Self Attempt 02

SKILL — Credit-Risk / Lending-Committee API Answering SOP

Self-evolved skill for task_group_011 credit-office evaluation. Distilled from reasoning over 5 train tasks (Redwood rating migration, Lakeview allocation, Civic NC fire/EMS segment posture, Summit watch-list stress, Harbor competing CRE) against the live remote credit-office public REST API. No gold answers were used — this is the METHOD.

The environment is a shared remote credit office exposed only via a read-only JSON REST API. Every task asks for a single committee-ready JSON object matching a per-task answer_template.json (which doubles as the field/enum/precision contract). The job is always: pull the right endpoints, re-derive a few quantified credit judgments from raw loan/application/sector/benchmark fields using the policy block, then emit exact-shape JSON with the right ordering, enums, and numeric precision.

0. Constants & conventions (apply to every task)

• API base: <remote-env-url> (from environment_access.md). Always GET, JSON, no auth. Pipe through jq. Never call /api/judge (none exists for you).
• Policy version: credit_policy_v2025Q1. Benchmark versions: FDIC fdic_q4_2024, NCUA ncua_q1_2025. Seed 11011.
• Date canon: review/as-of date is 2025-03-31 (committee date).
• branch_id values are UPPERCASE (REDWOOD, LAKEVIEW, SUMMIT, HARBOR, …). Credit-union "branches" are also segment_ids (e.g. CIVIC_NC_FIRE_EMS).
• Rating scale: integer, lower = better (1 best … 8 worst observed). "Worst numeric rating" = the max integer.
• Numeric precision (from templates — obey exactly):
  • USD / money / exposure / balance → 2 decimals.
  • Ratios (concentration %, NPA ratio, variance_ratio, delinquency ratio, LTV-as-ratio outputs) → 4 decimals (these are fractions, e.g. 0.4695, NOT 46.95).
  • bps (variance_bps, policy_variance_bps) → 2 decimals, signed (positive = adverse / branch worse than benchmark).
  • DSCR (base/stressed) → 2 decimals. weighted_cdfi_score → 1 decimal.
  • factor_score, loan_count, ratings, bps-from-NCUA-rows → integer.
• Ordering (obey per-field ordering):
  • loan_id, application_id, sector, action, reason_codes, conditions, trigger_id → ascending string/alpha.
  • final_rating, current_rating → ascending integer.
  • workout_queue → descending exposure, then ascending loan_id.
  • severe_bucket_counts → ascending current_rating, then payment_status.
• Output discipline: return ONLY a single JSON object. No narrative, no markdown fences, no trailing prose. Match required_top_level_keys exactly; do not add extra keys.

1. Remote API usage SOP

Endpoint map (what each task section reads)

Calling habits

1. GET /api/policies first — it owns every threshold/formula. Re-derive from it; do not hard-code.
2. GET /api/branches/{id} + /metrics?quarter=2025Q1 to anchor the branch.
3. Pull /loans (with the right ?min_current_rating= or ?loan_type= filter) and /sector-exposures and /applications as the task needs.
4. Pull the relevant benchmark (/benchmarks/fdic/q4-2024 or /benchmarks/ncua/q1-2025).
5. For CU tasks: /credit-union-segments/{segment_id} + NCUA rows for the segment's state_code, its peer_states, and US.
6. Always pass jq to shape/verify; sum balances with jq '[.[].outstanding_balance] | add'.

2. Transferable business rules (INFERRED from policy + data)

2.1 Risk-rating re-derivation (train_001 Redwood, also feeds train_004)

From /api/policies.risk_rating:

DSCR → rating (floor): >=1.5→3, >=1.25→4, >=1.05→5, >=1.0→6, <1.0→7. LTV → rating (floor): <=0.65→3, <=0.75→4, <=0.85→5, <=1.0→6, >1.0→7. Delinquency minimums (floor by payment_status): Current→null, 30 Days Past Due→4, 60 Days Past Due→5, 90+ Days Past Due→7, Nonaccrual→8.

Dominant-factor rule: final_rating = max( available factor ratings ) where the available factors are the DSCR band, the LTV band, and the delinquency floor (Current contributes no floor). When a loan has zero available factors (Current payment AND no dscr AND no ltv), carry the current_rating forward unchanged.

