Self Attempt 01

Credit-Risk / Lending-Committee Solver Skill

Self-contained operational skill for solving credit-risk committee evaluation tasks against the remote credit-office public REST API. A solver sees only: a task prompt, an answer_template.json, environment access, and this file. There are NO gold answers and NO judge endpoint available at solve time — all derivation logic must come from the API data and credit policy.

1. ENVIRONMENT ACCESS SOP

Base URL: <remote-env-url> — all endpoints return JSON, no auth, GET only.

Call pattern: use curl -s and pipe through python3 -m json.tool or python3 -c "import sys,json; ...". Save large responses to /tmp/*.json and parse with python to avoid truncation. Always retrieve FULL lists (never slice [0:2] — scenario-critical records use ID suffixes like -901, -902, -903 that sort AFTER normal IDs and may be missed by naive slicing).

Endpoint map (which endpoint feeds which answer section)

ID conventions

• branch_id: uppercase (e.g. REDWOOD, LAKEVIEW, SUMMIT, HARBOR). A task names its own target branch. Credit-union segments also appear in /api/branches with institution_type=credit_union (e.g. CIVIC_NC_FIRE_EMS).
• loan_id: {BRANCH_PREFIX}-LN-NNN (e.g. RED-LN-001). Scenario-injected loans use -901, -902, -903 suffixes.
• application_id: {BRANCH_PREFIX}-APP-NNN (e.g. HAR-APP-901). When a prompt names specific application IDs, fetch all applications for the branch and match — the -9xx IDs always exist.

2. DATA STRUCTURES REFERENCE

Loan object (from /loans)

loan_id, borrower_name, branch_id, loan_type, sector, current_rating, payment_status,
outstanding_balance, collateral_value, dscr, ltv, debt_to_asset, fico, liquidity_months,
guarantor_strength, interest_rate, annual_debt_service, days_past_due, annual_review_date,
notes

• dscr, ltv, fico, liquidity_months can be null (e.g. HELOC, residential mortgages often lack DSCR/LTV).
• current_rating: integer 1-8 (1=best, 8=worst).
• payment_status: Current | 30 Days Past Due | 60 Days Past Due | 90+ Days Past Due | Nonaccrual.

Application object (from /applications)

application_id, applicant_name, business_name, branch_id, loan_type, sector, purpose,
requested_amount, proposed_rate, term_months, dscr, ltv, dti, fico, total_assets,
total_debt, annual_revenue, net_income, collateral_value, co_guarantor_strength,
years_in_business, existing_relationship_years, prior_delinquencies_12m,
bankruptcy_months_ago, documentation_complete (0/1), sba_guaranty_pct,
relationship_deposit_balance, notes

• Applications do NOT have liquidity_months or debt_to_asset — compute debt_to_asset = total_debt / total_assets when needed.
• fico can be null (especially for business/CRE loans).
• co_guarantor_strength: none | limited | standard | strong.

Branch metrics (from /metrics — list of 2 quarters)

quarter ("2025Q1" | "2024Q4"), nonperforming_loans, total_loans_outstanding,
delinquency_30_plus_pct, net_charge_offs, allowance_for_loan_losses, total_deposits

• nonperforming_loans = sum of outstanding_balance for loans with payment_status = "Nonaccrual" (confirmed: matches exactly).
• total_loans_outstanding = sum of ALL loan outstanding_balances (confirmed: matches exactly).
• Always use the latest quarter (2025Q1) unless the task specifies otherwise.

Branch detail (from /branches/{id})

branch_id, branch_name, institution_type (bank|credit_union), state_code,
total_assets, lending_capacity_q1, sector_ceiling_pct, cre_policy_limit_pct,
fdic_benchmark_set

Credit-union segment (from /credit-union-segments/{id})

segment_id, segment_name, state_code, quarterly_capacity, current_outstanding,
member_profile, portfolio_focus[], minimum_checklist[], peer_states[],
risk_tolerance, internal_context {recent_delinquency_bps, control_issue,
staffing_constraint, portfolio_yield_pct}, notes

FDIC benchmark (from /benchmarks/fdic/q4-2024)

benchmark_version: "fdic_q4_2024"
total_loans_noncurrent_pct: 0.0098
total_real_estate_noncurrent_pct: 0.0121
construction_development_noncurrent_pct: 0.0076
total_real_estate_30_89_pct: 0.0051
construction_development_30_89_pct: 0.0042

NCUA benchmark (from /benchmarks/ncua/q1-2025)

benchmark_version: "ncua_q1_2025"
rows[]: each {state_code, delinquency_bps, loan_to_share_pct, roaa_bps, positive_net_income_pct}

• state_code = "US" is the national row. State rows: AL, FL, GA, MI, MN, NC, OH, PA, SC, TN, VA.
• All values are integers (report them exactly as-is, do NOT convert to decimals).

