Fewshot Attempt 02

Skill — Credit-Risk / Lending-Committee Answer Generation (task_group_011)

Transferable method for producing committee-ready JSON answers against the shared credit-office public REST API. Distilled from five fully-worked train tasks (Redwood regrade, Lakeview allocation, Civic NC segment posture, Summit watch-list, Harbor competing-CRE) and cross-checked against live API data.

0. Environment & ground rules

• Base URL: <remote-env-url>. Read-only GET only. No auth.
• Always start by reading /api/policies (the single source of credit rules: risk-rating thresholds, CDFI factor scores, CRE weighted score, stress formulas, capacity/concentration policy) and /api/manifest (versions).
• Never call /api/judge (no test-time judge; it is not part of the public surface). Do not read any local env/ source.
• Pipe responses through jq to inspect and project fields.
• Latest quarter of branch metrics is 2025Q1 (the second array element on these branches is 2024Q4). Use the latest quarter unless a prompt names a quarter. review_date / as-of date in prompts is 2025-03-31.
• Policy version is credit_policy_v2025Q1. FDIC set fdic_q4_2024; NCUA ncua_q1_2025.

1. API SOP — endpoints and how each answer section uses them

branch_id values are uppercase (REDWOOD, LAKEVIEW, SUMMIT, HARBOR, uppercase bank branch ids plus credit-union segment ids, which also appear in /api/branches with institution_type=credit_union.

2. Cross-cutting business rules

2.1 Risk-rating re-derivation (dominant-factor / worst-notch rule)

Policy risk_rating.dominant_factor_rule: the final re-derived rating is the worst (highest) numeric rating from the available DSCR, LTV/collateral, and delinquency factors. Compute each factor's rating, then take the max.

DSCR thresholds (risk_rating.dscr_thresholds, min is inclusive lower bound):

• dscr >= 1.5 → 3
• dscr >= 1.25 → 4
• dscr >= 1.05 → 5
• dscr >= 1.0 → 6
• dscr < 1.0 → 7
• dscr null/missing → factor contributes nothing (skip it).

LTV thresholds (risk_rating.ltv_thresholds, max inclusive upper bound):

• ltv <= 0.65 → 3
• ltv <= 0.75 → 4
• ltv <= 0.85 → 5
• ltv <= 1.0 → 6
• ltv > 1.0 → 7

Delinquency minimums (risk_rating.delinquency_minimums) — a hard floor:

• Current → no floor (null)
• 30 Days Past Due → at least 4
• 60 Days Past Due → at least 5
• 90+ Days Past Due → at least 7
• Nonaccrual → at least 8

final_rating = max(dscr_rating*, ltv_rating*, delinquency_floor*). Example verified on REDWOOD: RED-LN-002 (dscr 1.49→4, ltv 1.0219→7, 60 DPD→5) ⇒ final 7 (current 4, downgrade 3).

Population filter: regrade population = loans with current_rating >= target_current_rating_min (e.g. Redwood >=3). Pull with /api/branches/{id}/loans?min_current_rating=3. This is NOT the watch-list follow-up population (see 2.7).

Material downgrade: material_downgrade_notches = 2. A loan is a material downgrade iff current_rating - final_rating <= -2 (i.e. worsened by ≥2 notches). Only downgrades (final > current) by ≥2 notches are listed; list is ordered ascending by loan_id.

2.2 NPA & FDIC/NCUA variance

NPA variance (bank branches): use branch metrics latest quarter.

• branch_npa_exposure = metrics.nonperforming_loans (cross-check: sum of outstanding_balance where payment_status == "Nonaccrual"; the two agree).
• branch_total_loans = metrics.total_loans_outstanding.
• branch_npa_ratio = round(branch_npa_exposure / branch_total_loans, 4).
• fdic_benchmark_ratio = the chosen FDIC metric value (a decimal ratio, e.g. 0.0098).
• variance_ratio = round(npa_exposure/total_loans - fdic_ratio, 4) (positive = branch worse than benchmark).
• variance_bps = round((npa_exposure/total_loans - fdic_ratio) * 10000, 2), computed from the unrounded ratio (Redwood: 1037.49, not 1037.00).
• benchmark_version = "fdic_q4_2024".

