Self Attempt 03

SKILL — Credit-Risk / Lending-Committee Decision Packets (task_group_011)

A self-contained SOP for producing committee-ready JSON answers against the shared credit-office public REST API. Distilled from reasoning through five training archetypes (Redwood rating-migration, Lakeview allocation, Civic CU segment posture, Summit watch-list stress, Harbor competing-CRE). No gold answers used — this is the method, not the answers.

0. Environment & ground rules

• Base URL: <remote-env-url> (remote; no local env/ source, no DB).
• Read-only: GET only. No auth. Never call /api/judge.
• Always JSON. Pipe through jq. Health: GET /api/health. Catalog: GET /api/manifest.
• Policy version: credit_policy_v2025Q1 (fetch GET /api/policies once — it holds every numeric rule below). Benchmark versions: FDIC fdic_q4_2024, NCUA ncua_q1_2025.
• Generated/committee date = 2025-03-31 (use as review_date / packet date).
• Every answer must conform to that task's input/payloads/answer_template.json. Enums and orderings there are strict — do not invent values.

1. Remote API usage SOP — endpoints, keys, query params

Query-param semantics (critical)

• ?min_current_rating=N → returns loans with current_rating >= N (rating scale: 3 = best, 8 = worst; higher = worse). This is an inclusive lower bound on a "worse-or-equal" scale.
• ?loan_type= exact match (e.g. CRE, C%26I → URL-encode & as %26).
• ?payment_status= exact match (spaces → %20).
• ?state_code= 2-letter, returns rows[] with one element.

Which endpoint feeds which answer section

• Redwood (001): /branches/REDWOOD + /branches/REDWOOD/loans?min_current_rating=3 + /metrics[0] + /policies + /benchmarks/fdic/q4-2024.
• Lakeview (002): /branches/LAKEVIEW + /applications + /sector-exposures + /metrics[0] + /policies.
• Civic (003): /credit-union-segments/CIVIC_NC_FIRE_EMS + /benchmarks/ncua/q1-2025 (all rows + ?state_code=NC) + /policies.
• Summit (004): /branches/SUMMIT + /loans?min_current_rating=6 + /metrics[0] + /policies.
• Harbor (005): /branches/HARBOR + /applications?loan_type=CRE + /sector-exposures + /loans + /metrics[0] + /policies + /benchmarks/fdic/q4-2024.

2. Universal business rules (from GET /api/policies)

2.1 Risk-rating re-derivation (risk_rating)

Final re-derived rating = worst (highest numeric) of the available factor-derived ratings. Rating scale 3 (best) → 8 (worst).

• DSCR → rating: dscr >= 1.5 → 3; >= 1.25 → 4; >= 1.05 → 5; >= 1.0 → 6; < 1.0 → 7.
• LTV → rating: ltv <= 0.65 → 3; <= 0.75 → 4; <= 0.85 → 5; <= 1.0 → 6; > 1.0 → 7.
• Delinquency minimum (hard floor): Current → none; 30 DPD → 4; 60 DPD → 5; 90+ DPD → 7; Nonaccrual → 8.
• Dominant-factor rule: final_rating = max( current_rating, dscr_rating(if DSCR), ltv_rating(if LTV), delinquency_floor(if status has one) ). Treat current_rating itself as a floor so re-derivation only downgrades or holds, never spuriously upgrades (a Current loan with strong DSCR stays; a 30-DPD loan at 5 is not upgraded to 4 just because the delinquency floor is 4). When DSCR/LTV are null, that factor is simply omitted.
• Severe-delinquency override: 90+ DPD → 7 and Nonaccrual → 8 are mandatory floors that override even a strong DSCR (e.g. a Nonaccrual loan with DSCR 1.55 lands at 8, not 5).
• Material downgrade: final_rating - current_rating >= risk_rating.material_downgrade_notches (= 2). Only these go in material_downgrades.

