Fewshot Attempt 03

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

A transferable method for producing committee-ready JSON answers against the shared credit-office public API. Distilled from five fully-worked train tasks (Redwood rating migration, Lakeview allocation, Civic NC fire/EMS posture, Summit watch-list stress, Harbor competing-CRE). Use this as the operating procedure whenever a task asks for risk-rating re-derivation, NPA/FDIC/NCUA variance, capacity & concentration analysis, CDFI risk classification, +200bp DSCR stress, decline-reason coding, or watch-list/workout queuing.

The environment is REMOTE and read-only. Never read local env/ source, DB files, or scripts. Never call /api/judge (no judge endpoint is available to you).

1. Remote API usage SOP

Base URL: <remote-env-url> (all endpoints return JSON, no auth)

Call discipline: curl -s GET only. Pipe through jq to inspect/shape. Write large responses to /tmp/<name>.json then jq subsets, so you do not re-fetch.

Endpoint map (the only surfaces you need):

branch_id values are uppercase (REDWOOD, LAKEVIEW, SUMMIT, HARBOR, …). Segment ids look like CIVIC_NC_FIRE_EMS.

Always fetch the latest quarter from /metrics (sort by quarter, take last; observed 2025Q1). Do not assume the first array element is current.

Key object field names (do not guess):

• Loan exposure = outstanding_balance (NOT balance, which is null). Loans also carry dscr, ltv, debt_to_asset, fico, liquidity_months, payment_status, days_past_due, current_rating, sector, loan_type, annual_debt_service, collateral_value, borrower_name.
• Application requested size = requested_amount. Also: sba_guaranty_pct, bankruptcy_months_ago, years_in_business, co_guarantor_strength, documentation_complete, prior_delinquencies_12m, dti, existing_relationship_years.
• total_loans_outstanding lives on metrics, not on the branch object (the branch object has no such field). /loans sum and /metrics total can differ; prefer /metrics.total_loans_outstanding as the authoritative denominator.

2. Numeric conventions (apply everywhere)

variance_bps rule (critical): variance_bps = (branch_ratio − benchmark_ratio) × 10000, signed, rounded to 2dp. Compute branch_ratio at full precision from the raw numerics — do NOT recompute it from the 4dp-rounded display ratio. When the branch ratio is itself a stored metric (e.g. delinquency_30_plus_pct), that stored value is the full-precision input. The 4dp *_ratio fields in the output are display-rounded independently.

Example (verified): nonperforming 1,725,000 / total_loans 15,191,701.54 = 0.11354884… ; minus FDIC 0.0098 = 0.10374884… ; ×10000 = 1037.49 (not 1037.00 from the rounded 0.1037).

Ordering rules (apply to every list):

• Any loan_ids array: ascending string sort.
• material_downgrades, workout_queue tie-break, etc.: ascending loan_id.
• final_rating_exposure_totals, migration_from_current_rating_3: ascending by final_rating.
• watch_list_action_coverage.by_action: ascending by action name.
• workout_queue: descending exposure, then ascending loan_id.
• severe_bucket_counts: ascending current_rating, then payment_status.
• decisions, applications_compared, post_approval_concentrations: ascending application_id / sector.
• concentration_flags: ascending sector then application_id.
• escalation_triggers: ascending trigger_id.
• Any alphabetized enum list (reason_codes, conditions, unselected_reason_codes): ascending alphabetical.

3. The policy tables (from /api/policies)

Read /api/policies once and hold these tables.

3.1 Risk-rating re-derivation (risk_rating)

• dominant_factor_rule: final rating = the worst (maximum) numeric rating from the available DSCR, LTV, and delinquency factors. Current rating is NOT a factor when any objective factor is available; it is only the fallback when every factor is missing (dscr null AND ltv null AND payment Current).
• DSCR → rating: ≥1.5→3, ≥1.25→4, ≥1.05→5, ≥1.0→6, <1.0→7.
• LTV → rating: ≤0.65→3, ≤0.75→4, ≤0.85→5, ≤1.0→6, >1.0→7.
• Delinquency minimums (the severe-delinquency override — a floor, not a cap): 30 DPD→4, 60 DPD→5, 90+ DPD→7, Nonaccrual→8, Current→null.
• material_downgrade_notches = 2: a downgrade of ≥2 notches is "material" and goes in material_downgrades.

