Self Attempt 02

Credit Office Committee JSON Skill

Use this skill for Credit Office / lending committee tasks that ask for a JSON answer over the public credit office API.

Ground Rules

• Use only the task prompt, its input/payloads/answer_template.json, and the public API described by environment_access.md.
• Do not inspect local environment source/data, train outputs, test files, notes, prior runs, reports, other attempts, or judge endpoints.
• Prefer the remote public API directly. In this workspace the base URL is <environment_base_url>.
• Return exactly one valid JSON object. Do not add prose around it.
• Treat answer_template.json as the contract: required keys, enum spellings, ordering, and precision rules override intuition.

API Workflow

1. Read the prompt for target ids, review/as-of date, benchmark family, and requested analysis.
2. Read the answer template and list every required key, enum, sort order, and rounding instruction.
3. Fetch /api/manifest and /api/policies first to confirm policy and benchmark versions.
4. Fetch only the public records needed:
   • Branch base data: /api/branches/{branch_id}
   • Metrics: /api/branches/{branch_id}/metrics
   • Loans: /api/branches/{branch_id}/loans
   • Sector exposure: /api/branches/{branch_id}/sector-exposures
   • Applications: /api/branches/{branch_id}/applications
   • FDIC benchmark: /api/benchmarks/fdic/q4-2024
   • NCUA benchmark: /api/benchmarks/ncua/q1-2025
   • Credit-union segment: /api/credit-union-segments/{segment_id}
5. Compute with a script or structured JSON tooling, then validate the final object against the template before answering.

Field Conventions

• Loan exposure is outstanding_balance.
• Application exposure/capacity use is usually requested_amount, unless the decision explicitly reduces or participates the amount.
• Branch capacity is lending_capacity_q1.
• Branch sector limits come from sector-exposures.limit_pct; fall back to branches.sector_ceiling_pct only if no sector-specific row exists.
• CRE policy limit is branches.cre_policy_limit_pct.
• Metrics rows use quarter, not an as_of_date; select the row matching the review quarter.
• In metrics, map nonperforming_loans to NPA exposure and total_loans_outstanding to total loans.
• Ratios are decimals, not percentages. Basis points are ratio * 10000.
• Currency is rounded to 2 decimals. Ratio fields usually round to 4 decimals. DSCR fields usually round to 2 decimals. Weighted CRE score rounds to 1 decimal.

Rating Migration Tasks

For regrading loans, filter the exact population requested, commonly loans with current_rating >= target_min.

Apply the policy risk_rating dominant-factor rule: final rating is the worst numeric rating from available DSCR, LTV/collateral, and delinquency factors.

• DSCR rating: >=1.50 -> 3, >=1.25 -> 4, >=1.05 -> 5, >=1.00 -> 6, <1.00 -> 7.
• LTV rating: <=0.65 -> 3, <=0.75 -> 4, <=0.85 -> 5, <=1.00 -> 6, >1.00 -> 7.
• Delinquency minimum: 30 Days Past Due -> 4, 60 Days Past Due -> 5, 90+ Days Past Due -> 7, Nonaccrual -> 8, Current -> no floor.
• Material downgrade is final_rating - current_rating >= material_downgrade_notches.

Aggregate exposure by final rating, sort ascending by rating, and sort loan ids ascending where requested. For NPA benchmark variance, use:

• branch_npa_ratio = nonperforming_loans / total_loans_outstanding
• variance_ratio = branch_ratio - fdic_benchmark_ratio
• variance_bps = variance_ratio * 10000

Choose problem-credit and watch-list actions from the template enums only. Tie severity to final rating, payment status, projected loss/collateral weakness, and exposure.

Watch-List Stress Tasks

Adverse-rated watch lists usually mean loans with current_rating >= 6 unless the prompt gives another threshold.

For CDFI factor scoring, sum the available policy scores for fico, ltv, liquidity_months, and debt_to_asset.

