Self Attempt 01

Credit Office Lending-Committee SOP

This skill solves committee-packet JSON tasks against a remote, read-only "credit office" API. Every task: read prompt -> identify target (branch_id / segment_id / application_ids) -> pull policies + the referenced data from the API -> compute per the policy ruleset -> assemble JSON exactly to input/payloads/answer_template.json -> self-check shapes, ordering, enums, rounding. There is NO judge/feedback loop; get it right in one pass.

The five known task archetypes (use the closest one as a template):

1. Rating migration review (branch). Re-derive risk ratings, migration buckets, material downgrades, NPA-vs-FDIC variance, top problem credit, watch-list action coverage.
2. Q1 allocation package (branch). Pending applications -> decisions, capacity allocation, sector concentration flags, decline reason codes, post-approval concentration.
3. Credit-union segment posture (segment). NCUA state benchmarks -> posture, peer comparison, controls, escalation triggers, interpretation.
4. Watch-list stress packet (branch). Adverse-rated loans -> CDFI risk classes, +200bp DSCR stress, workout queue, severe-bucket payment-status counts.
5. Competing CRE decision (branch). Two CRE applications -> weighted CDFI score, dual stress, CRE concentration vs FDIC, pick stronger credit + reason codes for loser.

1. Using the remote API (the ONLY data source)

Base URL: <remote-env-url>. Use curl -s (Bash tool). The environment is remote and read-only. IGNORE any prompt instruction to run env/setup.sh or read local env/ files — there is no local env. Never read local data files; always fetch live.

Endpoints (GET):

• /api/health, /api/manifest — record counts, benchmark versions, policy_version.
• /api/policies — THE business ruleset. Always fetch first. (see section 2).
• /api/branches and /api/branches/{branch_id} — branch attributes: cre_policy_limit_pct, sector_ceiling_pct (default sector limit), lending_capacity_q1, total_assets, state_code, institution_type, fdic_benchmark_set.
• /api/branches/{id}/metrics[?quarter=2025Q1] — list of quarterly rows. Use 2025Q1 (current/as-of) unless told otherwise. Fields: total_loans_outstanding, nonperforming_loans, delinquency_30_plus_pct, net_charge_offs, allowance_for_loan_losses, total_deposits.
• /api/branches/{id}/loans[?loan_type=&payment_status=&min_current_rating=] — loan list. min_current_rating=N returns loans with current_rating >= N (server-side; verified).
• /api/branches/{id}/sector-exposures — per-sector current_exposure, limit_pct (sector-specific, may differ from branch default ceiling), grandfathered (0/1).
• /api/branches/{id}/applications[?loan_type=] — pending applications.
• /api/benchmarks/fdic/q4-2024 — FDIC ratios (version fdic_q4_2024).
• /api/benchmarks/ncua/q1-2025[?state_code=NC] — NCUA per-state rows (version ncua_q1_2025).
• /api/credit-union-segments/{segment_id} — segment posture inputs.

Gotchas:

• Branch ids are upper-cased by the server. Use REDWOOD, LAKEVIEW, SUMMIT, HARBOR, etc. Segment ids stay as given (e.g. CIVIC_NC_FIRE_EMS).
• min_current_rating is inclusive (>=). Equivalent to filtering the full list yourself.
• Metrics is a list of multiple quarters; pick the right quarter.
• Credit-union branches have empty fdic_benchmark_set (they use NCUA, not FDIC).

2. Policy ruleset (/api/policies) — the canonical rules

policy_version: credit_policy_v2025Q1. Structure and how to apply each block:

2.1 risk_rating — re-deriving a loan's rating (lower = better; scale ~2..8)

Compute a candidate rating from each available factor, then take the worst (max) numeric rating across available factors (dominant_factor_rule). Factors:

• DSCR thresholds (dscr_thresholds): dscr >=1.5 ->3; >=1.25 ->4; >=1.05 ->5;

  =1.0 ->6; below 1.0 ->7.

• LTV thresholds (ltv_thresholds, this is the "LTV or collateral" factor): ltv <=0.65 ->3; <=0.75 ->4; <=0.85 ->5; <=1.0 ->6; above 1.0 ->7.
• Delinquency minimums (delinquency_minimums, keyed by payment_status): a rating FLOOR. Current->null (no floor); 30 Days Past Due->4; 60 Days Past Due->5; 90+ Days Past Due->7; Nonaccrual->8.
• final_rating = max of all non-null factor ratings. If NO factor is available (dscr, ltv, payment_status all null/Current), keep the loan's current_rating.
• material_downgrade_notches: 2 -> a downgrade is "material" when final_rating - current_rating >= 2. (Re-derivation can also produce a lower rating; those are NOT material downgrades, only count >= +2.)