Re-derive each loan by taking max(dscr_rating?, ltv_rating?, delinquency_floor?) where ? = "only if present." This can produce upgrades (e.g. strong DSCR+LTV can lift a loan from 4 to 3). A loan with all factors missing keeps its current_rating. A Nonaccrual loan floors at 8 regardless of DSCR/LTV.

3.2 CDFI factor scores (cdfi_factor_scores) — 0 best, 6 worst

• 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
• fico: >720→0, 680–720→1, 580–679→3, <580→5
• liquidity_months: >12→0, 6–12→1, 3–6→3, <3→5
• factor_score = sum of available factor sub-scores (skip null factors).
• Class by total score: 0–5 Prime, 6–9 Desirable, 10–13 Satisfactory, 14–18 Watch, ≥19 Doubtful, Projected Loss = ≥19 AND ltv>1.0 — with the observed severe-delinquency override: an adversely-rated, Nonaccrual loan that is underwater (ltv>1.0) is classed Projected Loss even if its numeric score is below 19 (the loss is projected from collateral shortfall + nonaccrual).

3.3 CRE weighted score (cre_weighted_score) — lower is better

• Weights: capacity 0.45, collateral_exposure 0.36, conditions 0.11, character 0.05, capital 0.03 (sum 1.0).
• weighted_cdfi_score = Σ weight_c × sub-score_c, each sub-score on the 0–6 scale (0 best). Map the five Cs to application factors: collateral_exposure←ltv table, capital←debt_to_asset (total_debt/total_assets) table, character←fico table (or co_guarantor_strength: strong→0, standard→3, none→6 when fico is null), conditions←sector/loan-type risk (stable CRE→low, cyclical Hospitality→high), capacity←DSCR mapped to a 0/2/4/6 band (higher DSCR = lower score).
• score_class: ≤2.0 approve_quality, ≤3.0 conditional, >3.0 weak.

3.4 Stress (stress)

• coverage_breach_threshold = 1.0 (DSCR below 1.0 = breach).
• Watch-list stress (+200bp): stressed_dscr = dscr / (1 + 0.18) i.e. dscr/1.18. shock_label = "+200bp". Apply only to loans with DSCR available; omit null-DSCR loans from stress_results.results.
• CRE dual stress: stressed_dscr = dscr × 0.85 / 1.18. Reported formula string: "dscr * 0.85 / 1.18".
• Use the policy formula verbatim — the "+200bp" label is a name, the actual divisor is 1.18 (1 + 0.18).

3.5 Capacity & concentration (capacity_concentration)

• lending_capacity_field = branches.lending_capacity_q1.
• single_sector_default_field = branches.sector_ceiling_pct; per-sector overrides come from sector-exposures.limit_pct.
• allowed_mitigations = [participation_required, reduced_amount, board_exception].
• Grandfathering: existing over-ceiling exposure is grandfathered, but new approvals may not worsen an over-ceiling sector without mitigation.

4. Transferable business rules by domain

4.1 Risk-rating regrade (Redwood-style)

Population = all loans with current_rating ≥ target_current_rating_min (the prompt specifies the min, e.g. "rated 3 or worse" → min 3). Re-derive every loan per §3.1.

Output pieces:

• target_exposure = sum of outstanding_balance over the population.
• final_rating_exposure_totals: group population by final_rating (ascending), with loan_count and exposure.
• migration_from_current_rating_3: the subset whose current_rating == 3 (or whatever the "from" rating is), grouped by final_rating, with loan_ids ascending.
• material_downgrades: every loan whose downgrade_notches = current − final ≥ 2, with loan_id, current_rating, final_rating, downgrade_notches, exposure; ascending loan_id. (downgrade_notches is positive when rating worsened.)
• top_problem_credit: the single worst credit — highest final_rating (ties → largest exposure; a Nonaccrual/rating-8 loan with the largest exposure wins). recommended_action per the §4.7 action map.
• watch_list_action_coverage: covers loans with final_rating ≥ 6 (NOT the regrade population — only those needing follow-up). group by recommended_action, with loan_ids ascending. covered_loan_count/covered_exposure are the sum. Loans with final ≤5 are not covered (they are routine monitor).

