Reflect 3 Attempt 02

Credit Office API Workflow Skill

Overview

Tasks require querying a remote credit-office public API and producing a committee-ready JSON answer matching a strict template. The API base URL is provided in environment_access.md. Do not run local env/setup.sh.

API Endpoints (always use remote public API)

• GET /api/manifest – lists all available endpoints and benchmark versions
• GET /api/branches – all branches
• GET /api/branches/{branch_id} – branch details (capacity, limits, state)
• GET /api/branches/{branch_id}/metrics – quarterly metrics (use Q1 2025)
• GET /api/branches/{branch_id}/loans – loan portfolio
• GET /api/branches/{branch_id}/sector-exposures – sector concentrations
• GET /api/branches/{branch_id}/applications – pending applications
• GET /api/policies – rating thresholds, CDFI score tables, stress formulas
• GET /api/benchmarks/fdic/q4-2024 – FDIC benchmarks
• GET /api/benchmarks/ncua/q1-2025 – NCUA benchmarks (rows by state_code)
• GET /api/credit-union-segments/{segment_id} – segment data

Common Business Rules

Risk Rating Re-derivation (from 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 = worst of available DSCR, LTV, and delinquency factors.
• If a factor is missing (None), skip it.

CDFI Factor Score (from policies)

• debt_to_asset: <0.40→0, 0.40-0.60→2, 0.60-0.80→4, >0.80→6
• fico: >720→0, 680-720→1, 580-679→3, <580→5
• liquidity_months: >12→0, 6-12→1, 3-6→3, <3→5
• ltv: <0.40→0, 0.40-0.60→2, 0.60-0.80→4, >0.80→6
• Sum all available factor scores.
• Risk class: ≤5 Prime, ≤9 Desirable, ≤13 Satisfactory, ≤18 Watch, ≥19 & ltv>1.0 Projected Loss, ≥19 Doubtful.

Stress Formulas (from policies)

• Watch-list +200bp: stressed_dscr = dscr / (1 + 0.18)
• CRE dual-stress: stressed_dscr = dscr * 0.85 / (1 + 0.18)
• Breach threshold = 1.0

Material Downgrade

• Downgrade notches = final_rating − current_rating
• Material = ≥2 notches (policy: material_downgrade_notches: 2)

NPA Benchmark

• branch_npa_exposure = nonperforming_loans from branch metrics (Q1 2025)
• branch_total_loans = total_loans_outstanding from branch metrics
• branch_npa_ratio = branch_npa_exposure / branch_total_loans (4 decimals)
• Use total_loans_noncurrent_pct as the default FDIC benchmark metric unless the task specifically calls for real-estate or construction.

Output Conventions

Ordering

• Lists of loans/applications: ascending by loan_id or application_id
• Action buckets: ascending alphabetically by action name
• Sectors: ascending alphabetically
• Watch-list workout queue: descending by exposure, then ascending loan_id
• Severe bucket counts: ascending by current_rating, then by payment_status order (Current, 30 DPD, 60 DPD, 90+ DPD, Nonaccrual)

Precision

• Currency: round to 2 decimals
• Percentages/ratios: round to 4 decimals
• Basis points: round to 2 decimals
• DSCR: round to 2 decimals

Enum Values

• Always use exact enum strings from the template. Never invent values.
• Payment status: Current, 30 Days Past Due, 60 Days Past Due, 90+ Days Past Due, Nonaccrual
• Actions: monitor, watchlist, special_assets, workout, partial_chargeoff_review, legal_referral
• Decisions: approve, conditional_approve, decline, defer, participation_required
• Conditions: participation_required, reduced_amount, board_exception, sba_guaranty_required, startup_monitoring, none

Task-Specific Patterns

Rating Migration Review (e.g., train_001)

1. Fetch branch loans, filter by current_rating >= target_min
2. Re-derive final ratings using DSCR/LTV/delinquency rules
3. Build final_rating_exposure_totals and migration_from_current_rating_3
4. watch_list_action_coverage maps final ratings to actions (3→monitor, 4→watchlist, 5→special_assets, 6→workout, 7→partial_chargeoff_review, 8→legal_referral)
5. top_problem_credit = worst final rating, tie-break by highest exposure

Allocation Package (e.g., train_002)

1. Fetch applications, sector exposures, branch capacity
2. Compute post-approval sector % = (current_exposure + approved_amount) / total_loans_outstanding
3. Decision logic:
   • Clean + within limits → approve
   • Minor issue (weak DSCR, high LTV) → conditional_approve with relevant condition
   • Sector breach → conditional_approve with participation_required or reduced_amount
   • Multiple severe issues → decline
4. bank_capacity_used = approved_amount for approve/conditional; for participation_required use 50% of approved amount
5. priority_ranking = all approved/conditional apps sorted by amount descending
6. decline_reasons = sorted list of reason codes per declined app
7. post_approval_concentrations = every sector with exposure_after_approval

Credit-Union Posture (e.g., train_003)

1. Fetch segment and NCUA benchmarks
2. State metrics = exact integer values from NCUA row for the segment's state
3. Peer comparison: compute median of peer_states for each metric; direction = higher/lower/equal vs US and vs peer median
4. required_checklist_gates = segment's minimum_checklist (sorted)
5. added_operating_controls = choose controls that address segment's noted issues (e.g., insurance binder verification for insurance gaps)
6. Posture: if delinquency above peers but capacity available → continue_with_tighter_conditions
7. interpretation fields must be consistent with posture and metrics

Watch-List Stress (e.g., train_004)

1. Fetch loans with current_rating >= adverse_min
2. Compute CDFI factor scores and risk classes
3. monitoring_cadence: if any Watch/Doubtful/Projected Loss → monthly, else quarterly
4. Stress all loans with DSCR available; breaches_threshold if stressed_dscr < 1.0
5. Workout queue: map risk class to action; projected_loss = true for Nonaccrual or 90+ DPD

Competing CRE Decision (e.g., train_005)

1. Fetch both applications, branch loans, metrics, FDIC benchmark
2. Compute weighted CDFI score using policy weights (capacity 0.45, capital 0.03, character 0.05, collateral_exposure 0.36, conditions 0.11)
3. Score class: ≤2.0 approve_quality, ≤3.0 conditional, >3.0 weak
4. CRE dual-stress both applications; select the one that does NOT breach
5. Unselected gets decline with reason codes sorted alphabetically
6. Concentration: existing_cre_exposure = sum of CRE loans; post-approval = existing + selected amount
7. selected_policy_variance_bps = (post_conc − cre_policy_limit_pct) × 10000
8. fdic_benchmark_metric = total_real_estate_30_89_pct

Common Pitfalls

• Do not use localhost or env/setup.sh; always use the remote API URL from environment_access.md
• Round precisely: currency to 2 decimals, ratios to 4 decimals, bps to 2 decimals
• Sort correctly: ascending loan_id, alphabetical action, descending exposure for workout queue
• Skip missing factors: if DSCR or LTV is None, do not include it in rating derivation
• Use Q1 2025 metrics: metrics endpoint returns two quarters; pick the one with quarter="2025Q1"
• Enum exactness: any value not in the template's allowed enum will fail validation
• Do not include narrative outside the JSON object
• Total loans denominator: for concentrations, use total_loans_outstanding from metrics, not total_assets