2.2 Stress formulas (stress)

• coverage_breach_threshold = 1.0 (breach when stressed_dscr < 1.0).
• Watch-list (Summit): stressed_dscr = dscr / (1 + 0.18) → dscr / 1.18. Shock label +200bp.
• CRE dual stress (Harbor): stressed_dscr = dscr * 0.85 / (1 + 0.18) → dscr * 0.72034.
• Only loans/apps with a non-null DSCR appear in stress_results. Loans with null DSCR are excluded from the stress list (but stay in watch-list/workout). breaches_threshold is boolean.
• breach_loan_ids = the subset where stressed_dscr < 1.0, ascending loan_id.

2.3 CDFI factor scoring & risk classes (cdfi_factor_scores)

Sum available factor subscores (each 0–6) → factor_score (integer). Skip null factors (do not add a max penalty for missing data).

risk_class from factor_score:

• Prime 0–5 · Desirable 6–9 · Satisfactory 10–13 · Watch 14–18
• Doubtful >=19 · Projected Loss >=19 AND ltv > 1.0.

projected_loss (workout_queue boolean) = true iff the loan's risk_class == "Projected Loss" (i.e. factor_score >= 19 and ltv > 1.0). Apply the rule literally even for Nonaccrual loans that don't meet both conditions.

2.4 CRE weighted score (cre_weighted_score) — Harbor

Five-C weights: capacity 0.45, collateral_exposure 0.36, conditions 0.11, character 0.05, capital 0.03 (sum 1.0). Map app/loan factors to the C's (collateral_exposure←ltv, character←fico, capital←debt_to_asset, capacity/conditions←DSCR/sector-specific). Multiply each C's factor score by its weight and sum → weighted_cdfi_score (1 dp, lower is better).

• score_class: approve_quality if <= 2.0; conditional if <= 3.0; weak if > 3.0.
• The app with the lower weighted score (better DSCR + lower LTV) is the stronger credit.

2.5 Capacity & concentration (capacity_concentration)

• Q1 lending capacity = branches.lending_capacity_q1 (USD).
• Single-sector default ceiling = branches.sector_ceiling_pct (0.21–0.24). Per-sector overrides live in /sector-exposures (limit_pct); CRE-related sectors carry limit_pct = cre_policy_limit_pct (0.29 at HARBOR).
• CRE policy limit = branches.cre_policy_limit_pct (the field name — not cre_limit_pct).
• Concentration denominator = /metrics[0].total_loans_outstanding (NOT total_assets, NOT branch sum of balances). Critical pitfall.
• post_approval_pct = (sector current_exposure + approved app amount in that sector) / total_loans_outstanding, 4 dp.
• over_limit / sector flag = post_approval_pct > limit_pct.
• Grandfathering: existing over-ceiling exposure may stay (grandfathered=1), but new approvals may not worsen that sector without a mitigation. Allowed mitigations: participation_required, reduced_amount, board_exception. Non-mitigated breaches → decline or participation_required.

2.6 Decline reason codes & decision enums

Decline reason-code enum (12): 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. Sort ascending alphabetically.

Decision enum (5): approve, conditional_approve, decline, defer, participation_required.

Task-002 conditions enum (6): participation_required, reduced_amount, board_exception, sba_guaranty_required, startup_monitoring, none. (Use none for clean approvals.)

Decision heuristics (Lakeview): approve if DSCR≥1.25, LTV≤0.85, fico≥680, no sector breach, documentation_complete, within capacity; conditional_approve / reduced_amount when capacity-tight or LTV high but creditable; decline for DSCR<1.0, LTV>1.0, recent bankruptcy, sector breach w/o mitigation, documentation gap; defer for missing documentation that is curable; participation_required / sba_guaranty_required for SBA loans or capacity overruns.

2.7 Watch-list action coverage & workout queue (action enum, 6)

monitor, watchlist, special-assets, workout, partial_chargeoff_review, legal_referral.