2.2 cdfi_factor_scores — CDFI-style objective risk class (lower score = better)

Sum integer sub-scores from available factors, then map total to a class.

• 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.
• Treat bucket edges as: lower bucket is strict <, upper inclusive <= (e.g. ltv exactly 0.60 scores 2, exactly 0.80 scores 4; d2a 0.40 scores 2). Skip a factor when its field is null.
• Class by total score: 0-5 Prime; 6-9 Desirable; 10-13 Satisfactory; 14-18 Watch; >=19 Doubtful; Projected Loss only if score>=19 AND ltv>1.0 (the score>=19 gate must be met first; ltv>1.0 alone does NOT make Projected Loss).

2.3 cre_weighted_score — weighted CDFI score for CRE apps (lower = better)

• weights: capacity 0.45, collateral_exposure 0.36, conditions 0.11, character 0.05, capital 0.03.
• weighted_score = sum(weight_i * component_subscore_i), reported to precision 1.
• score_class: approve_quality <=2.0; conditional <=3.0; weak >3.0.
• Component sub-scores reuse the CDFI 0/2/4/6 and 0/1/3/5 scales: collateral_exposure <- LTV score (0/2/4/6); capital <- debt_to_asset score (debt_to_asset = total_debt/total_assets if not provided); character <- borrower quality (fico if present, else guarantor/relationship strength on a 0=strong .. 6=none scale); capacity <- a DSCR-strength score (stronger coverage = lower score, e.g.

  =1.5 ->0, >=1.25 ->2, >=1.0 ->4, <1.0 ->6); conditions <- macro/sector/exception posture (use 0 when no adverse condition flagged). NOTE: the policy fixes the weights and class cutoffs but NOT the exact capacity/character/ conditions sub-scales. Use CDFI-consistent 0/2/4/6 mappings; the decisive output is the relative ranking (stronger credit = lower weighted score) plus stress + concentration, so the chosen scales mainly need to rank the two apps correctly and land in the right class.

2.4 stress — DSCR stress formulas

• Watch-list / single shock (+200bp parallel shock): stressed_dscr = dscr / (1 + 0.18).
• CRE dual stress: stressed_dscr = dscr * 0.85 / (1 + 0.18).
• coverage_breach_threshold = 1.0. A loan/app breaches when stressed_dscr < 1.0.
• Round stressed_dscr to 2 decimals; base_dscr also 2 dp. Only include items where DSCR is available (skip null-DSCR loans from results/breach_loan_ids).

2.5 capacity_concentration — sector limits & capacity

• lending_capacity_q1 (branch field) is the quarter's lending capacity for allocation.
• Sector limit = sector_exposures.limit_pct for that sector, else branch sector_ceiling_pct default. CRE limit = branch cre_policy_limit_pct.
• Concentration denominator = total loan exposure = sum of all sector current_exposure = branch total_loans_outstanding (verified equal). NOT total_assets.
• Allowed mitigations: participation_required, reduced_amount, board_exception.
• grandfathering: existing over-ceiling exposure may be grandfathered (flag=1), but new approvals may not worsen that sector without a mitigation.

3. Benchmark metric selection & variance

Pick the FDIC/NCUA metric that matches the task's risk theme:

• NPA / noncurrent review -> FDIC total_loans_noncurrent_pct. Branch NPA population = loans with payment_status in {90+ Days Past Due, Nonaccrual} (i.e. "noncurrent"). branch_npa_exposure = sum of those balances (equals metric nonperforming_loans); branch_total_loans = sum of ALL loan balances (equals metric total_loans_outstanding).
• CRE / real-estate early delinquency -> FDIC total_real_estate_30_89_pct. The branch_delinquency_ratio comes directly from branch metric delinquency_30_plus_pct (2025Q1) — do NOT recompute from loans.
• Other FDIC metrics available: construction_development_noncurrent_pct, total_real_estate_noncurrent_pct, construction_development_30_89_pct.

Variance math (consistent everywhere):

• branch_ratio = branch value (4 dp). benchmark_ratio = FDIC/benchmark value (4 dp).
• variance_ratio = branch_ratio - benchmark_ratio (4 dp).
• variance_bps = variance_ratio * 10000 (2 dp).
• benchmark_version string: fdic_q4_2024 or ncua_q1_2025.