• "Re-derive ratings for loans currently rated N or worse" → population = loans with current_rating >= N (use /loans?min_current_rating=N).
• downgrade_notches = final_rating - current_rating.
• Material downgrade threshold = risk_rating.material_downgrade_notches = 2 (i.e. downgrade_notches >= 2).
• migration_from_current_rating_3 buckets ONLY loans whose current_rating == 3 (NOT all regrade loans — just the current-3 cohort), grouped by final_rating, with ascending loan_id lists.
• final_rating_exposure_totals buckets the WHOLE regrade population by final_rating.

PITFALL — delinquency floors are "minimums": a loan already rated worse than its delinquency floor (e.g. a 5-rated loan that is only 30-DPD, floor 4) would, under the literal max(factors) rule, upgrade to 4. The policy text supports the pure-factor max, but business logic of a downgrade review favors clamping with current: final = max(current_rating, dscr_floor, ltv_floor, delinquency_floor). The two readings only diverge for loans rated worse than all their floors; for current-3 loans they always agree (downgrade). Decide per task; default to the literal policy max and flag upgrades explicitly. Verify migration totals don't silently drop a loan.

2.2 NPA & FDIC/NCUA variance (train_001, train_005)

NPA exposure = sum of outstanding_balance for loans with payment_status ∈ {90+ Days Past Due, Nonaccrual} (the FDIC "noncurrent" definition: 90+ DPD plus nonaccrual). This equals metrics.nonperforming_loans — use that as a cross-check (REDWOOD: loan-level Nonaccrual 1,725,000 == metrics 1,725,000).

Denominator = metrics.total_loans_outstanding (2025Q1) — NOT total_assets, NOT the sum of sector exposures, NOT deposits. (REDWOOD 15,191,701.54; Harbor 14,933,688.02; Lakeview 14,334,094.87.)

branch_npa_ratio      = branch_npa_exposure / branch_total_loans          # 4dp
fdic_benchmark_ratio  = <chosen FDIC metric value>                        # 4dp
variance_ratio        = branch_npa_ratio - fdic_benchmark_ratio           # 4dp, signed
variance_bps          = variance_ratio * 10000                            # 2dp, signed (+ = adverse)

Choosing the FDIC metric (npa_benchmark.benchmark_metric enum): match the metric's universe + delinquency band to the ratio you computed:

• total_loans_noncurrent_pct (0.0098) — broadest; default for a general bank branch whose NPA ratio is over all loans (90+/Nonaccrual).
• total_real_estate_noncurrent_pct (0.0121) — use when the branch ratio is over real-estate-secured loans only.
• construction_development_noncurrent_pct (0.0076) — C&D-heavy branch. Available FDIC values: total_loans_noncurrent_pct=0.0098, total_real_estate_noncurrent_pct=0.0121, total_real_estate_30_89_pct=0.0051, construction_development_noncurrent_pct=0.0076, construction_development_30_89_pct=0.0042. (REDWOOD NPA over all loans → total_loans_noncurrent_pct; Harbor CRE page uses total_real_estate_30_89_pct.)

PITFALL — band/universe alignment: the FDIC *_30_89_pct metrics cover the 30–89 DPD band ONLY (30 Days Past Due + 60 Days Past Due), EXCLUDING 90+ and Nonaccrual. The *_noncurrent_pct metrics cover 90+ DPD + Nonaccrual. Never compare a 30-89 branch ratio to a noncurrent benchmark or vice-versa. And never use metrics.delinquency_30_plus_pct as a stand-in — it is all-loans 30+ (includes 90+ & nonaccrual) and is a fraction of a different universe.

FDIC real-estate 30-89 ratio (train_005 concentration block):

branch_delinquency_ratio = (Σ balance of RE loans w/ payment_status in {30DPD, 60DPD})
                         / (Σ balance of RE loans)
fdic_benchmark_ratio     = 0.0051   # total_real_estate_30_89_pct
fdic_variance_ratio      = branch_delinquency_ratio - fdic_benchmark_ratio   # 4dp
fdic_variance_bps        = fdic_variance_ratio * 10000                        # 2dp, signed

"RE loans" universe = loan_type ∈ {CRE, Residential Mortgage} (FDIC total real estate = 1-4 family + C&D + CRE + multifamily). SBA/Equipment loans secured by RE are a judgment call; default to excluding them unless clearly RE-collateralized.