Misjudgment guard: regrade population (current_rating ≥ min) ≠ watch-list coverage population (final_rating ≥ 6). A loan can be in the regrade but not covered (e.g. upgraded to ≤5), and vice versa is impossible (covered ⇒ in regrade when min≤6). Do not mix them.

4.2 NPA & FDIC/NCUA variance

FDIC (bank branches): choose the benchmark metric by the branch's character:

• mixed/general portfolio → total_loans_noncurrent_pct (0.0098)
• real-estate/CRE-weighted → total_real_estate_30_89_pct (0.0051) or total_real_estate_noncurrent_pct (0.0121)
• construction-heavy → construction_development_noncurrent_pct (0.0076) Full FDIC ratios: total_loans_noncurrent 0.0098, total_real_estate_noncurrent 0.0121, total_real_estate_30_89 0.0051, construction_development_noncurrent 0.0076, construction_development_30_89 0.0042.

Branch NPA ratio (NPA-style task):

• branch_npa_exposure = metrics.nonperforming_loans (latest quarter) — equivalently sum of outstanding_balance where payment_status ∈ {90+ Days Past Due, Nonaccrual}.
• branch_total_loans = metrics.total_loans_outstanding.
• branch_npa_ratio = branch_npa_exposure / branch_total_loans (4dp).
• variance_ratio = branch_npa_ratio − fdic_benchmark_ratio (4dp).
• variance_bps = (branch_npa_ratio_full − fdic_benchmark_ratio) × 10000, signed, 2dp.

Branch delinquency vs FDIC (CRE-style task): branch_delinquency_ratio = metrics.delinquency_30_plus_pct (already 4dp). fdic_variance_ratio = branch_delinquency_ratio − fdic_benchmark_ratio; fdic_variance_bps = (branch_delinquency_ratio − fdic_benchmark_ratio) × 10000 (signed, 2dp).

CRITICAL denominator rule: concentration/NPA ratios use total_loans_outstanding (from metrics), never total_assets (total_assets is ~30× larger and yields nonsense ratios). NPA/delinquency variance is reported in bps and is signed (positive = branch underperforming the benchmark).

NCUA (credit-union segments): pull the target-state row and the US row from /api/benchmarks/ncua/q1-2025; pull peer states from the segment's peer_states. For each of the four metrics (delinquency_bps, loan_to_share_pct, roaa_bps, positive_net_income_pct), set direction higher/lower/equal for NC-vs-US and NC-vs-peer-median (median = middle value of the sorted peer values). "Higher delinquency" and "lower roaa / lower positive_net_income" mean NC is weaker.

4.3 Capacity & concentration ceilings (Lakeview/Harbor-style)

Lending capacity:

• lending_capacity_q1 = branch field.
• For each approved/conditional app, bank_capacity_used (the capacity the bank actually consumes) depends on mitigation:
  • plain approve → bank_capacity_used = approved_amount (100% retained).
  • SBA guaranty → bank_capacity_used = approved_amount × (1 − sba_guaranty_pct).
  • participation_required → bank retains B sized so the retained sector exposure lands exactly on the sector limit: (existing_sector_exposure + B) / (total_loans_outstanding + Σ bank_capacity_used_of_all_approved_apps) = sector_limit_pct. Solve for B; the remainder (approved_amount − B) is participated out.
• gross_approved_amount = Σ approved_amount over approve+conditional apps (full amounts, before SBA/participation).
• committed_capacity_amount = Σ bank_capacity_used over approve+conditional apps.
• remaining_capacity = lending_capacity_q1 − committed_capacity_amount.
• Apps are processed in priority order; capacity is consumed in that order. priority_ranking = approve+conditional apps ordered by credit strength (observed: DSCR descending, missing-DSCR apps last; committee/9xx applications typically rank high), highest first. Declined apps are excluded from priority_ranking.