FDIC metric choice (npa_benchmark.benchmark_metric enum): total_loans_noncurrent_pct, total_real_estate_noncurrent_pct, construction_development_noncurrent_pct. Choose by branch/portfolio focus: general branch NPA review → total_loans_noncurrent_pct; CRE-heavy → total_real_estate_noncurrent_pct; construction-heavy → construction_development_noncurrent_pct. The benchmark value comes from the matching field of /api/benchmarks/fdic/q4-2024.

CRE/FDIC delinquency variance (Harbor-style CRE decision): here the variance pairs the branch delinquency_30_plus_pct (from metrics) against the FDIC total_real_estate_30_89_pct benchmark (template-hardcoded for that task).

• branch_delinquency_ratio = metrics.delinquency_30_plus_pct (e.g. 0.2853).
• fdic_benchmark_ratio = FDIC total_real_estate_30_89_pct (e.g. 0.0051).
• fdic_variance_ratio = round(branch - fdic, 4); fdic_variance_bps = round((branch - fdic) * 10000, 2) (e.g. 2802.0). Sign: positive = branch under-performs benchmark.

NCUA peer comparison (credit-union segment tasks):

• Fetch /api/benchmarks/ncua/q1-2025 (all rows, includes US).
• state_metrics = the segment state_code row; benchmark_version = "ncua_q1_2025". Values are integers exactly as reported (delinquency_bps, loan_to_share_pct, roaa_bps, positive_net_income_pct).
• Peer states come from the segment JSON peer_states (e.g. ["SC","TN","VA"]), already ascending; do not invent peers. Compute the median of the peer states' values per metric.
• nc_vs_us and nc_vs_peer_median: direction per metric ∈ {"higher","lower","equal"} (substitute the segment's state code for "nc"). Example: NC delinquency_bps 79 vs US 58 → higher; NC roaa_bps 44 vs peer median 59 → lower.

2.3 Capacity & concentration ceilings

Lending capacity (from branches.lending_capacity_q1):

• gross_approved_amount = Σ approved_amount over all approve + conditional_approve decisions (full approved principal, including amounts that are later participated / SBA-guaranteed).
• committed_capacity_amount = Σ bank_capacity_used over the same decisions (the bank's at-risk retained exposure, after mitigation).
• remaining_capacity = lending_capacity_q1 - committed_capacity_amount.

bank_capacity_used by decision type:

• approve → approved_amount (full retention).
• conditional_approve + sba_guaranty_required → approved_amount * (1 - sba_guaranty_pct). Verified: LAK-APP-902, approved 840,000 × (1 − 0.75) = 210,000.
• conditional_approve + participation_required → bank-retained amount sized so the affected sector sits at its limit_pct against the pro-forma book (total_loans_outstanding + committed bank-retained capacity); the excess is participated out. Verified: LAK-APP-901 retained 1,508,113.31 = 0.19 × (total_loans_outstanding + committed_capacity) − existing_healthcare.

Sector concentration limits: limit_pct comes from each sector-exposures row (per-sector override); default is branches.sector_ceiling_pct when a sector is not listed. CRE-wide ceiling = branches.cre_policy_limit_pct. Mitigations allowed by policy: participation_required, reduced_amount, board_exception. Grandfathering: existing over-ceiling exposure may be grandfathered, but a new approval may not worsen that sector without mitigation.

Existing CRE exposure (Harbor-style): existing_cre_exposure = Σ outstanding_balance of loans with loan_type == "CRE" (equivalently the sum of CRE sector exposures). existing_cre_concentration = round(existing_cre_exposure / total_loans_outstanding, 4). selected_post_approval_cre_concentration = round((existing_cre_exposure + selected_requested_amount) / (total_loans_outstanding + selected_requested_amount), 4) — both numerator and denominator grow by the new principal. selected_policy_variance_bps = round((post_unrounded - cre_policy_limit_pct) * 10000, 2) (Harbor: 0.5349 − 0.29 ⇒ 2449.15 bps).