2.3 Capacity & concentration ceilings (train_002, train_005)

From /api/policies.capacity_concentration:

• lending_capacity_field = branches.lending_capacity_q1 — the quarterly new-lending capacity (bank-retained). Sum of approved_amount (bank-retained portion) must not exceed it; breaches → capacity_limit / participation_required / reduced_amount.
• single_sector_default_field = branches.sector_ceiling_pct, BUT branch_sector_override_table = sector_exposures → per-sector limit_pct from /sector-exposures overrides the branch default (e.g. Lakeview Healthcare 0.19 vs default 0.21; Harbor Hospitality/Office 0.29 vs default 0.24).
• allowed_mitigations = participation_required, reduced_amount, board_exception.
• Grandfathering: sector_exposures.grandfathered == 1 means existing over-ceiling exposure is grandfathered, but new approvals may not worsen that sector without mitigation. A new app in a grandfathered-over sector → sector_breach / participation_required / decline.

Concentration denominator = metrics.total_loans_outstanding (same as NPA — not total_assets).

post_approval_pct (sector) = (existing_sector_exposure + approved_amount_in_sector)
                             / total_loans_outstanding        # 4dp
over_limit = post_approval_pct > limit_pct                     # boolean
cre_concentration           = existing_cre_exposure / total_loans_outstanding   # 4dp
policy_variance_bps         = (cre_concentration - cre_policy_limit_pct) * 10000 # 2dp signed

existing_cre_exposure = Σ outstanding_balance of loans with loan_type == "CRE" (use /loans?loan_type=CRE). Do not sum sector-exposure rows for CRE — they mix loan types (e.g. Harbor "Construction" sector includes an Equipment loan).

Decision/concentration handling matrix (inferred):

• Sector post-approval pct ≤ limit → approve.
• Post-approval pct > limit, mitigable (participation/reduction/board exception) → conditional_approve + mitigation, or participation_required.
• Post-approval pct > limit, unmitigable or grandfathered sector being worsened → decline with sector_breach.

2.4 CDFI-style risk classes & +200bp watch-list DSCR stress (train_004)

From /api/policies.cdfi_factor_scores:

• Four factors only: debt_to_asset, fico, liquidity_months, ltv. DSCR is NOT a CDFI factor (it feeds the stress test separately).
• Each factor maps to a score by band (debt_to_asset/ltv: 0/2/4/6; fico: 0/1/3/5; liquidity_months: 0/1/3/5). factor_score (integer) = sum of the four present factors.
• Class by total score: Prime 0–5, Desirable 6–9, Satisfactory 10–13, Watch 14–18, Doubtful >=19, Projected Loss >=19 AND ltv>1.0.

PITFALL — null CDFI factors: most adverse loans have fico=null (and sometimes null dta/liquidity). Two readings: (a) literal/zero: null contributes 0 (factor not applicable). Tends to class severe loans as Desirable/Satisfactory — feels too lenient for a watch-list. (b) conservative/worst-tier: null scores the worst band (fico→5, dta→6, liquidity→5, ltv→6). Produces meaningful Doubtful/Projected-Loss classes for the worst credits (e.g. Summit SUM-LN-902 Nonaccrual LTV 1.18 → Projected Loss only under (b)). The policy is silent. Lean (b) worst-tier for null factors in a watch-list context (unknown risk = high risk), because Projected Loss/projected_loss=true only materializes that way; verify a couple of loans against the severity narrative.

Watch-list DSCR stress (train_004): use the LITERAL formula from /api/policies.stress:

shock_label            = "+200bp"        # policy.stress.watch_list_parallel_shock
formula                = "stressed_dscr = dscr / (1 + 0.18)"
stressed_dscr          = base_dscr / 1.18
breach_threshold       = 1.00            # policy.stress.coverage_breach_threshold
breaches_threshold     = stressed_dscr < 1.00

Only run for loans where DSCR is available (skip nulls). breach_loan_ids = ascending list of loan_ids with breaches_threshold == true.

PITFALL — "+200bp" label ≠ ÷1.02: the parallel-shock label says +200bp but the policy's coefficient is 1.18. Use the formula string, not the label. Do NOT compute dscr×(1-0.02) or dscr/1.02. (The apps/loans have null annual_debt_service/interest_rate, so you cannot re-derive DSCR from first principles anyway — the closed-form formula is authoritative.)

