# Self Attempt 03 > Credit Office Committee JSON SOP - Skill: `prism-shadow/self-attempt-03-3` (Agent Skill) - Install (CLI): `npx skillmds@latest add prism-shadow/self-attempt-03-3` - Raw SKILL.md: https://api.skillmd.com/api/skills/prism-shadow/self-attempt-03-3/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: prism-shadow (https://skillmd.com/u/prism-shadow) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/prism-shadow/self-attempt-03-3 --- # Credit Office Committee JSON SOP Use this skill for credit-office tasks that ask for a committee-ready JSON answer about a bank branch, loan portfolio, pending applications, CRE comparison, watch-list stress, or credit-union segment posture. ## Source Discipline - Use the remote public API only: ``. - Do not start, inspect, or read any local environment service, database, generated data file, run output, train output, test file, notes, reports, judge endpoint, or prior skill attempt. - Read the task prompt and its `input/payloads/answer_template.json` first. The template is the output contract for top-level keys, enum spellings, precision, required empty arrays/maps, and ordering. - Fetch only public API data needed for the target branch or segment plus `/api/manifest`, `/api/policies`, and the named benchmark endpoint. - Return exactly one valid JSON object. Do not include prose outside JSON. ## API Checklist Start with: - `/api/manifest` for `policy_version` and benchmark versions. - `/api/policies` for rating thresholds, CDFI scoring, CRE weights, stress formulas, capacity rules, and material downgrade threshold. For branch tasks, fetch as needed: - `/api/branches/{branch_id}`: capacity, `sector_ceiling_pct`, `cre_policy_limit_pct`, total assets, state, benchmark set. - `/api/branches/{branch_id}/metrics`: quarter metrics. For a 2025-03-31 review, use `2025Q1`. - `/api/branches/{branch_id}/loans`: balances, ratings, payment status, DSCR, LTV, FICO, debt-to-asset, liquidity, loan type, sector. - `/api/branches/{branch_id}/sector-exposures`: current sector exposure, `limit_pct`, grandfathering. - `/api/branches/{branch_id}/applications`: pending underwriting factors and requested amounts. For benchmark/segment tasks: - `/api/benchmarks/fdic/q4-2024`: FDIC ratios such as `total_loans_noncurrent_pct`, `total_real_estate_30_89_pct`, and `total_real_estate_noncurrent_pct`. - `/api/benchmarks/ncua/q1-2025`: state rows with integer `delinquency_bps`, `loan_to_share_pct`, `roaa_bps`, and `positive_net_income_pct`. - `/api/credit-union-segments/{segment_id}`: state, peer states, checklist gates, internal context, capacity, current outstanding, and risk tolerance. ## Common Field Conventions - Use `outstanding_balance` for loan exposure and `requested_amount` for application gross amount. - `metrics.total_loans_outstanding` equals the sum of branch loan balances and is the denominator for branch portfolio, NPA, CRE, and sector concentration ratios unless the prompt says otherwise. - For post-approval concentration, use a post-approval denominator: current `total_loans_outstanding` plus retained approved exposure. - Sector limits come from `sector-exposures.limit_pct`; fall back to branch `sector_ceiling_pct` only when the sector has no override row. - CRE exposure is the sum of `outstanding_balance` for loans with `loan_type == "CRE"`. CRE concentration is CRE exposure divided by total loans. - Keep ratios as decimals, not whole percentages. Basis points are `ratio * 10000`. - Round only at the output boundary: currency to 2 decimals, ratios to 4 decimals, bps to 2 decimals, DSCR to 2 decimals, CRE weighted scores to 1 decimal, unless the template says otherwise. ## Risk Rating Regrade For a regrade population such as `current_rating >= 3`, filter by the requested current-rating minimum. Re-derive the final rating from objective policy factors, not from the prior rating. Policy thresholds: - 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`. If LTV is missing but balance and collateral value exist, compute balance divided by collateral value. - Delinquency minimums: `30 Days Past Due -> 4`, `60 Days Past Due -> 5`, `90+ Days Past Due -> 7`, `Nonaccrual -> 8`, `Current -> no minimum`. - Final rating is the worst numeric rating from available DSCR, LTV/collateral, and delinquency factors. - A material downgrade is `final_rating - current_rating >= policies.risk_rating.material_downgrade_notches`. For NPA benchmarks: - `branch_npa_exposure = metrics.nonperforming_loans`. - `branch_total_loans = metrics.total_loans_outstanding`. - `branch_npa_ratio = branch_npa_exposure / branch_total_loans`. - `variance_ratio = branch_ratio - benchmark_ratio`; `variance_bps = variance_ratio * 10000`. Preserve the sign unless the template asks for absolute variance. For problem-credit selection, rank first by most severe final rating, then by nonaccrual/90+ status, then by largest exposure. Use only the action enums allowed by the template; a practical severity ladder is `monitor`, `watchlist`, `special_assets`, `workout`, `partial_chargeoff_review`, `legal_referral`. ## CDFI Factor Scores And Watch-List Stress For adverse-rated or watch-list loans, compute `factor_score` from `policies.cdfi_factor_scores` using available objective factors: - FICO bins, LTV bins, debt-to-asset bins, and liquidity-month bins add together. - Exact boundary values belong to the explicit middle range where listed, such as `0.40-0.60` or `680-720`. - Class by total score: `0-5 Prime`, `6-9 Desirable`, `10-13 Satisfactory`, `14-18 Watch`, `>=19 Doubtful`. - Override to `Projected Loss` when score is `>=19` and LTV is `>1.0`. For watch-list DSCR stress: - Use the policy formula exactly: `stressed_dscr = dscr / (1 + 0.18)`, even when the label says `+200bp`. - Use only loans with DSCR available in stress result rows. - `breaches_threshold` is true when stressed DSCR is below `policies.stress.coverage_breach_threshold` unless the prompt defines a stricter comparison. - Sort stress rows and breach ID lists exactly as the template requires. ## Application Allocation And Concentration For pending application packages: - Capacity starts from `branches.lending_capacity_q1`. - `gross_approved_amount` is the sum of approved or conditionally approved gross requested amounts. - `bank_capacity_used` and `committed_capacity_amount` should reflect retained bank exposure after any SBA guaranty, participation, or reduced-amount condition. If no mitigation applies, retained exposure equals approved amount. - `remaining_capacity = lending_capacity_q1 - committed_capacity_amount`. - Rank only approved and conditionally approved applications in `priority_ranking`. For concentration flags: - Existing sector exposure is `sector-exposures.current_exposure`. - `exposure_after_approval = current_exposure + retained approved exposure for that sector`. - `post_approval_pct = exposure_after_approval / (current total loans + total retained approved exposure)`. - `over_limit` or flags should compare post-approval percentage to the sector-specific `limit_pct`. - Grandfathered existing over-limit exposure may be tolerated, but a new approval should not worsen it without a template-allowed mitigation such as participation, reduced amount, board exception, or decline. Reason codes should come only from the template. Typical mappings are: weak DSCR to `weak_dscr`, high LTV to `high_ltv`, low FICO to `low_fico`, recent bankruptcy to `recent_bankruptcy`, incomplete documentation to `documentation_gap`, capacity shortfall to `capacity_limit`, sector over-limit to `sector_breach`, adverse FDIC comparison to `fdic_adverse_variance`, and weak NCUA comparison to `ncua_peer_weakness`. ## CRE Competition For competing CRE applications: - Fetch branch details, metrics, loans, sector exposures, the target applications, policies, and FDIC benchmark data. - Use `policies.cre_weighted_score.weights`: capacity `0.45`, capital `0.03`, character `0.05`, collateral exposure `0.36`, conditions `0.11`. - Lower weighted score is better. Classify with policy thresholds: `<=2.0 approve_quality`, `<=3.0 conditional`, `>3.0 weak`. - Derive components from raw factors: DSCR for capacity, LTV for collateral exposure, total debt divided by total assets for capital when present, FICO/bankruptcy/prior delinquencies/relationship/guarantor for character, and documentation, sector concentration, CRE policy pressure, and FDIC variance for conditions. - CRE dual stress uses `stressed_dscr = dscr * 0.85 / (1 + 0.18)`. - Selected post-approval CRE concentration is `(existing CRE exposure + selected retained exposure) / (current total loans + selected retained exposure)`. - Policy variance bps is `(post_approval_cre_concentration - cre_policy_limit_pct) * 10000`. - Sort application comparisons by `application_id`; sort reason code and condition lists alphabetically when required. ## Credit-Union Segment Posture For segment posture pages: - Use the segment endpoint for `state_code`, `peer_states`, `minimum_checklist`, `internal_context`, `quarterly_capacity`, `current_outstanding`, notes, and risk tolerance. - Use the NCUA benchmark row for the segment state, the `US` row, and all named peer-state rows. - For `peer_comparison`, compare the state value to the US value and to the median of the peer-state values. Return only `higher`, `lower`, or `equal`. - State metrics in the template are integer values exactly as reported by NCUA, not ratios. - Required checklist gates usually come directly from `minimum_checklist`; add operating controls from current context, such as missing insurance or lien follow-up, state benchmark monitoring, second review, or capacity exceptions. - Choose escalation triggers and owners only from the template enums, aligning owner responsibility with the trigger type. ## Output Validation Before finalizing: - Confirm every required top-level key is present and no uncontrolled enum value appears. - Confirm all specified sort orders: IDs ascending, final ratings ascending, actions alphabetically, workout queues by descending exposure then ID, or template-specific order. - Confirm numeric values are numbers, booleans are booleans, and absent result groups are represented as empty arrays or objects if required. - Do not include candidate train answers, hidden reasoning, citations, markdown fences, or explanatory text outside the JSON response.