CRITICAL denominator rule: concentration denominators use metrics.total_loans_outstanding (e.g. 14,334,094.87 for Lakeview), never branches.total_assets (which is ~25× larger and would make every ratio trivially tiny).

post_approval_concentrations (reporting section):

• exposure_after_approval = existing sector exposure + Σ full approved_amount of approve/conditional apps in that sector (use approved principal, not bank_capacity_used).
• denominator = total_loans_outstanding + gross_approved_amount (pro-forma book including all newly approved principal). Verified on all four Lakeview sectors: Construction 3,934,283.26 / 18,908,348.32 = 0.2081, etc.
• post_approval_pct = round(exposure_after_approval / denom, 4).
• over_limit = post_approval_pct > limit_pct (boolean).

concentration_flags: for each approve/conditional app whose sector is at or near its ceiling, report sector, application_id, limit_pct, post_approval_pct (4dp), flag (true when the deal breaches/near-breaches the sector limit on a committed-capacity basis), handling ∈ {approve, conditional_approve, decline, participation_required, none}. Sort by sector then application_id.

2.4 CDFI factor scores & risk classes (watch-list)

CDFI factor score = sum of four sub-scores (lower = better); null/missing factors score 0.

• ltv: <0.40→0, 0.40–0.60→2, 0.60–0.80→4, >0.80→6
• fico: >720→0, 680–720→1, 580–679→3, <580→5 (null→0)
• debt_to_asset: <0.40→0, 0.40–0.60→2, 0.60–0.80→4, >0.80→6 (null→0)
• liquidity_months: >12→0, 6–12→1, 3–6→3, <3→5 (null→0)

Verified: SUM-LN-003 (ltv 0.9511→6, fico null→0, dta 0.3579→0, liq 11.6→1) ⇒ factor_score 7 ⇒ Desirable. SUM-LN-902 (ltv 1.18→6, dta 0.88→6, liq 1.6→5) ⇒ 17.

Risk class from factor_score (policies.cdfi_factor_scores.classes):

• Prime: 0–5
• Desirable: 6–9
• Satisfactory: 10–13
• Watch: 14–18
• Doubtful: ≥19
• Projected Loss: ≥19 and ltv > 1.0

Severe-delinquency override: a loan with payment_status == "Nonaccrual" is classed Projected Loss regardless of its factor score (it overrides the score-range mapping). Verified on SUM-LN-902 (factor 17, Nonaccrual ⇒ Projected Loss). 90+ Days Past Due does not override — SUM-LN-011/015 stay at their score-derived class.

2.5 +200bp DSCR stress (watch-list) and CRE dual stress

From policies.stress:

• watch_list_parallel_shock = "+200bp" ⇒ shock_label = "+200bp".
• Watch-list formula: stressed_dscr = dscr / (1 + 0.18) = dscr / 1.18. Verified: SUM-LN-003 1.59/1.18 = 1.35; SUM-LN-004 1.01/1.18 = 0.86.
• CRE dual-stress formula (cre_dual_stress_formula): stressed_dscr = dscr * 0.85 / (1 + 0.18) = dscr * 0.85 / 1.18. Verified: HAR-APP-901 1.47×0.85/1.18 = 1.06; HAR-APP-902 1.32×0.85/1.18 = 0.95. Report formula as the compact expression, e.g. "dscr * 0.85 / 1.18".
• coverage_breach_threshold = 1.0. breaches_threshold = stressed_dscr < 1.0 (i.e. stressed < 1.0). Use strict < (Harbor 901 stressed 1.06 ⇒ false; 902 stressed 0.95 ⇒ true).
• Stress results include only loans with DSCR available (skip null-dscr loans such as SUM-LN-010). DSCR values round to 2dp.
• breach_loan_ids = ascending loan_id list of loans whose stressed DSCR breaches.

2.6 Decline reason codes & decision enums

Decision enum: approve, conditional_approve, decline, defer, participation_required (used as a per-application decision and as the selected path).

Conditions enum: participation_required, reduced_amount, board_exception, sba_guaranty_required, startup_monitoring, none. Approve ⇒ ["none"]; conditional_approve lists the mitigations in effect.

