Self Attempt 01

Skill: CRE Competing-Application Committee Decision

1. Task Overview

Receive a prompt naming a branch_id and two CRE application IDs to compare. Fetch all relevant data from the shared public credit-office API, compute stressed coverage, concentration, FDIC/NCUA variance, and weighted CRE credit scores, then emit a single JSON object matching the task's answer_template.json.

2. API Workflow

Use the remote base URL from environment_access.md (do not run env/setup.sh or use localhost).

Required endpoints (call in this order):

1. GET /api/manifest – verify environment version and endpoint list.
2. GET /api/branches – map the prompt branch name to the actual branch_id.
3. GET /api/branches/{branch_id} – branch profile (cre_policy_limit_pct, sector_ceiling_pct, lending_capacity_q1, institution_type, fdic_benchmark_set).
4. GET /api/branches/{branch_id}/metrics – use the latest quarter (usually 2025Q1).
5. GET /api/branches/{branch_id}/loans – full loan tape.
6. GET /api/branches/{branch_id}/sector-exposures – per-sector limits and grandfathering flags.
7. GET /api/branches/{branch_id}/applications – filter to the two target application IDs.
8. GET /api/policies – scoring weights, stress formula, rating thresholds, concentration rules.
9. GET /api/benchmarks/fdic/q4-2024 – for bank branches (institution_type == "bank").
10. GET /api/benchmarks/ncua/q1-2025 – for credit-union branches (institution_type == "credit_union").

3. Key Computations

3.1 CRE Dual-Stress DSCR

Formula (from policies):

stressed_dscr = dscr * 0.85 / (1 + 0.18)

• coverage_breach_threshold = 1.0
• breaches_threshold = true when stressed_dscr < 1.0
• Round base_dscr and stressed_dscr to 2 decimal places.

3.2 Existing CRE Exposure

existing_cre_exposure = sum(outstanding_balance for loan in loans if loan['loan_type'] == 'CRE')

• Do not include Residential Mortgage, C&I, SBA, Equipment, or Consumer loans.
• Round to 2 decimal places.

3.3 CRE Concentration

existing_cre_concentration = existing_cre_exposure / total_loans_outstanding
selected_post_approval_cre_concentration = (existing_cre_exposure + requested_amount) / (total_loans_outstanding + requested_amount)
selected_policy_variance_bps = (selected_post_approval_cre_concentration - cre_policy_limit_pct) * 10000

• Round concentrations to 4 decimal places.
• Round selected_policy_variance_bps to 2 decimal places.
• If the selected app is not CRE, still add its requested amount to the denominator (it becomes part of the total portfolio) but do not add it to the CRE numerator.

3.4 Sector Exposure Check

For the selected application's sector, look up sector-exposures:

post_sector_concentration = (current_exposure + requested_amount) / (total_loans_outstanding + requested_amount)
sector_limit = limit_pct from sector-exposures (or sector_ceiling_pct if sector not listed)

• If post_sector_concentration > sector_limit, the app triggers sector_breach.
• Respect grandfathered flags: grandfathered sectors may already exceed the default ceiling, but new money still may not worsen the breach without mitigation.

3.5 FDIC / NCUA Benchmark Variance (for bank branches)

branch_delinquency_ratio = sum(balance for loan in loans
    if loan['loan_type'] in ('CRE', 'Residential Mortgage')
    and 30 <= loan['days_past_due'] <= 89) / sum(balance for loan in loans
    if loan['loan_type'] in ('CRE', 'Residential Mortgage'))

• Do not use metrics['delinquency_30_plus_pct'] directly; that metric is portfolio-wide, not real-estate-specific.
• fdic_benchmark_metric = "total_real_estate_30_89_pct"
• fdic_benchmark_ratio = value from FDIC Q4-2024 endpoint.
• fdic_variance_ratio = branch_delinquency_ratio - fdic_benchmark_ratio
• fdic_variance_bps = fdic_variance_ratio * 10000
• Round ratios to 4 decimal places, bps to 2 decimal places.

