Task Group 011 Fewshot Attempt 01

Input-Output Credit JSON

Guardrails

• Use only the prompt, its input/payloads/answer_template.json, and public HTTP API endpoints supplied by the user or printed by setup.
• Do not read local env/, test_tasks/, notes, reports, run artifacts, other generated skills, or task metadata files.
• Treat answer_template.json as authoritative for required keys, enums, ordering, precision, and empty-list/object behavior.
• Return one valid JSON object only. Do not include Markdown, comments, or explanatory prose outside the JSON.

Public API Workflow

1. Read the task prompt and template. Extract target branch_id, segment_id, application IDs, as-of date, review date, rating cutoff, and requested benchmark period.
2. Discover the public API from /api/manifest when needed. Expected surfaces include:
   • /api/branches/{branch_id}
   • /api/branches/{branch_id}/metrics
   • /api/branches/{branch_id}/loans
   • /api/branches/{branch_id}/sector-exposures
   • /api/branches/{branch_id}/applications
   • /api/policies
   • /api/benchmarks/fdic/q4-2024
   • /api/benchmarks/ncua/q1-2025
   • /api/credit-union-segments/{segment_id}
3. Fetch only the surfaces relevant to the prompt. Prefer the branch/segment specified by the task rather than listing all records.
4. Use the latest relevant metric row unless the prompt specifies a quarter. For 2025-03-31 tasks, this is generally 2025Q1; FDIC benchmark tasks use fdic_q4_2024; NCUA tasks use ncua_q1_2025.
5. Build calculations in a scratch table, then shape the final JSON from the template. Sort every list exactly as the template says.

Numeric Conventions

• Currency fields: sum raw values, then round final outputs to 2 decimals.
• Ratio fields: output decimals, not percentages; round final outputs to 4 decimals unless the template says otherwise.
• Basis points: (branch_ratio - benchmark_ratio) * 10000, rounded to 2 decimals.
• Use outstanding_balance for loan exposure and existing portfolio sums. Use requested_amount, approved amount, and bank-retained amount separately for applications.
• For boolean fields, emit JSON booleans true/false, not strings.

Rating Migration Tasks

• Filter the branch loans by the prompt's current-rating cutoff, such as current_rating >= 3 or adversely rated current_rating >= 6.
• Re-derive final risk rating from public policy rules. Compute available component ratings and take the worst numeric rating:
  • DSCR: >=1.50 -> 3, >=1.25 -> 4, >=1.05 -> 5, >=1.00 -> 6, <1.00 -> 7.
  • LTV: <=0.65 -> 3, <=0.75 -> 4, <=0.85 -> 5, <=1.00 -> 6, >1.00 -> 7.
  • Payment status minimums: 30 Days Past Due -> 4, 60 Days Past Due -> 5, 90+ Days Past Due -> 7, Nonaccrual -> 8; Current has no minimum.
• A material downgrade is normally final_rating - current_rating >= policy.risk_rating.material_downgrade_notches.
• Group rating exposure totals by final rating; group migrations by the requested starting current rating when the template asks for it.
• For NPA/noncurrent benchmark comparisons, identify nonperforming exposure from nonaccrual or public metric fields as the prompt/template implies. Divide by total loans outstanding and compare to the named FDIC metric.
• Pick the top problem credit by severity first, then exposure: nonaccrual/projected-loss indicators, highest final rating, most severe payment status, and largest exposure.

CDFI Factor Scoring And Watch Lists

• Score objective factors from policy.cdfi_factor_scores:
  • Debt-to-asset, FICO, liquidity months, and LTV each contribute the bucket score shown by policy.
  • Sum only available factor scores; do not invent missing FICO or DSCR factor scores.
• Map total score to policy classes: Prime, Desirable, Satisfactory, Watch, Doubtful; apply Projected Loss when policy conditions and credit facts show loss severity, especially underwater collateral plus nonaccrual/projected-loss notes.
• For watch-list DSCR stress, use policy.stress.watch_list_formula: stressed_dscr = dscr / (1 + 0.18). Include only loans with DSCR available, round DSCRs to 2 decimals, and flag breaches below the policy threshold.
• Queue actions by severity using controlled enum values. Use public policy, payment status, notes, and risk class together:
  • partial_chargeoff_review for projected loss, nonaccrual, or clear loss-review facts.
  • special_assets for 90+ past due, rating 7+ severity, or structurally severe credits needing active handling.
  • watchlist for rating 6 or stressed/weak credits needing close monitoring.
  • Use monitor, workout, or legal_referral only when the facts support those exact paths.

