benchmark (credit unions):
  • Match branch state_code to NCUA state row; use delinquency_bps.

Output Conventions (Pitfall Prevention)

1. Precision: Respect the decimal precision in the answer template (e.g., precision: 4 → four decimal places, precision: 2 → two, precision: 1 → one). Do not truncate incorrectly; round half-up.
2. Alphabetical ordering:
   • reason_codes arrays must be sorted ascending alphabetically.
   • conditions arrays must be sorted ascending alphabetically.
3. Application ordering:
   • applications_compared and nested results lists must be ascending by application_id.
4. Required keys: Do not omit any key marked required or required_keys in the template.
5. Enums: Use only the exact allowed values from the template (e.g., score_class values, decision values).
6. JSON only: Do not wrap output in markdown code fences or add narrative text.

Branch Name Resolution

Prompts may use fictional aliases. Resolve by:

• Searching /api/branches for the branch whose applications contain the IDs referenced in the prompt (e.g., HAR-APP-901 → HARBOR).
• If the prompt’s branch_id does not match any API branch_id, fall back to matching by application_id prefix or by state_code when the prompt mentions a location.

Common Pitfalls

• Using 2024Q4 instead of 2025Q1 metrics — always pick the latest quarter.
• Confusing sector_ceiling_pct with cre_policy_limit_pct — CRE sectors (Multifamily, Hospitality, Retail CRE, Office, Industrial CRE, Construction) often use the CRE limit; non-CRE sectors use the general sector ceiling.
• Forgetting to include the requested amount in post-approval concentration — add the selected application’s requested_amount to existing_cre_exposure before dividing by total_assets.
• Wrong stress formula — use the exact dscr * 0.85 / (1 + 0.18) from the policy; do not approximate.
• Ignoring grandfathering — the policy states existing over-ceiling exposure may be grandfathered, but new approvals may not worsen that sector without mitigation. If post-approval exposure increases the breach, flag it.
• Omitting documentation_gap or policy_floor_missing — if documentation_complete == 0 or if a required policy floor (e.g., minimum DSCR covenant) is absent, add the corresponding reason code.