Workout queue (train_004): order descending exposure, then ascending loan_id. recommended_action from the action enum graded by severity (see 2.7). projected_loss (boolean) = true iff the loan's CDFI class is Projected Loss (score>=19 AND ltv>1.0). monitoring_cadence: monthly for an adverse watch-list (rating>=6 population), quarterly/semiannual only for milder books.

2.5 CRE weighted score & dual stress (train_005)

From /api/policies.cre_weighted_score:

• Weights (sum=1.0): capacity 0.45, collateral_exposure 0.36, conditions 0.11, character 0.05, capital 0.03.
• weighted_cdfi_score (1dp, lower is better) = Σ(weight × sub-score).
• score_class: approve_quality if score ≤ 2.0, conditional if ≤ 3.0, weak if > 3.0.

Mapping the 5 C's to application fields (inferred — applications lack explicit C's): capacity ← dscr, collateral_exposure ← ltv, character ← fico, capital ← debt_to_asset, conditions ← qualitative/sector-stress. Sub-score scale is not pinned: likely the risk-rating band (3–7) from the corresponding threshold table for dscr/ltv, and the CDFI fico/dta bands otherwise. Null sub-scores → 0 or worst-tier (same fork as 2.4). Because the class cutoffs (2.0/3.0) sit low, the rating-band (3–7) reading pushes most apps to weak; the CDFI-band (0–5) reading is gentler. Pick one reading and apply consistently; the relative ranking between the two competing apps is robust either way (HAR-APP-901 dscr 1.47/ltv 0.68 beats HAR-APP-902 dscr 1.32/ltv 0.76 → 901 lower score).

CRE dual stress (train_005 stress block):

formula                      = "stressed_dscr = dscr * 0.85 / (1 + 0.18)"  # policy.stress.cre_dual_stress_formula
coverage_breach_threshold    = 1.00
stressed_dscr                = base_dscr * 0.85 / 1.18                     # 2dp
breaches_threshold           = stressed_dscr < 1.00

(0.85 = 15% NOI haircut; 1.18 = the same rate-shock coefficient as watch-list.)

2.6 Decline reason codes & decision enums (train_002, train_005)

decision enum (all decision fields): approve, conditional_approve, decline, defer, participation_required.

conditions enum (train_002 per-app conditions): participation_required, reduced_amount, board_exception, sba_guaranty_required, startup_monitoring, none. (Note: participation_required is BOTH a decision and a condition.)

concentration handling enum (train_002): approve, conditional_approve, decline, participation_required, none.

reason_code enum (decline_reasons / applications_compared reason_codes / unselected_reason_codes): 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. Always sort ascending alphabetically.

Inferred reason-code triggers:

unselected_reason_codes (train_005) is restricted to the subset {sector_breach, weak_dscr, high_ltv, fdic_adverse_variance}.

2.7 Watch-list action coverage & workout queues (train_001, train_004)

recommended_action / coverage action enum (note the hyphen): monitor, watchlist, special-assets, workout, partial_chargeoff_review, legal_referral.

Inferred severity → action mapping (grade by final_rating + payment_status):

watch_list_action_coverage (train_001): population = the regrade population (loans current_rating >= target_current_rating_min), each assigned an action; the monitor bucket holds the stable ones so the coverage sums to the whole regrade population. by_action groups, ordered ascending by action, with ascending loan_id lists. top_problem_credit = the single worst credit (lowest final_rating number? no — highest final_rating, tie-break by exposure desc, then severity of payment_status); its recommended_action uses the same enum.

PITFALL — regrade population vs watch-list vs severe buckets: these are DIFFERENT populations. (a) Regrade population = current_rating >= N where N is the task's target (REDWOOD N=3, SUMMIT N=6). (b) severe_bucket_counts (train_004) groups the adverse population (rating>=6) by (current_rating, payment_status). (c) Watch-list action coverage (train_001) covers the regrade population. Don't conflate them.

2.8 Credit-union segment posture (NCUA) — train_003

Driven by /credit-union-segments/{segment_id} + /benchmarks/ncua/q1-2025 rows.

