Reflect 3 Attempt 01

Credit Office Committee JSON Skill

Overview

Generate committee-ready JSON answers for credit risk decisions by querying the public API and strictly conforming to the task's answer_template.json schema.

Test-Time Workflow

1. Read the prompt to identify the target branch/segment, review date, and specific deliverable.
2. Read input/payloads/answer_template.json before fetching any data. Record:
   • required_top_level_keys
   • Every enum field and its allowed values
   • Ordering rules for lists
   • Numeric precision requirements
3. Fetch data from the public API using the base URL in environment_access.md:
   • Start with /api/manifest to discover endpoints
   • Fetch branch/segment details, loans, applications, metrics, sector exposures, policies, benchmarks
4. Apply policy rules from /api/policies to derive ratings, scores, and decisions.
5. Build JSON output matching the template exactly. No narrative text outside the JSON.
6. Validate before returning:
   • All required top-level keys present
   • Enum values match exactly (case-sensitive)
   • Lists sorted per template ordering rules
   • Numbers rounded to specified precision
   • No extra keys at any level

API Endpoints (Common)

• GET /api/manifest — discover endpoints and record counts
• GET /api/branches — list branches
• GET /api/branches/{branch_id} — branch details (capacity, limits)
• GET /api/branches/{branch_id}/loans — loan portfolio
• GET /api/branches/{branch_id}/metrics — quarterly metrics
• GET /api/branches/{branch_id}/sector-exposures — sector limits and exposure
• GET /api/branches/{branch_id}/applications — pending applications
• GET /api/policies — risk rating, CDFI scoring, stress formulas, concentration rules
• GET /api/benchmarks/fdic/q4-2024 — FDIC benchmarks
• GET /api/benchmarks/ncua/q1-2025 — NCUA benchmarks
• GET /api/credit-union-segments/{segment_id} — segment details

Key Field Definitions & Business Rules

Risk Rating Re-derivation (/api/policies → risk_rating)

• DSCR thresholds: ≥1.5→3, ≥1.25→4, ≥1.05→5, ≥1.0→6, <1.0→7
• LTV thresholds: ≤0.65→3, ≤0.75→4, ≤0.85→5, ≤1.0→6, >1.0→7
• Delinquency floors: Current→null, 30 DPD→4, 60 DPD→5, 90+ DPD→7, Nonaccrual→8
• Dominant factor rule: final rating = max(available DSCR rating, LTV rating, delinquency floor)
• Material downgrade: migration ≥ 2 notches

CDFI Factor Scoring (/api/policies → cdfi_factor_scores)

• Only use available factors. Do NOT penalize missing fields.
• Score thresholds: Prime(0–5), Desirable(6–9), Satisfactory(10–13), Watch(14–18), Doubtful(≥19), Projected Loss(≥19 and LTV>1.0)
• Factor tables:
  • FICO: >720→0, 680–720→1, 580–679→3, <580→5
  • LTV: <0.40→0, 0.40–0.60→2, 0.60–0.80→4, >0.80→6
  • Debt-to-asset: <0.40→0, 0.40–0.60→2, 0.60–0.80→4, >0.80→6
  • Liquidity months: >12→0, 6–12→1, 3–6→3, <3→5

CRE Weighted Score (/api/policies → cre_weighted_score)

• Weights: capacity 0.45, capital 0.03, character 0.05, collateral_exposure 0.36, conditions 0.11
• Map each factor to a 1–5 score, then compute weighted sum.
• Classes: approve_quality (≤2.0), conditional (≤3.0), weak (>3.0)

Stress Formulas (/api/policies → stress)

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

Concentration & Capacity

• Sector limit: use sector_ceiling_pct from branch or per-sector limit_pct from sector-exposures
• Post-approval % = (current_exposure + requested_amount) / total_loans_outstanding
• CRE exposure = sum of loan_type == 'CRE' outstanding balances only
• CRE concentration = existing_cre_exposure / total_loans_outstanding

NPA / Benchmark Variance

• Branch NPA ratio = nonperforming_loans / total_loans_outstanding from branch metrics
• FDIC benchmark: use the metric specified in the template (e.g., total_loans_noncurrent_pct, total_real_estate_30_89_pct)
• Variance = branch_ratio − fdic_ratio; express in bps (×10000)

Output-Field Conventions

Common Pitfalls

1. Wrong template structure — Always match required_top_level_keys exactly. Adding or omitting a key at any level often drops the score to 0.
2. Missing enum value — A single invalid enum value invalidates the entire answer.
3. Incorrect list ordering — Judges check ordering rules strictly. Use Python's sorted() with the correct key.
4. Treating missing data as worst-case — For CDFI scoring, policy-derived ratings, and DSCR/LTV thresholds, only apply rules when the underlying field is present. Do not invent defaults.
5. Wrong denominator for concentration — Use total_loans_outstanding from branch metrics, not total assets.
6. NPA definition — Use branch metrics nonperforming_loans divided by total_loans_outstanding. Do not include 30/60 DPD loans unless they are in the metrics NPA figure.
7. Watch-list action coverage — Include only loans that actually need follow-up (material downgrade, derived rating ≥5, or non-current status). Do not blanket-cover all reviewed loans.
8. Priority ranking — In allocation tasks, include only approved and conditional_approved applications; sort by priority logic (quality first, then amount).
9. Concentration flags — Include ALL applications, not just breached ones. Set flag: true only when post-approval % > limit. Set handling to the app's decision when flagged, otherwise "none".
10. Post-approval concentrations — Include ALL sectors from sector-exposures, even sectors with no pending applications.
11. Trigger IDs — Use simple numeric IDs like "01", "02", "03" (not "TR-01").
12. Peer comparison — peer_states must exactly match the segment's peer_states list, sorted ascending.
13. State metrics — Report exact integers from the NCUA benchmark table; do not round or recalculate.
14. CRE scoring factor mapping — Capacity maps to DSCR, Capital to debt-to-asset, Character to years-in-business/delinquencies, Collateral to LTV, Conditions to documentation/relationship length.
15. Unselected reason codes — In competing-CRE decisions, the unselected app's reason_codes are restricted to a smaller allowed set (e.g., sector_breach, weak_dscr, high_ltv, fdic_adverse_variance).