• Classes: Prime 0-5, Desirable 6-9, Satisfactory 10-13, Watch 14-18, Doubtful 19+, and Projected Loss when score is 19+ and ltv > 1.0.
• Use only the risk class enum in the template.
• If DSCR is missing, omit that loan from DSCR stress results but keep it in population and workout aggregates.
• Watch-list stress formula: stressed_dscr = dscr / (1 + 0.18).
• Breach threshold is the policy coverage_breach_threshold, normally 1.00.

Workout queues are typically sorted by descending exposure, then ascending loan id. Severe bucket counts group by current_rating and payment_status, sorted as the template requires.

Application Allocation Tasks

For pending-application packages:

• Fetch branch, sector exposures, and all pending applications for the target branch.
• Evaluate application-level policy issues: DSCR, LTV, FICO, recent bankruptcy, startup risk, documentation, SBA guaranty, capacity, and sector concentration.
• Keep decisions and conditions to template enums exactly.
• Capacity math:
  • gross_approved_amount is total approved or conditionally approved face amount.
  • committed_capacity_amount is the bank-retained amount after reductions or participation.
  • remaining_capacity = lending_capacity_q1 - committed_capacity_amount.
• Sector post-approval concentration:
  • exposure_after_approval = current_exposure + bank_capacity_used for that sector
  • post_approval_pct = exposure_after_approval / total_assets
  • compare to the sector-specific limit_pct.
• Existing over-ceiling exposure can be grandfathered, but a new approval should not worsen it without mitigation.
• Sort decisions by application_id, concentration flags by sector then application_id, post-approval concentrations by sector, and reason-code lists alphabetically.

Competing CRE Decision Tasks

When comparing named CRE applications, filter the application list to the exact ids in the prompt.

• Use policy cre_weighted_score.weights: capacity .45, capital .03, character .05, collateral/exposure .36, conditions .11.
• Lower weighted score is better. Classify by policy: approve_quality <= 2.0, conditional <= 3.0, weak > 3.0.
• If component scores are not explicit fields, derive them transparently from objective application factors: DSCR for capacity, debt-to-assets/net income for capital, relationship/delinquency/bankruptcy for character, LTV/collateral for collateral exposure, and documentation/concentration/benchmark environment for conditions.
• CRE stress formula: stressed_dscr = dscr * 0.85 / (1 + 0.18).
• Use the policy coverage breach threshold, normally 1.00.
• CRE concentration after selecting an application:
  • existing CRE exposure comes from existing CRE/CRE-like sector exposure as requested by the prompt.
  • post-approval concentration adds the selected bank-retained amount and divides by total_assets.
  • policy variance bps is (post_approval_concentration - cre_policy_limit_pct) * 10000.
• FDIC 30-89 or noncurrent benchmark variance uses branch_ratio - benchmark_ratio; keep the exact metric enum from the template.

Credit-Union Segment Posture Tasks

For segment posture pages:

• Fetch the segment endpoint, policies, manifest, and NCUA benchmark table.
• Use the segment state_code, peer_states, quarterly_capacity, current_outstanding, risk_tolerance, minimum_checklist, notes, and internal_context.
• State metrics come directly from the NCUA row for the segment state; do not convert bps or integer percentages unless the template says to.
• Peer comparison directions are relative to US and to the median of named peer states. Return only higher, lower, or equal.
• Required checklist gates should include the segment minimum checklist, restricted to template choices.
• Added operating controls should respond to the actual context: insurance/lien gaps, elevated delinquency, capacity exceptions, staffing/second review, or benchmark monitoring.
• Escalation triggers must use allowed trigger ids/conditions/owners only and be sorted ascending by trigger_id.
• Match posture and interpretation enums to the evidence: capacity availability, external risk status, and risk tolerance.

Final QA Checklist

• All required top-level keys are present and no narrative text is outside the JSON.
• Every enum value is copied exactly from the template.
• Every list follows the requested ordering.
• Currency, ratios, DSCRs, scores, and bps use the requested precision.
• Benchmark versions and policy versions come from the API, not memory.
• Denominators are correct: total assets for concentration, total loans outstanding for NPA ratios, and capacity for allocation.
• Nulls, missing DSCR values, and missing FICO values are handled deliberately rather than coerced to zero.
• The answer contains no training-output values or test-answer leakage.