• state_metrics = the NCUA row for the segment's state_code, reported as integers exactly (delinquency_bps, loan_to_share_pct, roaa_bps, positive_net_income_pct).
• peer_comparison.peer_states = the segment's peer_states, sorted ascending (e.g. CIVIC → [SC, TN, VA]). peer_median = median of each metric across peer_states.
• nc_vs_us / nc_vs_peer_median: per-metric direction enum higher/lower/equal (NC value vs US / vs peer median).
• controls.required_checklist_gates = the segment's minimum_checklist set (CIVIC: board_authorization, equipment_invoice, public_contract_or_tax_support, proof_of_insurance, ucc_or_title_lien). Only add a template-enum gate if the segment data supports it (e.g. fleet_replacement_plan only for fleet/vehicle focus).
• controls.added_operating_controls: choose from the enum, mapped from internal_context:
  • control_issue about insurance binders → pre_close_insurance_binder_verification
  • staffing constraint / one senior underwriter → senior_underwriter_second_review + lien_perfection_prior_to_funding
  • external state delinquency > national → quarterly_state_benchmark_monitoring + monthly_segment_delinquency_watch
  • capacity pressure → committee_exception_for_capacity_overrun
• escalation_triggers: list all relevant conditions from the enum, each with an owner:
  • segment_recent_delinquency_ge_90_bps → credit_risk_manager
  • missing_insurance_or_lien_exception → operations_control_manager
  • quarterly_capacity_exceeded_or_exception_requested → lending_committee_chair
  • state_delinquency_gap_widens_25_bps → credit_risk_manager (trigger_id ascending; use T1/T2/… or numeric.)
• interpretation: derive from data —
  • capacity_status: capacity_available if quarterly_capacity not exhausted by current_outstanding trajectory; else capacity_constrained / no_capacity.
  • external_risk_status: compare segment state (state_code) vs US and vs all peer_states on delinquency_bps (primary) + others. weaker_than_national_and_peers if NC delinquency > US AND > peer median; stronger_… if lower than both; mixed_… otherwise.
  • risk_tolerance: take directly from segment's risk_tolerance field.
  • committee_message: tie to the combination — capacity_available_but_external_risk_weaker when capacity ok but state delinquency elevated; pause_until_state_metrics_recover when risk severe; routine_approval_path_supported when both clean.
  • posture: continue_with_tighter_conditions if capacity ok but external risk weaker
    • control issues (CIVIC case: NC delinq 79 > US 58 and > all peers; recent_delinquency 86 bps < 90 trigger threshold so not paused); continue_approving if all clean; temporarily_pause if external risk severe / capacity gone.

PITFALL — segment_recent_delinquency_ge_90_bps threshold: the segment's internal_context.recent_delinquency_bps (CIVIC 86) is BELOW 90, so that escalation trigger is armed but not currently breached — list it as a monitored trigger, do not report the segment as paused. Compare against 90, not against the state's 79.

3. Output field definitions & exact enums (consolidated)

All enums are closed — emit ONLY listed values. payment_status enum (used across tasks): Current, 30 Days Past Due, 60 Days Past Due, 90+ Days Past Due, Nonaccrual.