Action mapping by final risk:

• Current / held rating, factor Watch-or-better → monitor or watchlist.
• Substandard (final 6) → watchlist/special-assets.
• 90+ DPD / final 7 → special-assets/workout.
• Nonaccrual / final 8 → legal_referral (or partial_chargeoff_review if projected loss).

Workout queue (workout_queue) ordering: descending exposure, then ascending loan_id. Watch-list action coverage (watch_list_action_coverage) is grouped by_action ascending by action, each with loan_count, exposure, ascending loan_ids.

3. Field-name & data-source gotchas (verified against live data)

1. /metrics is a list — index [0] for the current quarter. Field is total_loans_outstanding, branch record's same-named field is null.
2. Branch CRE field = cre_policy_limit_pct (0.29 Harbor). Do not query cre_limit_pct.
3. /sector-exposures balance field = current_exposure; ceiling field = limit_pct (per-sector, may equal sector_ceiling_pct or cre_policy_limit_pct); grandfathered is 0/1.
4. Loans/applications carry many nullable factor fields (dscr, ltv, fico, debt_to_asset, liquidity_months). Always null-check before scoring; null DSCR ⇒ exclude from stress list.
5. loan_type C&I must be URL-encoded C%26I when passed as a query param.
6. NCUA endpoint always returns {benchmark_version, rows:[...]} — even ?state_code=NC gives a 1-element rows[]. Do not assume a flat object.
7. Rating scale direction: higher current_rating = worse. min_current_rating=3 = "3 or worse" (the whole regrade review set); min_current_rating=6 = "adverse/watch-list" (Summit).
8. nonperforming_loans (NPA exposure) is a USD field in /metrics; it is not a count. nonperforming_loans at REDWOOD (1,725,000.00) equals the Nonaccrual loan balance — confirms NPA exposure comes from the metrics row, not a hand recount.

4. Per-task output field map & orderings

Task 001 — Redwood rating migration (review_date = 2025-03-31)

Top-level: branch_id, review_date, portfolio_regrade, npa_benchmark, material_downgrades, top_problem_credit.

portfolio_regrade:

• target_current_rating_min = 3 (the min_current_rating used).
• target_loan_count / target_exposure = count and 2-dp USD sum of the rating>=3 population.
• final_rating_exposure_totals: all regrade loans grouped by final_rating (asc), with loan_count, exposure.
• migration_from_current_rating_3: only loans whose current_rating == 3 (literal 3 — subset, not the whole population!), grouped by final_rating (asc), each row {final_rating, loan_count, exposure, loan_ids(asc)}. Common trap: do not use the whole >=3 population here.
• watch_list_action_coverage: {covered_loan_count, covered_exposure, by_action[]} — the downgraded loans needing follow-up, bucketed by the §2.7 action enum (asc by action).

npa_benchmark:

• benchmark_version = "fdic_q4_2024".
• benchmark_metric ∈ {total_loans_noncurrent_pct, total_real_estate_noncurrent_pct, construction_development_noncurrent_pct}. For Redwood (general bank) use total_loans_noncurrent_pct = 0.0098.
• branch_npa_exposure = metrics[0].nonperforming_loans; branch_total_loans = metrics[0].total_loans_outstanding.
• branch_npa_ratio = npa_exposure / total_loans (4dp); fdic_benchmark_ratio = 0.0098 (4dp).
• variance_ratio = branch_npa_ratio − fdic_benchmark_ratio (4dp, signed).
• variance_bps = variance_ratio × 10000 (2dp, signed; positive = adverse).

material_downgrades: loans where final_rating − current_rating >= 2, asc by loan_id; each {loan_id, current_rating, final_rating, downgrade_notches, exposure}.

top_problem_credit: the single worst credit — pick by severe delinquency (Nonaccrual/90+), then largest exposure, then largest downgrade. {loan_id, borrower_name, exposure, current_rating, final_rating, payment_status(enum), recommended_action(enum)}. (REDWOOD: the Nonaccrual CRE loan whose balance equals nonperforming_loans is the natural pick.)