Sector concentration (post-approval view):

• exposure_after_approval (sector) = existing_sector_exposure + full_approved_amount of the app(s) in that sector (full amount is booked to the sector even when participation/SBA reduces capacity — participation is a funding/capacity mitigation, not a sector removal).
• Denominator = total_loans_outstanding + gross_approved_amount (existing loans plus all full approved amounts).
• post_approval_pct = exposure_after_approval / denominator (4dp).
• over_limit = post_approval_pct > limit_pct (strict).
• limit_pct per sector from sector-exposures (override) or branch sector_ceiling_pct.

concentration_flags: flag an application whose post-approval sector concentration reaches at or very near the sector limit (within ~10 bps, or over). handling = participation_required (themitigation). Sectors comfortably below the limit are not flagged. Order by sector then application_id.

CRE-portfolio concentration (Harbor-style):

• cre_policy_limit_pct = branch.cre_policy_limit_pct.
• existing_cre_exposure = Σ outstanding_balance where loan_type == "CRE".
• existing_cre_concentration = existing_cre_exposure / total_loans_outstanding (4dp).
• selected_post_approval_cre_concentration = (existing_cre_exposure + selected_amount) / (total_loans_outstanding + selected_amount) (4dp; full selected amount).
• selected_policy_variance_bps = (selected_post_approval_cre_concentration_full − cre_policy_limit_pct) × 10000, signed, 2dp.

4.4 Decline reason codes & decision enums

Evaluate each application against credit gates; decline reasons list every unmitigated gate that fails (sorted ascending alphabetically).

Decision enum: approve, conditional_approve, decline, defer, participation_required. Conditions enum: participation_required, reduced_amount, board_exception, sba_guaranty_required, startup_monitoring, none. Handling enum: approve, conditional_approve, decline, participation_required, none.

Decision flow: documentation gate → hard credit gates (decline if unmitigated failure) → sector-capacity gate (decline capacity_limit/sector_breach if sector over ceiling unmitigated) → lending-capacity gate (decline capacity_limit if exhausted) → else conditional_approve when a mitigant is required (SBA guaranty, participation, startup monitoring, reduced amount) → else approve.

• Declined and plain-approved apps get conditions: ["none"].
• conditional_approve lists the mitigants, e.g. ["sba_guaranty_required","startup_monitoring"].
• priority_ranking includes approve + conditional_approve only (not declines/defers).

4.5 CDFI risk class & +200bp watch-list stress (Summit-style)

Population = loans with current_rating ≥ adverse_rating_min (prompt-specified, e.g. "6 or worse" → min 6).

• adverse_balance = Σ outstanding_balance over the population.
• For each loan, compute factor_score (§3.2) and risk_class (§3.2), applying the Projected Loss override for underwater (ltv>1.0) Nonaccrual loans. risk_classes list ascending loan_id.
• monitoring_cadence: monthly for an adverse watch-list (ratings ≥6); quarterly /semiannual only for milder populations.
• DSCR stress (§3.4): stressed_dscr = dscr / 1.18, threshold 1.0. Include only loans with DSCR available. breach_loan_ids = those with stressed_dscr < 1.0, ascending.
• workout_queue: all adverse loans, ordered descending exposure, then ascending loan_id. recommended_action per §4.7. projected_loss = (risk_class == "Projected Loss").
• severe_bucket_counts: group the adverse population by (current_rating, payment_status), ascending rating then payment_status, with count + exposure.

4.6 Watch-list action coverage & workout queues

Distinct from §4.5's CDFI queue — this is the risk-rating-driven action map (used in regrade tasks). For loans needing follow-up (final_rating ≥ 6), assign:

Escalation by payment status: a 90+ Days Past Due loan escalates one action level (e.g. Desirable-class 90+DPD → special_assets); Nonaccrual + underwater → partial_chargeoff_review with projected_loss = true. In the workout queue (§4.5), projected_loss is true only for Projected Loss class.

4.7 Recommended-action enum (all tasks)

monitor, watchlist, special_assets, workout, partial_chargeoff_review, legal_referral — ascending severity. payment_status enum: Current, 30 Days Past Due, 60 Days Past Due, 90+ Days Past Due, Nonaccrual.