• train_001 top keys: branch_id, review_date, portfolio_regrade, npa_benchmark, material_downgrades, top_problem_credit.
  • portfolio_regrade: target_current_rating_min, target_loan_count, target_exposure, final_rating_exposure_totals (asc final_rating), migration_from_current_rating_3 (current-3 cohort, asc final_rating, loan_ids asc), watch_list_action_coverage{covered_loan_count,covered_exposure,by_action}.
  • npa_benchmark: benchmark_version, benchmark_metric, branch_npa_exposure, branch_total_loans, branch_npa_ratio, fdic_benchmark_ratio, variance_ratio, variance_bps.
  • material_downgrades: list (asc loan_id) of {loan_id,current_rating,final_rating, downgrade_notches,exposure}, downgrade_notches ≥ 2.
• train_002 top keys: branch_id, allocation, decisions, concentration_flags, decline_reasons, post_approval_concentrations.
  • allocation: lending_capacity_q1, gross_approved_amount, committed_capacity_amount, remaining_capacity, priority_ranking (approved + conditional_approve app_ids, highest priority first).
  • decisions: asc application_id; {application_id,decision,approved_amount, bank_capacity_used,conditions}; money 2dp.
  • concentration_flags: sort by sector then application_id; {sector,application_id, limit_pct,post_approval_pct,flag,handling}; pct 4dp.
  • decline_reasons: {app_id: [sorted reason_codes]}.
  • post_approval_concentrations: asc sector; {sector,exposure_after_approval, post_approval_pct,limit_pct,over_limit}.
• train_003 top keys: segment_id, posture, state_metrics, peer_comparison, controls, escalation_triggers, interpretation (enums in §2.8).
• train_004 top keys: branch_id, watch_list_summary, stress_results, workout_queue, severe_bucket_counts.
  • watch_list_summary: adverse_rating_min, adverse_loan_count, adverse_balance, risk_classes (asc loan_id; {loan_id,risk_class,factor_score}), monitoring_cadence.
  • stress_results: shock_label, breach_threshold, results (asc loan_id, DSCR-only; {loan_id,base_dscr,stressed_dscr,breaches_threshold}), breach_loan_ids (asc).
  • workout_queue: desc exposure then asc loan_id; {loan_id,exposure,risk_class, payment_status,recommended_action,projected_loss}.
  • severe_bucket_counts: asc current_rating then payment_status; {current_rating, payment_status,loan_count,exposure}.
• train_005 top keys: branch_id, applications_compared, recommended_path, stress, concentration, conditions.
  • applications_compared: asc application_id; {application_id,weighted_cdfi_score, score_class,decision,reason_codes} (score 1dp; reason_codes asc alpha).
  • recommended_path: {selected_application_id,path,unselected_application_id, unselected_disposition,unselected_reason_codes} (unselected_disposition ∈ {decline,defer}; unselected_reason_codes ⊆ {sector_breach,weak_dscr,high_ltv, fdic_adverse_variance}, asc alpha).
  • stress: formula, coverage_breach_threshold, results (asc application_id; {application_id,base_dscr,stressed_dscr,breaches_threshold}).
  • concentration: cre_policy_limit_pct, existing_cre_exposure, existing_cre_concentration, selected_post_approval_cre_concentration, selected_policy_variance_bps, fdic_benchmark_metric(=total_real_estate_30_89_pct), branch_delinquency_ratio, fdic_benchmark_ratio, fdic_variance_ratio, fdic_variance_bps.
  • conditions: list (asc alpha) from 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.

4. Anticipated misjudgments & exclusion rules (checklist)

1. Concentration/NPA denominator: use metrics.total_loans_outstanding — never total_assets, never sum of sector_exposures, never total_deposits.
2. CRE exposure: sum loan_type=="CRE" loan balances from /loans, NOT sector rows (sectors mix Equipment/SBA/CRE).
3. NPA exposure: 90+ DPD + Nonaccrual balances == metrics.nonperforming_loans (cross-check). Do not include 30/60 DPD.
4. FDIC band alignment: *_noncurrent_pct ↔ (90+ + Nonaccrual); *_30_89_pct ↔ (30 DPD + 60 DPD only). Never mix. metrics.delinquency_30_plus_pct ≠ any FDIC metric.
5. regrade vs watch-list vs severe buckets: different populations — regrade = current_rating >= task_min (REDWOOD 3, SUMMIT 6); migration_from_current_rating_3 = only the current_rating==3 cohort; severe_bucket_counts = the adverse (>=6) group.
6. delinquency floor upgrade edge: a loan rated worse than its delinquency floor can spuriously "upgrade" under literal max(factors). Decide clamp-vs-literal and verify migration totals.
7. +200bp label vs ÷1.18 formula: use the policy formula coefficient (1.18), not the label. Same for CRE dual stress (×0.85/1.18).
8. null CDFI factors: pick zero-vs-worst-tier and apply consistently; worst-tier is more sensible for a watch-list (makes Projected Loss reachable).
9. grandfathered sectors: existing over-ceiling is OK, but a NEW approval worsening it is a sector_breach — don't auto-approve into a grandfathered-over sector.
10. policy_sector override: use /sector-exposures per-sector limit_pct, not the branch default sector_ceiling_pct.
11. ascending loan_id / application_id: string sort, zero-padded already (RED-LN-001 < RED-LN-002 < RED-LN-901). Don't numeric-sort.
12. money 2dp / ratios 4dp / bps 2dp-signed: ratios are fractions (0.4695), not percentages (46.95). bps = ratio-diff × 10000.
13. special-assets has a hyphen; watchlist is one word; partial_chargeoff_review and legal_referral use underscores. Match exactly.
14. Severe-delinquency override: 90+ DPD forces rating ≥ 7, Nonaccrual ≥ 8 — these override any better DSCR/LTV band (max rule). A loan with dscr 1.59 but Nonaccrual is still ≥ 8.
15. CU segment vs branch: CIVIC_NC_FIRE_EMS is BOTH a branch_id (institution_type credit_union) and a segment_id; train_003 reads the segment endpoint + NCUA, not FDIC (its fdic_benchmark_set is empty).
16. Compare only the two named apps in train_005 (HAR-APP-901 vs HAR-APP-902) — the branch has other pending apps; ignore them for applications_compared.
17. selected vs unselected: the selected app is the stronger credit (lower weighted_cdfi_score, survives stress, cleaner sector). unselected gets decline/defer + the restricted reason-code subset.