3.6 NCUA Variance (for credit-union branches)

• Use /api/benchmarks/ncua/q1-2025.
• Credit unions lack FDIC data (fdic_benchmark_set == "").
• The template may still ask for FDIC fields; if the branch is a credit union, compute the same real-estate delinquency ratio and compare against the relevant NCUA peer metric, or note the institution type mismatch.

3.7 Weighted CRE Credit Score (weighted_cdfi_score)

The policy defines:

{
  "capacity": 0.45,
  "capital": 0.03,
  "character": 0.05,
  "collateral_exposure": 0.36,
  "conditions": 0.11
}

Map application data to component scores (0 = best, higher = worse) using the cdfi_factor_scores tables in /api/policies as guidance:

• Capacity → DSCR (higher is better; score inversely).
• Capital → total_debt / total_assets (debt_to_asset table: 0/2/4/6).
• Collateral → LTV (ltv table: 0/2/4/6).
• Character → prior_delinquencies_12m, fico, bankruptcy_months_ago (fico table: 0/1/3/5).
• Conditions → documentation_complete, years_in_business, sector risk, guarantor strength.

Multiply each component score by its weight and sum. Round the final weighted_cdfi_score to 1 decimal place.

Score class thresholds:

• approve_quality → score ≤ 2.0
• conditional → score ≤ 3.0
• weak → score > 3.0

4. Decision Logic

1. Select the stronger credit: compare weighted_cdfi_score (lower is better), then stressed DSCR (higher is better), then LTV (lower is better).
2. Selected application path (from recommended_path.path):
   • If concentration is within policy → approve or conditional_approve based on score class.
   • If CRE or sector limit is breached → participation_required, conditional_approve, or defer.
   • The policy explicitly lists participation_required, reduced_amount, and board_exception as allowed mitigations.
3. Unselected application (recommended_path.unselected_disposition):
   • Allowed: decline or defer.
   • unselected_reason_codes must be from the restricted subset: [sector_breach, weak_dscr, high_ltv, fdic_adverse_variance], sorted alphabetically.
4. Per-application decisions (applications_compared[].decision):
   • The selected app's decision should match recommended_path.path.
   • The unselected app's decision should match unselected_disposition.
   • reason_codes for each app can draw from the full enum; sort alphabetically.

5. Output Conventions

• Ordering:
  • applications_compared: ascending by application_id.
  • stress.results: ascending by application_id.
  • reason_codes and conditions: ascending alphabetically.
• Precision:
  • USD amounts: 2 decimals.
  • Concentration percentages: 4 decimals.
  • BPS values: 2 decimals.
  • DSCR values: 2 decimals.
  • Weighted score: 1 decimal.
• Required top-level keys: branch_id, applications_compared, recommended_path, stress, concentration, conditions.
• No narrative text outside the JSON.

6. Common Pitfalls

• Missing applications: The prompt may reference application IDs that do not exist in the API (e.g., CEN-APP-101 when only CEN-APP-001–005 exist). Always verify by scanning /api/branches/{branch_id}/applications.
• Branch name aliases: Prompts may use shorthand names (NORTHERN, SOUTHERN, WESTERN) that map to different API branch IDs (NORTHSTAR, SOUTHPORT, etc.). Start with /api/branches to resolve the correct ID.
• Using wrong delinquency metric: metrics['delinquency_30_plus_pct'] is portfolio-wide. For FDIC/NCUA comparison, compute the real-estate-only 30–89 DPD ratio from loan-level data.
• CRE vs. total concentration: Only loans with loan_type == 'CRE' count toward CRE exposure. Do not include Residential Mortgage in the CRE numerator.
• Post-approval denominator: Always add the requested amount to total_loans_outstanding when calculating post-approval concentrations.
• Credit-union branches: They have fdic_benchmark_set == "" and use NCUA benchmarks instead. Do not blindly apply FDIC Q4-2024 data.
• Alphabetical sorting: reason_codes and conditions must be sorted ascending alphabetically. The evaluator checks ordering.
• Stressed DSCR rounding: Round to 2 decimals before comparing to the 1.0 threshold (e.g., 0.949 → 0.95, which is still < 1.0).