Decline reason codes (assign to declined apps; sorted ascending alphabetically in output):

• high_ltv — ltv elevated past the 0.80 boundary (borderline) on an app declined for compounding weakness; clearly high when ltv > 0.85.
• weak_dscr — base DSCR < 1.25 (when DSCR applicable). In CRE-decision contexts, weak_dscr is instead triggered by stressed DSCR < 1.0 (the coverage breach) — verified: HAR-APP-902 (stressed 0.95) gets weak_dscr, HAR-APP-901 (stressed 1.06) does not.
• low_fico — fico < 580.
• recent_bankruptcy — bankruptcy_months_ago not null and < 24 months.
• startup_risk — years_in_business < 2.
• sector_breach — the app's sector post-approval concentration breaches (or would breach) limit_pct without mitigation.
• capacity_limit — the app is credit-acceptable but cannot be funded within remaining lending capacity after higher-priority approvals (e.g. LAK-APP-006: ltv 0.7749, fico 668 — passes credit, declined only on capacity).
• fdic_adverse_variance — branch FDIC delinquency materially exceeds benchmark (applies to all apps at an under-performing branch in a CRE decision; both Harbor apps carry it).
• ncua_peer_weakness — credit-union segment's state is weaker than national and peer median.
• underwater_collateral, policy_floor_missing, documentation_gap — use when collateral/value, policy floor, or documentation gaps are the binding weakness (documentation_complete == 0 ⇒ documentation_gap).

Decision logic (transferable):

• approve — passes every credit test (ltv ≤ 0.85 band, dscr ≥ 1.25 where applicable, fico ≥ 580, no recent bankruptcy, years ≥ 2, documentation complete) AND fits capacity and sector ceilings. approved_amount = requested_amount; bank_capacity_used = approved_amount.
• conditional_approve — credit-acceptable but needs a mitigation: participation_required (sector near ceiling), sba_guaranty_required (startup risk partly offset by SBA guaranty), startup_monitoring, reduced_amount, or board_exception.
• decline — compounding credit weaknesses (e.g. high_ltv + weak_dscr; or low_fico + recent_bankruptcy) or capacity_limit.
• defer — for a competing-credit scenario, the weaker credit that is not selected but not outright declined (remediation possible). unselected_disposition ∈ {decline, defer}.

2.7 Watch-list action coverage & workout queues

Watch-list action coverage (regrade tasks, e.g. Redwood): covers the subset of the regrade population whose final (re-derived) rating ≥ 6. Loans regraded to ≤ 5 are dropped from follow-up (they are no longer adverse). covered_loan_count / covered_exposure = Σ over final ≥ 6. Group by_action, each action sorted ascending alphabetically; loan_ids ascending.

Action mapping by final_rating (Redwood pattern):

• final 8 (Nonaccrual) → partial_chargeoff_review
• final 7 → special_assets
• final 6 → watchlist

Workout queue (watch-list tasks, e.g. Summit): includes all adverse loans (current_rating >= adverse_rating_min), each assigned a recommended_action based on a payment-status-first cascade, then risk class. Sort the queue descending by exposure, then ascending loan_id.

Action cascade (verified on Summit):

1. Nonaccrual → partial_chargeoff_review (projected_loss = true)
2. 90+ Days Past Due → special_assets (projected_loss = false)
3. Current + class Watch → special_assets
4. Current + class Desirable → watchlist
5. (Current + Prime/mild → monitor; Doubtful/Projected Loss current → workout — extend by analogy for unseen buckets.)

watch_list_summary:

• adverse_rating_min = the threshold from the prompt (e.g. 6).
• adverse_loan_count = count of loans with current_rating >= adverse_rating_min.
• adverse_balance = Σ outstanding_balance of those loans (verified: Summit 7,675,179.41).
• risk_classes = one row per adverse loan, sorted ascending loan_id, with factor_score (integer) and risk_class (with Nonaccrual ⇒ Projected Loss override).
• monitoring_cadence = monthly for an adverse watch-list.

stress_results: all DSCR-bearing adverse loans, ascending loan_id, with base_dscr, stressed_dscr = base/1.18 (2dp), breaches_threshold. breach_loan_ids ascending.

severe_bucket_counts: group adverse loans by (current_rating, payment_status), sorted ascending current_rating then payment_status. Note payment_status ascending is string-ascending, so "90+ Days Past Due" sorts before "Current" (digit 9 < C), and "Current" before "Nonaccrual". loan_count and exposure (Σ outstanding_balance) per bucket.