5. Per-task SOP walkthroughs (the method, end to end)

train_001 — Redwood rating migration (branch_id REDWOOD, review_date 2025-03-31)

1. GET /branches/REDWOOD, GET /branches/REDWOOD/metrics?quarter=2025Q1 → total_loans_outstanding, nonperforming_loans.
2. GET /branches/REDWOOD/loans?min_current_rating=3 (15 loans) → regrade population.
3. GET /api/policies → re-derive each loan's final_rating = max(dscr_band, ltv_band, delinquency_floor); carry current_rating when no factors (e.g. RED-LN-003).
4. target_current_rating_min=3, target_loan_count=15, target_exposure=Σ balance.
5. final_rating_exposure_totals: bucket all 15 by final_rating (asc).
6. migration_from_current_rating_3: bucket ONLY current_rating==3 loans by final_rating (asc), with asc loan_ids.
7. material_downgrades: loans with final-current >= 2, asc loan_id.
8. npa_benchmark: branch_npa_exposure = Nonaccrual+90+ balances (RED-LN-901 1,725,000) == metrics.nonperforming_loans; metric = total_loans_noncurrent_pct (0.0098); variance_bps signed.
9. watch_list_action_coverage: assign action per §2.7 to all 15; group by_action.
10. top_problem_credit: worst final_rating (8 if any Nonaccrual) — RED-LN-901 Cedar Harbor Properties, Nonaccrual, → partial_chargeoff_review/legal_referral.

train_002 — Lakeview Q1 allocation (branch_id LAKEVIEW)

1. GET /branches/LAKEVIEW (capacity 5,900,000), /metrics?quarter=2025Q1 (total_loans_outstanding 14,334,094.87), /sector-exposures (per-sector limit_pct; Healthcare 0.19), /applications (9 apps).
2. GET /api/policies for capacity/concentration + decline rules.
3. Per app: compute post_approval sector pct, check capacity/sector/dscr/ltv/fico; assign decision + approved_amount (reduce if needed) + conditions; bank_capacity_used = bank-retained portion of approved.
4. allocation: gross_approved = Σ approved_amount; committed_capacity; remaining = capacity − committed; priority_ranking = approved+conditional app_ids by priority.
5. decline_reasons: map each declined app → sorted reason codes.
6. concentration_flags: per (sector, app) where the app touches a sector, post_approval_pct vs limit_pct, handling.
7. post_approval_concentrations: per sector after all approvals, over_limit bool.

train_003 — Civic NC Fire/EMS segment posture (segment_id CIVIC_NC_FIRE_EMS)

1. GET /credit-union-segments/CIVIC_NC_FIRE_EMS (peer_states SC/TN/VA, risk_tolerance moderate, quarterly_capacity 2,900,000, current_outstanding 16,850,000, recent_delinquency_bps 86, control_issue, staffing_constraint).
2. GET /benchmarks/ncua/q1-2025 → NC row (delinq 79, lts 76, roaa 44, pni 76), US (delinq 58, lts 69, roaa 65, pni 85), peers (SC 72/73/51/79, TN 64/71/59/81, VA 53/67/65/85).
3. state_metrics = NC integers. peer_comparison: peer_states asc; nc_vs_us and nc_vs_peer_median directions (NC delinq higher, roaa/pni lower, lts higher).
4. controls: required = segment.minimum_checklist; added from internal_context (insurance binder → pre_close_insurance_binder_verification; staffing → senior_underwriter_second_review; external delinq → quarterly_state_benchmark_monitoring
   • monthly_segment_delinquency_watch; lien → lien_perfection_prior_to_funding).
5. escalation_triggers: 4 conditions with owners (asc trigger_id).
6. interpretation: capacity_available + weaker_than_national_and_peers (NC delinq exceeds US and all peers) + moderate → committee_message capacity_available_but_external_risk_weaker; posture continue_with_tighter_conditions (recent_delinquency 86 < 90, so not paused).