5. Output field definitions & exact enums (per task archetype)

5.1 Rating-migration review (Redwood)

Top-level: branch_id, review_date(YYYY-MM-DD), portfolio_regrade, npa_benchmark, material_downgrades, top_problem_credit.

• portfolio_regrade: target_current_rating_min(int), target_loan_count, target_exposure(2dp), final_rating_exposure_totals[](asc final_rating), migration_from_current_rating_3[](asc final_rating, loan_ids asc), watch_list_action_coverage{covered_loan_count, covered_exposure, by_action[](asc action)}.
• npa_benchmark: benchmark_version, benchmark_metric(enum: total_loans_noncurrent_pct | total_real_estate_noncurrent_pct | construction_development_noncurrent_pct), branch_npa_exposure, branch_total_loans, branch_npa_ratio(4dp), fdic_benchmark_ratio(4dp), variance_ratio(4dp), variance_bps(2dp signed).
• material_downgrades[](asc loan_id): loan_id, current_rating, final_rating, downgrade_notches, exposure.
• top_problem_credit: loan_id, borrower_name, exposure, current_rating, final_rating, payment_status(enum), recommended_action(enum).
• watch_list_action_coverage.by_action action enum: monitor | watchlist | special_assets | workout | partial_chargeoff_review | legal_referral.

5.2 Allocation package (Lakeview)

Top-level: 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[](application_id, approve+conditional only).
• decisions[](asc application_id): application_id, decision(enum), approved_amount, bank_capacity_used, conditions[](enum).
• concentration_flags[](asc sector, application_id): sector, application_id, limit_pct, post_approval_pct, flag(bool), handling(enum).
• decline_reasons{ app_id: [reason_codes sorted asc] } — map only declined apps.
• post_approval_concentrations[](asc sector): sector, exposure_after_approval, post_approval_pct, limit_pct, over_limit(bool).
• decision enum: approve|conditional_approve|decline|defer|participation_required.
• conditions enum: participation_required|reduced_amount|board_exception| sba_guaranty_required|startup_monitoring|none.
• reason_code enum: 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.

5.3 Credit-union segment posture (Civic NC fire/EMS)

Top-level: segment_id, posture, state_metrics, peer_comparison, controls, escalation_triggers, interpretation.

• posture enum: continue_approving | continue_with_tighter_conditions | temporarily_pause.
• state_metrics: state_code, benchmark_version, delinquency_bps, loan_to_share_pct, roaa_bps, positive_net_income_pct (integers exactly as NCUA reports).
• peer_comparison: peer_states[](asc), nc_vs_us{4 directions}, nc_vs_peer_median{4 directions}; direction enum higher|lower|equal (higher delinquency / lower roaa / lower positive_net_income = NC weaker).
• controls: required_checklist_gates(set, from segment.minimum_checklist), added_operating_controls(set, chosen from enum based on segment.internal_context). required enum: board_authorization|equipment_invoice|fleet_replacement_plan| payer_contract_summary|public_contract_or_tax_support|proof_of_insurance| ucc_or_title_lien. added enum: pre_close_insurance_binder_verification|lien_perfection_prior_to_funding| senior_underwriter_second_review|quarterly_state_benchmark_monitoring| monthly_segment_delinquency_watch|committee_exception_for_capacity_overrun.
• escalation_triggers[](asc trigger_id, e.g. ET001/ET002/ET003): trigger_id, condition(enum), owner(enum). condition enum: segment_recent_delinquency_ge_90_bps| missing_insurance_or_lien_exception|quarterly_capacity_exceeded_or_exception_requested| state_delinquency_gap_widens_25_bps. owner enum: credit_risk_manager|operations_control_manager|lending_committee_chair. Owner mapping: delinquency trigger → credit_risk_manager; insurance/lien exception → operations_control_manager; capacity/exception → lending_committee_chair.
• interpretation: 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(capacity_available_but_external_risk_weaker| pause_until_state_metrics_recover|routine_approval_path_supported).
• Decision logic: capacity available + external risk weaker (NC delinquency above US and peers, roaa/pni below) + risk_tolerance moderate → continue_with_tighter_conditions, message capacity_available_but_external_risk_weaker. Pick added controls that respond to the segment's specific internal_context (insurance-binder misses → pre_close_insurance_binder_verification + lien_perfection_prior_to_funding; staffing constraint → senior_underwriter_second_review; external state risk → quarterly_state_benchmark_monitoring + monthly_segment_delinquency_watch).