Task 002 — Lakeview allocation

Top-level: branch_id, allocation, decisions, concentration_flags, decline_reasons, post_approval_concentrations.

allocation: lending_capacity_q1 (branch), gross_approved_amount (sum of approve+conditional approved_amounts), committed_capacity_amount, remaining_capacity = capacity − committed, priority_ranking (application_ids highest priority first, approve+conditional only).

decisions: asc by application_id; each {application_id, decision(enum), approved_amount(2dp), bank_capacity_used(2dp), conditions(enum)}. Declined apps → approved_amount 0.00; bank_capacity_used 0.00.

concentration_flags: asc by (sector, application_id); each {sector, application_id, limit_pct(4dp), post_approval_pct(4dp), flag, handling(enum)}. handling: approve/conditional_approve/decline/participation_required/none.

decline_reasons: object mapping each declined application_id → sorted list of reason codes (§2.6 enum).

post_approval_concentrations: asc by sector; {sector, exposure_after_approval(2dp), post_approval_pct(4dp), limit_pct(4dp), over_limit(bool)}. Denominator = total_loans_outstanding.

Task 003 — Civic CU segment posture

Use NCUA rows for NC, US, and peer states (segment.peer_states = [SC, TN, VA]).

state_metrics: state_code="NC", benchmark_version="ncua_q1_2025", and the four NC integer values exactly as reported (delinquency_bps=79, loan_to_share_pct=76, roaa_bps=44, positive_net_income_pct=76).

peer_comparison: peer_states asc ([SC,TN,VA]); nc_vs_us and nc_vs_peer_median each = {delinquency_bps, loan_to_share_pct, roaa_bps, positive_net_income_pct} with values higher|lower|equal (direction of NC vs comparator). Peer median = median of the 3 peer-state values per metric.

• NC vs US: delinquency higher, loan_to_share higher, roaa lower, positive_net_income lower.
• NC vs peer median: delinquency higher, loan_to_share higher, roaa lower, positive_net_income lower.

controls.required_checklist_gates = segment's minimum_checklist (board_authorization, equipment_invoice, public_contract_or_tax_support, proof_of_insurance, ucc_or_title_lien — note fleet_replacement_plan & payer_contract_summary are allowed by the template but NOT in this segment's checklist, so omit them). added_operating_controls chosen from the template set to address the segment's internal_context (insurance-binder control_issue → pre_close_insurance_binder_verification; staffing_constraint → senior_underwriter_second_review; external risk → quarterly_state_benchmark_monitoring, monthly_segment_delinquency_watch, lien_perfection_prior_to_funding).

escalation_triggers: asc by trigger_id; each {trigger_id, condition, owner} using the condition & owner enums. Standing set of 4 triggers with owners among credit_risk_manager / operations_control_manager / lending_committee_chair (segment_recent_delinquency_ge_90_bps → note NC bps 79 & segment recent 86 are both < 90, so it is a standing trigger, not currently breached).

interpretation: capacity_status=capacity_available (segment quarterly_capacity 2.9M, notes say available); external_risk_status=weaker_than_national_and_peers; risk_tolerance=moderate (from segment); committee_message=capacity_available_but_external_risk_weaker. posture=continue_with_tighter_conditions.