Lending Allocation Tasks

• Start with branch.lending_capacity_q1. Approved applications consume bank capacity unless participation, guaranty, decline, or deferral changes the retained exposure.
• Apply policy floors and reason codes from objective facts:
  • weak_dscr for DSCR below policy/task minimums or stress breach.
  • high_ltv or underwater_collateral for collateral weakness.
  • low_fico, recent_bankruptcy, startup_risk, documentation_gap, or policy_floor_missing when those fields trigger.
  • capacity_limit when the remaining branch capacity cannot support the request.
  • sector_breach when a new approval would exceed the sector limit without mitigation.
  • fdic_adverse_variance or ncua_peer_weakness only when the relevant benchmark comparison is adverse.
• Participation may leave approved_amount equal to the full request while bank_capacity_used is capped. For sector caps, solve retained exposure so (current_sector_exposure + retained_amount) / (current_total_loans + other_committed_retained_amounts + retained_amount) <= limit_pct.
• For SBA guaranty, bank capacity used is usually requested_amount * (1 - sba_guaranty_pct) when the guaranty is the risk mitigant.
• Priority ranking includes approved and conditionally approved applications only, ordered by credit strength, strategic fit, mitigated concentration impact, and capacity use.
• Post-approval concentration views generally use full approved exposure for sector exposure and the post-approval total loan base. Mark over_limit after rounding only if the raw ratio exceeds the limit.

Competing CRE Tasks

• Compare only the application IDs named by the prompt.
• Calculate the CRE weighted score from policy.cre_weighted_score.weights; lower is better. Use the public application facts for capacity/DSCR, capital leverage, character/relationship/delinquency, collateral/LTV/exposure, and external conditions. Round the weighted result to 1 decimal and map it to policy score classes.
• Apply the CRE dual stress from policy: stressed_dscr = dscr * 0.85 / (1 + 0.18); breach threshold is usually 1.00.
• Evaluate concentration using branch CRE limits, existing CRE/sector exposure, selected approval amount, and FDIC underperformance. Existing over-limit exposure may be grandfathered, but new approvals should not worsen it without participation or committee exception.
• Select the stronger credit by lower weighted score, stress result, collateral quality, relationship strength, and manageable concentration treatment. Give the unselected credit only enum reason codes allowed by the template.

Credit-Union Segment Posture Tasks

• Fetch the segment and NCUA benchmark rows. Use segment.state_code, segment.peer_states, quarterly capacity, current outstanding, checklist, internal context, and risk tolerance.
• Compare the target state against US and the median of peer states for delinquency_bps, loan_to_share_pct, roaa_bps, and positive_net_income_pct. Output directions as higher, lower, or equal.
• Choose posture from capacity and external risk:
  • Use continue_approving only when capacity and benchmark posture are routine.
  • Use continue_with_tighter_conditions when capacity exists but peer/national metrics or internal controls are weaker.
  • Use temporarily_pause when capacity is unavailable or external/control risk is unacceptable.
• Preserve segment checklist gates. Add operating controls that directly respond to benchmark weakness, missing close controls, lien/insurance risk, capacity exceptions, or senior-review needs.
• Escalation triggers should have stable ascending IDs and owners that match the control area: credit risk for delinquency/benchmarks, operations for lien/insurance exceptions, committee chair for capacity exceptions.

Final Validation

• Confirm all required top-level keys exist and no extra narrative is present.
• Confirm all enum strings exactly match the template.
• Confirm every list's ordering rule, including nested loan_ids, reason codes, trigger IDs, ratings, sectors, and application IDs.
• Confirm numeric precision after all calculations.
• Confirm output parses as JSON before submitting.