3. CORE BUSINESS RULES (from /api/policies)

3.1 Risk rating re-derivation (tasks 001, general)

Policy key: risk_rating. Re-derive each loan's rating from objective factors, then take the worst (highest) numeric rating across all available factors (dominant_factor_rule).

DSCR thresholds (apply only if dscr is not null):

LTV thresholds (apply only if ltv is not null):

Delinquency minimums (always apply — this is the severe-delinquency override):

Final rating = max(dscr_rating, ltv_rating, delinquency_minimum) over available factors. When a factor is null, skip it. The delinquency minimum ACTS AS A FLOOR — if DSCR/LTV would give a better (lower) rating, the delinquency floor overrides it. This is the "severe-delinquency override."

Material downgrade: downgrade_notches = final_rating - current_rating. Material if >= material_downgrade_notches (2).

3.2 NPA benchmark variance (task 001, 005)

branch_npa_exposure   = metrics.nonperforming_loans  (latest quarter)
branch_total_loans    = metrics.total_loans_outstanding  (latest quarter)
branch_npa_ratio      = branch_npa_exposure / branch_total_loans   (4dp)
fdic_benchmark_ratio  = <selected FDIC metric value>              (4dp)
variance_ratio        = branch_npa_ratio - fdic_benchmark_ratio   (4dp, SIGNED)
variance_bps          = variance_ratio * 10000                    (2dp, SIGNED)

• Positive variance = branch is WORSE than benchmark (branch ratio exceeds benchmark).
• FDIC metric selection (task 001 allows 3 choices): use total_loans_noncurrent_pct for mixed-portfolio bank branches; use total_real_estate_noncurrent_pct for CRE-heavy branches; use construction_development_noncurrent_pct for construction-heavy branches. Pick the one whose loan-type composition best matches the branch portfolio.
• Task 005 template FORCES fdic_benchmark_metric = "total_real_estate_30_89_pct" (value 0.0051). The branch_delinquency_ratio in task 005 = metrics.delinquency_30_plus_pct from the latest quarter.

3.3 Lending capacity & sector concentration (task 002, 005)

Capacity:

lending_capacity_q1    = branch.lending_capacity_q1
gross_approved_amount  = sum of approved_amount for approve + conditional_approve decisions
committed_capacity_amount = sum of bank_capacity_used (= approved_amount after mitigations/reductions)
remaining_capacity     = lending_capacity_q1 - committed_capacity_amount

Sector concentration (the denominator is total_loans_outstanding from branch metrics, NOT total_assets):

existing_sector_pct     = sector_exposure.current_exposure / total_loans_outstanding
post_approval_pct       = (current_exposure + approved_amount) / total_loans_outstanding   (4dp)
limit_pct               = branch.sector_ceiling_pct  (default per-sector ceiling)
over_limit              = post_approval_pct > limit_pct   (boolean)

CRE concentration (task 005):

cre_policy_limit_pct            = branch.cre_policy_limit_pct
existing_cre_exposure           = sum of outstanding_balance for all loans where loan_type == "CRE"  (2dp, USD)
existing_cre_concentration      = existing_cre_exposure / total_loans_outstanding   (4dp)
selected_post_approval_cre_concentration = (existing_cre_exposure + selected.requested_amount) / total_loans_outstanding  (4dp)
selected_policy_variance_bps    = (selected_post_approval_cre_concentration - cre_policy_limit_pct) * 10000  (2dp, SIGNED)

• Some sectors in sector-exposures have limit_pct == cre_policy_limit_pct (these are CRE-designated sectors: e.g. Hospitality, Office). Others use the sector_ceiling_pct. Use loan_type == "CRE" from the loans endpoint as the authoritative CRE exposure measure — it's more precise than aggregating real-estate-named sectors.

Grandfathering rule: existing over-ceiling exposure (where grandfathered == 1 in sector-exposures) may be retained, but NEW approvals may not worsen an over-limit sector without a mitigation (participation_required, reduced_amount, or board_exception).

Allowed mitigations (from policy capacity_concentration.allowed_mitigations): participation_required, reduced_amount, board_exception.