NCUA (segment/credit-union tasks): rows per state_code (plus a US row). Metrics are integers exactly as reported: delinquency_bps, loan_to_share_pct, roaa_bps, positive_net_income_pct. Peer states come from the segment's peer_states list.

• direction = NC value vs comparison: higher if NC>comp, lower if NC<comp, equal.
• Compare NC to US row (nc_vs_us) and to the median of peer-state values per metric (nc_vs_peer_median). Order peer_states ascending by state code.
• Polarity: for delinquency, lower is better (higher = weaker); for roaa_bps, loan_to_share_pct (within prudent range), positive_net_income_pct, higher is generally stronger. Net "external_risk_status" reflects whether NC is broadly weaker/stronger/mixed.

4. Action / disposition mappings

recommended_action enum (severity order): monitor < watchlist < special_assets < workout < partial_chargeoff_review < legal_referral. Map from severity of the (final) rating / CDFI class / payment status. Reasonable mapping:

• Performing & low-adverse (rating ~3-4, Prime/Desirable, Current) -> monitor.
• Adverse but performing (rating 5, Watch, Current) -> watchlist.
• Severe / early-default (rating 6, 60-day, Doubtful) -> special_assets.
• Nonaccrual / 90+ / Doubtful -> workout.
• Confirmed loss exposure (ltv>1.0 + nonaccrual, Projected Loss) -> partial_chargeoff_review.
• Legal/charge-off track for the worst -> legal_referral. The top/worst problem credit = highest final_rating, tie-break by largest exposure; its recommended_action is the most severe consistent with its status (Nonaccrual + ltv>1.0 => legal_referral or partial_chargeoff_review).

monitoring_cadence for adverse populations: monthly (most adverse) / quarterly / semiannual. Adverse watch-list populations -> monthly.

projected_loss boolean (workout queue) = true when CDFI class is Projected Loss (score>=19 and ltv>1.0) or loan is Nonaccrual with ltv>1.0.

5. Output formatting & assembly conventions

Always match input/payloads/answer_template.json exactly: top-level keys, nested required_keys, enum value sets, list orderings.

Precision/rounding (round half-to-even or standard 2dp is fine; match precision):

• Money / USD amounts: 2 decimals.
• Ratios / percentages-as-ratios (concentration, npa_ratio, variance_ratio): 4 decimals expressed as a fraction (e.g. 0.1135, NOT 11.35).
• bps fields: 2 decimals (variance_bps = ratio*10000).
• DSCR (base/stressed): 2 decimals.
• weighted_cdfi_score: 1 decimal.
• factor_score, ratings, counts: integers.
• NCUA state metrics: integers exactly as reported (no rounding/scaling).

List ordering (read the template per field; common rules seen):

• migration/final_rating lists: ascending by final_rating.
• loan lists / material_downgrades / risk_classes / stress results / severe buckets: ascending by loan_id (severe buckets: ascending current_rating, then payment_status).
• workout_queue: descending exposure, then ascending loan_id.
• decisions: ascending application_id; concentration_flags: by sector then application_id.
• post_approval_concentrations: ascending sector.
• reason_codes / conditions / loan_ids within an item: ascending alphabetically/loan_id.
• escalation_triggers: ascending trigger_id; peer_states: ascending state_code.
• priority_ranking: ordered application_id list, highest priority first, approved + conditionally approved only (exclude declined/deferred).

Enum discipline: use ONLY values listed in the template. Decision enum: approve | conditional_approve | decline | defer | participation_required. score_class: approve_quality | conditional | weak. CDFI risk_class: Prime | Desirable | Satisfactory | Watch | Doubtful | Projected Loss. payment_status: Current | 30 Days Past Due | 60 Days Past Due | 90+ Days Past Due | Nonaccrual. Reason codes (decline): capacity_limit, sector_breach, weak_dscr, high_ltv, low_fico, recent_bankruptcy, startup_risk, underwater_collateral, policy_floor_missing, documentation_gap, fdic_adverse_variance, ncua_peer_weakness. Output ONLY the JSON object — no narrative text outside it.