5.4 Watch-list stress packet (Summit)

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

• watch_list_summary: adverse_rating_min, adverse_loan_count, adverse_balance(2dp), risk_classes[](asc loan_id: loan_id, risk_class(enum), factor_score(int)), monitoring_cadence(monthly|quarterly|semiannual).
• stress_results: shock_label("+200bp"), breach_threshold(1.0), results[](asc loan_id, DSCR-available 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(bool).
• severe_bucket_counts[](asc current_rating, then payment_status): current_rating, payment_status, loan_count, exposure.
• risk_class enum: Prime|Desirable|Satisfactory|Watch|Doubtful|Projected Loss.

5.5 Competing-CRE decision (Harbor)

Top-level: branch_id, applications_compared, recommended_path, stress, concentration, conditions.

• applications_compared[](asc application_id): application_id, weighted_cdfi_score(1dp, lower better), score_class(approve_quality|conditional|weak), decision(enum), reason_codes[](asc, reason enum).
• recommended_path: selected_application_id, path(enum), unselected_application_id, unselected_disposition(decline|defer), unselected_reason_codes[](asc, subset enum: sector_breach|weak_dscr|high_ltv|fdic_adverse_variance).
• stress: formula("dscr * 0.85 / 1.18"), coverage_breach_threshold(1.0), results[](asc application_id: application_id, base_dscr, stressed_dscr, breaches_threshold).
• concentration: cre_policy_limit_pct(4dp), existing_cre_exposure(2dp), existing_cre_concentration(4dp), selected_post_approval_cre_concentration(4dp), selected_policy_variance_bps(2dp signed), fdic_benchmark_metric(total_real_estate_30_89_pct), branch_delinquency_ratio(4dp), fdic_benchmark_ratio(4dp), fdic_variance_ratio(4dp), fdic_variance_bps(2dp signed).
• conditions[](asc, enum): 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.
• Decision logic: compute weighted_cdfi_score per app; the lower score is stronger. Run the CRE dual stress on both; the selected credit is the stronger one that also passes the stress (stressed_dscr ≥ 1.0). Because the branch CRE concentration already exceeds the policy limit, the selected path is participation_required (mitigate via participation + covenants). The unselected credit gets unselected_disposition = defer (weak, revisitable) with its reason codes (e.g. fdic_adverse_variance, sector_breach, weak_dscr when it also breaches stress). Apply the full conditions list when the path is participation_required on an over-limit CRE branch.

6. Common misjudgments & exclusion rules (avoid these)

1. Regrade population vs watch-list coverage. Regrade population = current_rating ≥ target_min. Watch-list action coverage = final_rating ≥ 6 (post-regrade). They are different sets. Do not put final-≤5 loans in watch_list_action_coverage; do not count final-≤5 loans as "covered."
2. Severe-delinquency override. Payment status sets a RATING FLOOR (90+→7, Nonaccrual→8) and a CLASS override (Nonaccrual + ltv>1.0 → Projected Loss). A good DSCR/LTV does NOT override a severe delinquency downward. Take the max, not the average.
3. Ascending loan_id everywhere it appears in a list. String sort (RED-LN-008 before RED-LN-013 before RED-LN-901). Workout queue is the exception: descending exposure first, ascending loan_id only as tie-break.
4. Concentration denominator = total_loans_outstanding, NOT total_assets. total_assets is ~30× larger and yields meaningless tiny ratios. Also: the post-approval denominator = total_loans_outstanding + gross_approved_amount (include new full approvals), and existing_cre_exposure uses loan_type=="CRE" (not a sector-name guess).
5. variance_bps uses full-precision inputs. Do not derive bps from the 4dp rounded ratio. Compute ratio from raw numerics, subtract benchmark, ×10000, then round to 2dp. variance_bps is signed (negative if branch beats benchmark).
6. Participation does not remove sector exposure. exposure_after_approval uses the FULL approved amount; only bank_capacity_used is reduced by SBA-guaranty/participation. Don't subtract the participated portion from sector exposure.
7. bank_capacity_used for SBA = approved_amount × (1 − sba_guaranty_pct), not the full amount. For participation, solve the limit equation (§4.3), don't guess.
8. Include only DSCR-available loans in stress_results.results; null-DSCR loans are omitted (but still appear in workout_queue/risk_classes).
9. priority_ranking excludes declines/defers — approve + conditional_approve only.
10. projected_loss boolean is true only for risk_class == "Projected Loss" (underwater + nonaccrual), not for every 90+DPD/nonaccrual loan.
11. over_limit is strict (post_approval_pct > limit_pct); a sector exactly at the limit is over_limit=false but may still trigger a concentration_flags entry (near-limit).
12. Don't call /api/judge and don't read env/, test tasks, or evaluator internals. Stick to the public GET endpoints in §1.