Task 004 — Summit watch-list stress

Top-level: branch_id, watch_list_summary, stress_results, workout_queue, severe_bucket_counts.

watch_list_summary:

• adverse_rating_min = 6 (the requested threshold; adverse = current_rating >= 6).
• adverse_loan_count, adverse_balance (2dp) over the rating>=6 set.
• risk_classes: asc by loan_id; each {loan_id, risk_class(enum §2.3), factor_score}.
• monitoring_cadence ∈ {monthly, quarterly, semiannual} — pick by worst risk class present (Nonaccrual/Projected-Loss → monthly; otherwise quarterly).

stress_results: shock_label ("+200bp"/"watch_list_+200bp"), breach_threshold=1.0, results asc by loan_id for loans with DSCR only: {loan_id, base_dscr(2dp), stressed_dscr(2dp), breaches_threshold(bool)} using stressed_dscr = dscr/1.18. breach_loan_ids asc.

workout_queue: asc/desc by exposure desc, then loan_id asc: {loan_id, exposure(2dp), risk_class, payment_status(enum), recommended_action(enum §2.7), projected_loss(bool)}. Includes all adverse loans (even those without DSCR).

severe_bucket_counts: asc by (current_rating, payment_status); each {current_rating, payment_status, loan_count, exposure} summarizing the adverse population.

Task 005 — Harbor competing-CRE

applications_compared = the two competing CRE apps (the larger 9xx-series sentinel apps, e.g. HAR-APP-901 / HAR-APP-902), asc by application_id. Each {application_id, weighted_cdfi_score(1dp), score_class(enum), decision(enum), reason_codes(sorted)}.

recommended_path: {selected_application_id, path(enum), unselected_application_id, unselected_disposition ∈ {decline, defer}, unselected_reason_codes(sorted, restricted to sector_breach/weak_dscr/high_ltv/fdic_adverse_variance)}. Select = lower weighted score (survives the dual stress stressed_dscr = dscr*0.85/1.18 >= 1.0); the unselected typically breaches stress and/or sector.

stress: formula ("cre_dual_stress"/"stressed_dscr = dscr * 0.85 / (1 + 0.18)"), coverage_breach_threshold=1.0, results asc by application_id: {application_id, base_dscr(2dp), stressed_dscr(2dp), breaches_threshold(bool)}.

concentration: cre_policy_limit_pct(4dp = branch.cre_policy_limit_pct), existing_cre_exposure(2dp = sum of CRE-sector current_exposure; CRE sectors are those whose limit_pct == cre_policy_limit_pct), existing_cre_concentration(4dp = exposure/total_loans_outstanding), selected_post_approval_cre_concentration(4dp = (existing + selected requested_amount)/total_loans_outstanding), selected_policy_variance_bps(2dp signed = (post − limit)×10000), fdic_benchmark_metric=total_real_estate_30_89_pct, branch_delinquency_ratio(4dp, branch analog of that metric — derive from metrics/loans), fdic_benchmark_ratio=0.0051(4dp), fdic_variance_ratio(4dp signed = branch − benchmark), fdic_variance_bps(2dp signed = ratio×10000).

conditions: subset of the 7 allowed enum values, asc alphabetically (bank_retained_exposure_cap, committee_cre_exception, updated_appraisal_before_close, tenant_roll_and_lease_review, minimum_dscr_covenant_1_25, quarterly_financial_reporting, no_additional_cre_without_committee_review).

5. Numeric conventions

• Money / USD: round to 2 decimals. Sum balances/exposures with 2-dp rounding at the end.
• Ratios / concentrations / percentages expressed as ratios: 4 decimals (e.g. 0.1135).
• Basis-point variances: 2 decimals, signed (positive = adverse / over-limit). variance_bps = (branch_ratio − benchmark_ratio) × 10000.
• DSCR (base & stressed): 2 decimals. weighted_cdfi_score: 1 decimal (lower better).
• Counts, ratings, factor_score: integers.
• breach / over_limit / projected_loss: JSON booleans (true/false).
• Always honour the precision stated in each template field — it is checked.

