Reflect 3 Attempt 03

Credit Office Committee JSON Packet Skill

API Workflow

Fetch data in this order for every branch/segment task:

1. GET /api/branches/{branch_id} — branch config (capacity, limits, state, assets)
2. GET /api/branches/{branch_id}/metrics — latest quarter metrics (nonperforming_loans, total_loans_outstanding, delinquency_30_plus_pct)
3. GET /api/branches/{branch_id}/loans — full loan portfolio
4. GET /api/branches/{branch_id}/sector-exposures — per-sector exposure and limit_pct
5. GET /api/branches/{branch_id}/applications — pending applications (for allocation/CRE tasks)
6. GET /api/policies — credit policy thresholds and formulas
7. GET /api/benchmarks/fdic/q4-2024 or /api/benchmarks/ncua/q1-2025 — benchmark data
8. GET /api/credit-union-segments/{segment_id} — segment config, peer_states, minimum_checklist, internal_context

Use the latest metric row (max quarter string like "2025Q1").

Risk Rating Re-derivation (Universal Rule)

From /api/policies:

• DSCR: ≥1.5→3, ≥1.25→4, ≥1.05→5, ≥1.0→6, <1.0→7
• LTV: ≤0.65→3, ≤0.75→4, ≤0.85→5, ≤1.0→6, >1.0→7
• Delinquency: Current→null, 30 DPD→4, 60 DPD→5, 90+ DPD→7, Nonaccrual→8
• Final rating = max of all available factor ratings. If no DSCR/LTV/delinquency factors are available, fallback to current_rating.
• Material downgrade = final_rating − current_rating ≥ 2 (policy value material_downgrade_notches).

Output-Field Conventions

Precision

• Currency (USD): round to exactly 2 decimals (round(value, 2))
• Ratios: 4 decimals
• Basis points (bps): 2 decimals
• Integer fields: exact integers (no .0)

Ordering

• loan_id lists: ascending alphanumeric
• application_id lists: ascending alphanumeric
• sector lists: ascending alphabetically
• state lists: ascending alphabetically
• exposure-descending queues: (-exposure, loan_id)

Enums

Use only the exact strings listed in answer_template.json. Never invent values.

Task-Specific Business Rules

1. Branch Rating Migration Review (e.g., REDWOOD)

• Regrade population: loans with current_rating >= 3 ("3 or worse")
• NPA benchmark: use nonperforming_loans and total_loans_outstanding from branch metrics, do NOT sum non-current loans from the loan list
• watch_list_action_coverage.by_action must include all actions (including monitor). covered_loan_count and covered_exposure should reflect loans with non-monitor actions only
• Action mapping by final_rating: 3→monitor, 4→watchlist, 5→special_assets, 6→workout, 7→partial_chargeoff_review, 8→legal_referral

2. Lending Allocation Package (e.g., LAKEVIEW)

• Sector concentration denominator is total_loans_outstanding from branch metrics, NOT total_assets
• bank_capacity_used for SBA loans should be reduced by the guaranty percentage (approved_amount * (1 - sba_guaranty_pct))
• concentration_flags is per-application: post_approval_pct = (existing_sector[sector] + this_app_approved_amount) / total_loans_outstanding
• Only applications whose own approved amount pushes their sector over limit_pct get flag=true
• committed_capacity_amount = sum of bank_capacity_used (not gross approved amounts)
• Declined applications get approved_amount=0.0, bank_capacity_used=0.0, conditions=[]

3. Credit-Union Segment Posture (e.g., CIVIC_NC_FIRE_EMS)

• peer_states: use the exact list from the segment endpoint (segment["peer_states"])
• required_checklist_gates: use the exact list from segment["minimum_checklist"]
• state_metrics values must come from the NCUA benchmark table (not segment internal_context)
• risk_tolerance: use segment["risk_tolerance"] when available
• For peer median comparisons, compute median of each metric across peer states

4. Watch-List Stress Packet (e.g., SUMMIT)

• Adverse population: loans with current_rating >= 6
• CDFI risk class: compute factor score from available debt_to_asset, fico, liquidity_months, ltv using the tables in /api/policies
  • Score ≤5→Prime, ≤9→Desirable, ≤13→Satisfactory, ≤18→Watch, >18 & ltv>1.0→Projected Loss, otherwise Doubtful
• Watch-list stress formula: stressed_dscr = round(dscr / 1.18, 2) (policy watch_list_formula)
• Shock label: +200bp watch-list parallel shock
• Breach threshold: 1.0
• Workout queue actions: risk-class driven—Projected Loss→legal_referral, Doubtful→partial_chargeoff_review, Watch→workout, Satisfactory→special_assets, Desirable→watchlist, Prime→monitor
• monitoring_cadence: monthly if any Projected Loss / Doubtful / Nonaccrual exists, otherwise quarterly

5. Competing CRE Decision (e.g., HARBOR)

• Compare only the two specified CRE application IDs
• CRE weighted score: approximate the 5 C's with 1–4 scale, weight by policy (capacity 0.45, collateral_exposure 0.36, etc.), round to 1 decimal
  • capacity from DSCR, collateral from LTV, capital from debt-to-asset or equity ratio, character from FICO/delinquencies/bankruptcy
• CRE dual-stress: stressed_dscr = round(dscr * 0.85 / 1.18, 2)
• existing_cre_exposure includes all real-estate-related sectors: Construction, Hospitality, Office, Multifamily, Residential, Retail CRE, Industrial CRE
• Concentration denominator: total_loans_outstanding
• unselected_reason_codes allowed values: sector_breach, weak_dscr, high_ltv, fdic_adverse_variance. Include fdic_adverse_variance whenever branch delinquency exceeds the FDIC benchmark.
• Conditions: always include no_additional_cre_without_committee_review when branch CRE is already near/above policy limit

Common Pitfalls

• Do NOT use total_assets for sector or CRE concentration denominators; use total_loans_outstanding
• Do NOT compute NPA by summing all non-current loans; use the branch metric nonperforming_loans
• Do NOT guess peer states for credit unions; read them from the segment endpoint
• Do NOT omit monitor from watch_list_action_coverage.by_action
• Do NOT compute concentration_flags post_approval_pct using aggregate after ALL approvals; compute per-application
• Do NOT forget to sort every list to the ordering specified in the template
• Do NOT include narrative outside the JSON object