3. Per-task archetypes & method

A. Rating-migration review (Redwood)

1. GET /api/policies, /api/branches/REDWOOD, /api/branches/REDWOOD/loans?min_current_rating=3, /api/branches/REDWOOD/metrics, /api/benchmarks/fdic/q4-2024.
2. Re-derive final_rating per loan (2.1). Build final_rating_exposure_totals (all target loans, ascending final_rating — count + Σ exposure) and migration_from_current_rating_3 (only loans whose current_rating == 3, grouped by final_rating, with ascending loan_ids).
3. watch_list_action_coverage over final ≥ 6 (2.7).
4. npa_benchmark (2.2) with total_loans_noncurrent_pct.
5. material_downgrades = loans with final - current >= 2 notches worsening, ascending loan_id.
6. top_problem_credit = the worst final_rating loan (prefer Nonaccrual / highest final / largest exposure); include borrower_name, payment_status, recommended_action.

B. Quarterly allocation package (Lakeview)

1. GET /api/branches/LAKEVIEW, /applications, /sector-exposures, /metrics, /api/policies.
2. Decide each application (2.6); compute approved_amount and bank_capacity_used per decision (2.3). Decisions sorted ascending application_id.
3. Aggregate allocation (2.3); priority_ranking = approved + conditional application_ids, highest committee funding priority first (strategic/stronger credits lead).
4. concentration_flags per near-ceiling conditional app (2.3).
5. decline_reasons map (only declined apps) with sorted reason-code lists.
6. post_approval_concentrations for every sector touched by an approval (reporting denominator = total_loans_outstanding + gross_approved_amount; 2.3), sorted ascending sector.

C. Credit-union segment posture (Civic NC)

1. GET /api/credit-union-segments/{segment_id}, /api/policies, /api/benchmarks/ncua/q1-2025, /api/manifest.
2. state_metrics from the segment state_code NCUA row (integer values); benchmark_version = "ncua_q1_2025".
3. peer_comparison: peer_states from segment (ascending); direction vs US row and vs median of peer rows (2.2), each ∈ {higher, lower, equal}.
4. posture from capacity vs external risk: capacity available + external risk weaker ⇒ continue_with_tighter_conditions; capacity constrained ⇒ temporarily_pause; external risk strong ⇒ continue_approving.
5. controls.required_checklist_gates = segment minimum_checklist; added_operating_controls = the standard tighter set (pre-close insurance binder verification, lien perfection prior to funding, senior underwriter second review, quarterly state benchmark monitoring, monthly segment delinquency watch); include committee_exception_for_capacity_overrun only when capacity is constrained.
6. escalation_triggers: ascending trigger_id (ET001 delinquency ≥ 90 bps → credit_risk_manager; ET002 missing insurance/lien exception → operations_control_manager; ET003 quarterly capacity exceeded/exception → lending_committee_chair). Condition/owner enums only.
7. interpretation: derive capacity_status, external_risk_status, risk_tolerance (from segment), and the matching committee_message.

D. Watch-list stress packet (Summit)

1. GET /api/branches/SUMMIT, /api/branches/SUMMIT/loans?min_current_rating=6, /api/policies.
2. watch_list_summary (2.7): factor scores, risk classes (Nonaccrual ⇒ Projected Loss), adverse_balance, monitoring_cadence = monthly.
3. stress_results (2.5): +200bp, dscr/1.18, threshold 1.0, DSCR-bearing loans only.
4. workout_queue (2.7): all adverse loans, action cascade, descending exposure then ascending loan_id.
5. severe_bucket_counts (2.7) with the string-ascending payment_status order.