6. Anticipated misjudgments & exclusion rules

1. Regrade population vs watch-list. Redwood regrades current_rating >= 3 (min_current_rating=3); Summit watch-list is current_rating >= 6. Don't confuse the two thresholds. Within Redwood, target_current_rating_min=3 covers almost the whole book; the migration block is the current_rating == 3 subset only.
2. Severe-delinquency override. 90+ DPD forces ≥7, Nonaccrual forces 8, even if DSCR is strong. Don't let a good DSCR "rescue" a Nonaccrual.
3. No spurious upgrades. final = max(current, factors) — re-derivation only downgrades/holds. Do not upgrade a 5 to 4 because only the 30-DPD floor (4) is available.
4. Concentration denominator. Always total_loans_outstanding from /metrics[0]. Never total_assets (the branch record's, ~5× larger) — that would shrink every concentration % and hide breaches.
5. NPA exposure source. Use metrics[0].nonperforming_loans (USD), not a manual count, not total_assets.
6. Stress-list DSCR filter. Loans with null DSCR are excluded from stress_results but still appear in workout_queue / watch_list_summary. Don't drop them from the portfolio.
7. final_rating_exposure_totals vs migration_from_current_rating_3. Former = all rating>=3 loans by final_rating; latter = only current_rating==3 loans by final_rating (and includes loan_ids).
8. Orderings. ascending loan_id is string sort ("RED-LN-001" < "RED-LN-011" < "RED-LN-015" < "RED-LN-901"). workout_queue is exposure-desc then loan_id-asc. applications_compared, material_downgrades, risk_classes, breach_loan_ids, severe_bucket_counts, post_approval_concentrations — each has its own stated ordering; follow the template literally.
9. Reason-code lists sorted ascending alphabetically; unselected_reason_codes is restricted to the 4-value subset, not the full 12.
10. CRE sectors. Identify CRE sectors as those whose /sector-exposures limit_pct equals cre_policy_limit_pct (0.29), not by name heuristics. Respect grandfathered — don't worsen a grandfathered over-ceiling sector without a mitigation.
11. NCUA shape. Always .rows[]; a ?state_code= call returns a 1-element rows array, not a flat object.
12. benchmark_metric choice. Match the FDIC metric to the branch/sector mix: general bank NPA review → total_loans_noncurrent_pct; CRE delinquency variance (Harbor) → total_real_estate_30_89_pct; construction-heavy → construction metrics. The template enum pins Harbor to total_real_estate_30_89_pct.
13. Two competing apps, not all CRE apps. Harbor has several CRE-typed applications; the "competing" pair is the two large 9xx-series sentinel apps. Confirm via amounts / the prompt's "two CRE requests".
14. score_class thresholds use <= (approve_quality ≤2.0, conditional ≤3.0, weak >3.0). Lower weighted score is better; pick the lower score as selected.
15. recent_bankruptcy trigger: years_in_business small + bankruptcy_months_ago present → recent_bankruptcy/startup_risk. low_fico only when fico is non-null and below band. documentation_gap only when documentation_complete is false.

7. Step-by-step derivation SOP (apply per task)

1. Read the prompt → identify branch_id / segment_id, the threshold param (min_current_rating), and which benchmark (FDIC vs NCUA) applies. Open the answer_template; list every required key, enum, and ordering.
2. Fetch /api/policies once. Fetch /api/branches/{id}, /api/branches/{id}/metrics (take [0]), the relevant loans/applications/sector-exposures with the right ? filter, and the benchmark.
3. Derive per-field using §2 rules. Compute factor-derived ratings, stress DSCRs, factors scores, concentrations, and variances with the exact formulas and precisions in §5.
4. Filter & group with the exact subset rules (regrade pop vs migration subset; DSCR-only stress list; adverse rating>=6; decline-only reason map; competing-pair apps).
5. Sort each list to the template's ordering (string asc loan_id; exposure-desc; alpha asc reason codes; asc by final_rating / sector / trigger_id / application_id).
6. Round to the field's precision (money 2dp, ratios 4dp, bps 2dp signed, scores 1dp).
7. Emit a single JSON object with exactly the required top-level keys and per-item keys; omit extras.
8. Self-check against §6 pitfalls before finishing: denominator? NPA source? override applied? no upgrade? enum exact? ordering literal? precision honoured?