7. Worked example fragments (method only — not copyable gold)

• Risk regrade: a loan current_rating 3, dscr 0.87, ltv 0.82, payment Current → DSCR <1.0→7, LTV ≤0.85→5, delinquency floor null → final max(7,5)=7, downgrade 4 notches → material. A loan current_rating 5, dscr 1.79, ltv 0.63, Current → max(3,3)=3 (upgrade, not material, not in watch-list coverage).
• NPA variance: nonperforming 1,725,000 / total_loans 15,191,701.54 = 0.11355; FDIC total_loans_noncurrent_pct 0.0098; variance_bps = 1037.49 (signed positive).
• SBA capacity: requested 840,000, sba_guaranty_pct 0.75 → bank_capacity_used = 840,000 × 0.25 = 210,000; full 840,000 still counts in gross_approved_amount and sector exposure.
• Participation sizing: existing healthcare 1,937,814.40, sector limit 0.19, approved 1,650,000, total_loans 14,334,094.87, other approved committed 2,294,253.45 → solve (1,937,814.40 + B)/(14,334,094.87 + 2,294,253.45 + B) = 0.19 → B ≈ 1,508,113.31 retained, ≈ 141,886.69 participated.
• +200bp stress: base DSCR 1.59 → 1.59/1.18 = 1.35 (passes ≥1.0); base 1.01 → 0.86 (breaches).
• CRE dual stress: base 1.47 → 1.47×0.85/1.18 = 1.06 (passes); base 1.32 → 0.95 (breaches) → weaker credit deferred.
• CRE concentration: existing CRE 7,011,570.24 / total_loans 14,933,688.02 = 0.4695 (already over the 0.29 policy limit → grandfathered but no worsening without mitigation → participation_required); selected post = (7,011,570.24 + 2,100,000) / (14,933,688.02 + 2,100,000) = 0.5349; selected_policy_variance_bps = 2449.15.
• CDFI class: ltv 1.18→6, debt_to_asset 0.88→6, liquidity 1.6→5, fico null → score 17 (Watch band) but Nonaccrual + ltv>1.0 → Projected Loss override, projected_loss = true, action partial_chargeoff_review.

8. Execution checklist (per task)

1. Parse the prompt: identify branch_id / segment_id, review date, the population threshold (e.g. "rated 3 or worse," "current_rating 6 or worse"), and which archetype (§5) applies. Load the matching answer_template.json.
2. GET /api/policies and hold the tables (§3).
3. Fetch the branch /metrics (latest quarter), /loans (with the right ?min_current_rating=/?loan_type=/?payment_status= filter), /sector-exposures, /applications as the archetype needs.
4. Fetch the relevant benchmark (/api/benchmarks/fdic/q4-2024 or /api/benchmarks/ncua/q1-2025); for NCUA, fetch the state row, the US row, and the segment's peer_states rows.
5. Re-derive / classify / stress per §3–§4, preserving exact enums and orderings.
6. Compute money 2dp, ratios 4dp, bps 2dp signed (full-precision inputs for bps).
7. Emit a single JSON object matching the template — no narrative outside JSON. Verify every list is sorted per §2 and every enum value is from the template's allowed set.