E. Competing CRE decision (Harbor)

1. GET /api/branches/HARBOR, /metrics, /loans?loan_type=CRE, /sector-exposures, /applications, /api/policies, /api/benchmarks/fdic/q4-2024.
2. For each competing CRE app compute weighted_cdfi_score (1dp, lower better) using cre_weighted_score weights (capacity 0.45, collateral_exposure 0.36, conditions 0.11, character 0.05, capital 0.03) — capacity (DSCR) and collateral (LTV) carry 81% of the weight; map to score_class (approve_quality ≤ 2.0, conditional ≤ 3.0, weak > 3.0).
3. stress (2.5 CRE dual): formula = "dscr * 0.85 / 1.18", threshold 1.0. weak_dscr reason fires when stressed < 1.0.
4. concentration (2.2/2.3): existing & post-approval CRE concentration vs cre_policy_limit_pct; FDIC variance vs total_real_estate_30_89_pct.
5. applications_compared (ascending application_id): each with score, class, decision, sorted reason_codes. Reasons include fdic_adverse_variance and sector_breach (both apps at an under-performing, CRE-heavy branch) plus weak_dscr only for the stressed-breach app.
6. recommended_path: select the lower (better) weighted score app; its decision becomes path. unselected_disposition ∈ {decline, defer}; unselected reason_codes ascending.
7. conditions: for a participation_required selected CRE path, the full set (all 7 enum values), sorted ascending alphabetically: bank_retained_exposure_cap, committee_cre_exception, minimum_dscr_covenant_1_25, no_additional_cre_without_committee_review, quarterly_financial_reporting, tenant_roll_and_lease_review, updated_appraisal_before_close.

4. Output field definitions & exact enums (consolidated)

• payment_status: Current, 30 Days Past Due, 60 Days Past Due, 90+ Days Past Due, Nonaccrual.
• recommended_action / watch-list action: monitor, watchlist, special_assets (underscore — this is the spelling used by both the Redwood and Summit templates and their gold answers), workout, partial_chargeoff_review, legal_referral. Still copy the exact token from the task's answer_template.json enum in case a future template differs.
• decision: approve, conditional_approve, decline, defer, participation_required.
• conditions: participation_required, reduced_amount, board_exception, sba_guaranty_required, startup_monitoring, none.
• handling: approve, conditional_approve, decline, participation_required, none.
• risk_class: Prime, Desirable, Satisfactory, Watch, Doubtful, Projected Loss.
• score_class: approve_quality, conditional, weak.
• posture: continue_approving, continue_with_tighter_conditions, temporarily_pause.
• capacity_status: capacity_available, capacity_constrained, no_capacity.
• external_risk_status: stronger_than_national_and_peers, mixed_vs_national_and_peers, weaker_than_national_and_peers.
• risk_tolerance: restrained, moderate, expansive.
• committee_message (segment): capacity_available_but_external_risk_weaker, pause_until_state_metrics_recover, routine_approval_path_supported.
• unselected_disposition: decline, defer.
• Decline reason_code set: 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.

Ordering rules (apply unless the template says otherwise):

• lists of loans/applications/actions: ascending loan_id / application_id;
• final_rating ascending; sector ascending; action/condition/reason-code ascending alphabetically; trigger_id ascending;
• material_downgrades ascending loan_id; workout_queue descending exposure then ascending loan_id; severe_bucket_counts ascending current_rating then string-ascending payment_status;
• priority_ranking: highest priority first (approved + conditional only);
• numeric lists inside objects follow the ordering field in the template.

5. Numeric conventions

• Money / exposure / balance: 2 decimals USD (e.g. 1725000.00).
• Ratios (concentration, NPA, FDIC variance, post_approval_pct, limit_pct): 4 decimals as a ratio (0.1897, not 18.97%).
• variance_bps / *_variance_bps / policy_variance_bps: 2 decimals, signed (positive = branch/app worse than benchmark/policy); computed from the unrounded ratio difference × 10000.
• dscr / base_dscr / stressed_dscr: 2 decimals.
• weighted_cdfi_score: 1 decimal (lower is better).
• factor_score, ratings, counts, bps-in-state-metrics: integers.
• Percentages in policy/templates are ratios (0.19), never percent integers.
• Round only at output; keep unrounded intermediates for variance_bps so the bps figure stays consistent with the rounded ratio (Redwood 0.1135 ratio but 1037.49 bps, not 1037.00).