train_004 — Summit watch-list stress (branch_id SUMMIT, adverse_rating_min 6)

1. GET /branches/SUMMIT/loans?min_current_rating=6 (7 loans) → adverse population; adverse_balance=Σ balance.
2. GET /api/policies → CDFI factor bands + watch_list stress formula.
3. risk_classes: per loan, factor_score from dta/fico/liquidity/ltv (handle nulls — lean worst-tier); class by score range; asc loan_id.
4. stress_results: for loans with dscr (skip SUM-LN-010 null), stressed = dscr/1.18; breach < 1.00 (SUM-LN-004/011/901/902 breach; 003/015 don't); breach_loan_ids asc.
5. workout_queue: desc exposure then asc loan_id; recommended_action by severity; projected_loss = (class==Projected Loss).
6. severe_bucket_counts: group 7 loans by (current_rating, payment_status), asc.
7. monitoring_cadence: monthly (adverse watch-list).

train_005 — Harbor competing CRE (branch_id HARBOR; apps HAR-APP-901, HAR-APP-902)

1. GET /branches/HARBOR (cre_policy_limit_pct 0.29), /metrics?quarter=2025Q1 (total_loans_outstanding 14,933,688.02, nonperforming 0), /applications (pick only 901/902), /loans?loan_type=CRE (existing CRE 7,011,570.24), /sector-exposures (Hospitality grandfathered=1, limit 0.29).
2. GET /benchmarks/fdic/q4-2024 (total_real_estate_30_89_pct 0.0051).
3. applications_compared: weighted_cdfi_score per app (capacity←dscr, collateral←ltv; lower=better; 901 < 902); score_class; decision; reason_codes asc.
4. recommended_path: selected = HAR-APP-901 (stronger dscr/ltv, new Industrial CRE sector, survives stress); path approve/conditional_approve; unselected = HAR-APP-902 (decline; reason_codes ⊆ {sector_breach(Hospitality grandfathered), weak_dscr(stressed <1.0), fdic_adverse_variance} asc).
5. stress: cre_dual formula dscr×0.85/1.18; 901→1.06 (no breach), 902→0.95 (breach); threshold 1.00.
6. concentration: existing_cre_concentration 0.4695 (already > 0.29); selected_post_approval_cre_concentration = (7,011,570.24+2,100,000)/14,933,688.02; selected_policy_variance_bps signed; branch_delinquency_ratio = RE 30-89 / RE total (loan-level); fdic_variance_bps signed (large adverse).
7. conditions: asc alpha from the enum (e.g. bank_retained_exposure_cap, minimum_dscr_covenant_1_25, no_additional_cre_without_committee_review, quarterly_financial_reporting, updated_appraisal_before_close, tenant_roll_and_lease_review …).

6. Quick reference — pinned values observed

• REDWOOD: total_loans_outstanding 15,191,701.54; NPA 1,725,000 (RED-LN-901 Nonaccrual); regrade pop 15/18 loans.
• LAKEVIEW: total_loans_outstanding 14,334,094.87; capacity 5,900,000; Healthcare limit 0.19.
• CIVIC: quarterly_capacity 2,900,000; current_outstanding 16,850,000; recent_delinquency_bps 86 (<90); NC delinq 79 > US 58 >? no, 79>58 and > all peers.
• SUMMIT: adverse pop 7 loans (rating 6-8); SUM-LN-902 Nonaccrual LTV 1.18 (worst); SUM-LN-010 has no dscr (excluded from stress).
• HARBOR: total_loans_outstanding 14,933,688.02; existing CRE 7,011,570.24 (46.95% > 29%); nonperforming 0; Hospitality grandfathered; HAR-APP-901 selected, HAR-APP-902 declined.
• FDIC q4-2024: noncurrent total 0.0098 / RE 0.0121 / C&D 0.0076; 30-89 RE 0.0051 / C&D 0.0042.
• Policy: material_downgrade_notches 2; coverage_breach_threshold 1.0; watch_list_formula dscr/1.18; cre_dual dscr×0.85/1.18; CRE weights cap 0.45 / coll 0.36 / cond 0.11 / char 0.05 / cap'l 0.03.