6. Population / exclusion rules (common misjudgments)

• "rated 3 or worse" / "current_rating 3 or worse" = current_rating >= 3 (numbers are severity; larger = worse). "6 or worse" = >= 6. Use min_current_rating.
• Rating-2 (and better) loans are OUT of regrade/adverse populations.
• NPA / noncurrent population = 90+ Days Past Due + Nonaccrual only. 30/60-day are NOT NPA.
• "30-89 day" real-estate delinquency = 30 Days Past Due + 60 Days Past Due (NOT 90+/Nonaccrual).
• Material downgrades: only final - current >= 2; ignore unchanged or improved loans.
• "migration_from_current_rating_3" = subset whose current_rating == 3 exactly (not >=3).
• Loans with all factors null/Current keep current_rating (not auto-downgraded).
• CDFI Projected Loss requires score>=19 AND ltv>1.0 — high ltv alone is just Watch/Doubtful.
• Concentration uses total LOAN exposure as denominator, not total assets.
• Grandfathered (flag=1) sectors: existing overage tolerated; new worsening needs mitigation.
• Stress results exclude loans/apps without a DSCR value.
• For competing-credit tasks the winner has the LOWER weighted score / fewer breaches / better concentration impact; the loser gets decline or defer + the matching reason codes (e.g. weak_dscr if stressed breach, high_ltv, sector_breach, fdic_adverse_variance).

7. Step-by-step SOP for a NEW task

1. Read the prompt. Note target id(s), as-of/review date, and which template file to match. Read input/payloads/answer_template.json fully (keys, enums, ordering, precision).
2. GET /api/policies and GET /api/manifest. Cache the policy blocks you need.
3. Fetch the target's data: branch detail + metrics (right quarter) + loans and/or sector-exposures and/or applications; segment endpoint for segments; the matching FDIC or NCUA benchmark.
4. Identify the population (rating threshold / two named apps / segment state) and apply exclusion rules in section 6.
5. Compute per policy: re-derive ratings (2.1), CDFI scores/classes (2.2), CRE weighted score (2.3), DSCR stress (2.4), concentration/capacity (2.5), benchmark variance (3), actions/dispositions (4). Use a small script for sums/ordering to avoid arithmetic slips.
6. Assemble the JSON to the template: correct keys, enums, list ordering, and per-field rounding (section 5). Drop nothing required; add nothing extra.
7. Self-check: every required key present? every enum value legal? lists ordered as specified? money 2dp / ratios 4dp / bps 2dp / dscr 2dp / weighted 1dp / metrics integer? ratios expressed as fractions not percents? priority_ranking excludes declines? Output pure JSON only.

8. Worked patterns (transferable, no memorized answers)

• Rating review: pop = loans with current_rating>=3; for each compute final_rating (max of dscr/ltv/delinquency factor ratings); build final_rating_exposure_totals (group whole pop by final_rating), migration_from_current_rating_3 (group the current==3 subset by final_rating with loan_ids), material_downgrades (final-current>=2), npa_benchmark (noncurrent exposure / total loans vs FDIC total_loans_noncurrent_pct), top_problem_credit (worst final_rating, then largest exposure), watch_list_action_coverage (group post-regrade follow-up loans by recommended_action).
• Watch-list stress: pop = current_rating>=6; assign CDFI class+factor_score; +200bp stress (dscr/1.18, threshold 1.0) over DSCR-available loans; workout_queue ordered by exposure desc; severe_bucket_counts grouped by (current_rating, payment_status).
• CRE compare: weighted CDFI score per app (lower=better, class cutoffs 2.0/3.0); dual stress (dscr*0.85/1.18); existing CRE concentration = CRE-loan exposure / total loans vs cre_policy_limit_pct; selected post-approval concentration adds requested amount to both numerator and denominator; FDIC metric total_real_estate_30_89_pct, branch_delinquency_ratio from metric delinquency_30_plus_pct; pick the lower-score, non-breaching, less-concentrating credit; loser -> decline/defer with reason codes; winner path likely conditional_approve / participation_required when CRE already over limit, with conditions like bank_retained_exposure_cap, committee_cre_exception, minimum_dscr_covenant_1_25.
• Allocation: capacity = lending_capacity_q1; rank/approve apps within capacity; screen each for decline reasons (weak_dscr, high_ltv, low_fico, recent_bankruptcy, startup_risk, documentation_gap); concentration_flags where post-approval sector pct > limit_pct; post_approval_concentrations per sector; priority_ranking = approved+conditional only.
• Segment posture: NCUA state metrics (integers); peer median vs US directions; capacity from quarterly_capacity vs pipeline; controls from minimum_checklist + control_issue (insurance/lien gaps -> binder verification + lien perfection); escalation triggers w/ owners; interpretation (capacity_available + weaker external risk -> continue_with_tighter_conditions, committee_message capacity_available_but_external_risk_weaker).