6. Common misjudgments & exclusion rules

1. Regrade population ≠ watch-list coverage. Regrade population = current_rating >= target_current_rating_min. Watch-list action coverage = subset with re-derived final_rating ≥ 6. Do not list final ≤ 5 loans in coverage, and do not use current_rating to decide coverage.
2. Severe-delinquency override is Nonaccrual only. Nonaccrual ⇒ Projected Loss regardless of factor_score. 90+ Days Past Due does not override the score-derived class (SUM-LN-011/015 stay Desirable).
3. Concentration denominator = total_loans_outstanding, not total_assets. total_assets is ~25–35× larger and yields trivially small ratios. Always pull total_loans_outstanding from metrics (latest quarter).
4. Pro-forma denominator grows with new principal. For post_approval_concentrations the denominator = total_loans_outstanding + gross_approved_amount; for CRE selected_post_approval_cre_concentration both numerator and denominator grow by the selected requested_amount. Keep numerator and denominator consistent.
5. approved_amount is always full principal; bank_capacity_used is the reduced, retained figure. SBA-guaranty ⇒ approved × (1 − sba_guaranty_pct); participation ⇒ sized to sector limit. Never substitute bank_capacity_used into gross_approved_amount or sector exposure_after_approval.
6. variance_bps uses the unrounded ratio. Recompute from the raw exposure/total fraction before multiplying by 10000, then round to 2dp. This avoids the 0.5–1 bps drift that matching judges penalize.
7. Stress excludes null-DSCR loans. stress_results lists only loans with a DSCR; null-dscr loans (often Consumer/Equipment) still appear in risk_classes and workout_queue.
8. Material downgrade threshold is 2 notches. Include only loans worsened by ≥ 2 (final - current >= 2). Single-notch downgrades are not "material".
9. benchmark_metric choice tracks portfolio focus. General NPA review ⇒ total_loans_noncurrent_pct; CRE decision task ⇒ total_real_estate_30_89_pct paired with delinquency_30_plus_pct. Do not mix a noncurrent metric with a 30-89 branch figure.
10. Payment-status ordering is string-ascending, not severity-ascending. "90+ Days Past Due" < "Current" < "Nonaccrual" because '9' < 'C'. In severe_bucket_counts, a rating-7 "90+ DPD" bucket sorts before a rating-7 "Current" bucket.
11. top_problem_credit should be the genuinely worst credit (highest final_rating, prefer Nonaccrual, then largest exposure), not the largest balance regardless of rating.
12. Selection in a competing-credit task = lower weighted score. The selected app is the one with the lower weighted_cdfi_score (better quality); its decision becomes the path; the other is defer/decline.
13. Always honor the exact enum spelling in the target answer_template.json. The action enum is consistently special_assets (underscore) across the train templates and gold answers; still verify against the target template's allowed_values rather than assuming.
14. Do not invent peer states or benchmark versions. Peer states come from the segment JSON; benchmark versions come from /api/manifest. Use only the enum choices and identifiers the template allows.

7. End-to-end execution checklist

1. Read the prompt; fix the branch_id / segment_id, as-of date, and which answer_template to mirror.
2. GET /api/policies and /api/manifest first; cache the rating tables, stress formulas, score weights, and benchmark versions.
3. Pull the branch, metrics (latest quarter), loans (?min_current_rating= as needed), sector-exposures, applications, and the relevant benchmark (FDIC and/or NCUA all-rows).
4. Re-derive ratings / scores / stresses with the rules above; never trust the stored current_rating as the final answer.
5. Assemble the JSON exactly matching the template's required_top_level_keys and per-section required_keys, applying ordering and precision rules.
6. Emit only valid JSON — no narrative outside the object. Money 2dp, ratios 4dp, bps 2dp signed, scores 1dp, DSCR 2dp.