3.4 CDFI factor scoring & risk classes (task 004, 005)

Policy key: cdfi_factor_scores. Four factor tables, each yielding an integer score. Lower score = better.

CDFI total factor_score = sum of the 4 available factor scores (integer).

Risk class (from total score):

Edge case — null factors: When fico is null (common for business loans), assign the maximum penalty (fico→5) since a missing credit score is a risk indicator. When liquidity_months is null (applications don't have it), assign max penalty (5) or use an available proxy. When debt_to_asset is null, compute it from total_debt / total_assets if both are available; otherwise max penalty (6).

3.5 CRE weighted score (task 005)

Policy key: cre_weighted_score. Five C's with weights:

capacity:           0.45
capital:            0.03
character:          0.05
collateral_exposure: 0.36
conditions:         0.11

Inferred factor-to-C mapping (NOT explicitly in policy — derive each C's 0-6 score from the application data):

weighted_cdfi_score = capacity_score*0.45 + capital_score*0.03 + character_score*0.05
                    + collateral_exposure_score*0.36 + conditions_score*0.11   (1dp)

Lower is better. Score class:

3.6 DSCR stress formulas (task 004, 005)

From policy stress:

• Watch-list stress (task 004): stressed_dscr = dscr / (1 + 0.18) — the +200bp parallel shock. shock_label = "+200bp".
• CRE dual stress (task 005): stressed_dscr = dscr * 0.85 / (1 + 0.18) — combines a 15% NOI haircut with the 200bp rate shock. Report as the formula field.
• Breach threshold = coverage_breach_threshold = 1.00 (from policy). breaches_threshold = stressed_dscr < 1.00.
• Only compute stress for loans/applications where dscr is not null. List ordering: ascending by loan_id (or application_id).
• breach_loan_ids: ascending sorted list of IDs where breaches_threshold == true.

3.7 Watch-list action & workout assignment (task 001, 004)

Action enum: monitor, watchlist, special-assets, workout, partial_chargeoff_review, legal_referral.

Inferred action mapping (from final rating + payment_status — not gold-confirmed, use as default):

Workout queue (task 004): include all adverse loans (rating >= 6) or loans requiring active management. Order: descending exposure, then ascending loan_id. Fields: loan_id, exposure, risk_class, payment_status, recommended_action, projected_loss (boolean = true when risk_class == "Projected Loss").

3.8 Decision logic for pending applications (task 002, 005)

Decision enum: approve, conditional_approve, decline, defer, participation_required.

Decline 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.

Inferred decision thresholds (from policy + credit-scoring logic):

Conditions enum (task 002): participation_required, reduced_amount, board_exception, sba_guaranty_required, startup_monitoring, none.

Priority ranking (task 002): ordered list of application_id values, highest priority first, including approved AND conditionally approved applications only (excludes declined/deferred). Rank by credit quality (DSCR, LTV, FICO) — strongest credits first.

3.9 Concentration flags handling (task 002)

For each application that would push a sector over its limit_pct:

flag = "breach" (if post_approval_pct > limit_pct)  -- but flag is actually boolean "over_limit" in post_approval_concentrations
handling = approve | conditional_approve | decline | participation_required | none

• If the sector is already grandfathered (over ceiling), new approval in that sector requires participation_required or reduced_amount mitigation or must decline.
• concentration_flags list ordering: sort by sector, then application_id.

4. TASK-TYPE PLAYBOOKS

Task type A: Rating migration review (task 001 pattern)

Inputs needed: branch loans (filtered ?min_current_rating=3), branch metrics (latest quarter), FDIC benchmark, policies.

Steps:

1. Fetch loans?min_current_rating=3 → regrade population. target_current_rating_min = 3, target_loan_count = len(population), target_exposure = sum(outstanding_balance).
2. For each loan, re-derive rating using §3.1 rules. Skip null factors.
3. final_rating_exposure_totals: group ALL regrade loans by final_rating (ascending), show count + exposure.
4. migration_from_current_rating_3: group ONLY loans with current_rating == 3 by final_rating (ascending), show count + exposure + loan_ids (ascending).
5. material_downgrades: loans where final_rating - current_rating >= 2, ordered ascending by loan_id. Fields: loan_id, current_rating, final_rating, downgrade_notches, exposure.
6. npa_benchmark: per §3.2. Choose FDIC metric matching portfolio. benchmark_version = "fdic_q4_2024".
7. top_problem_credit: the loan with worst final_rating (highest), worst payment_status, highest exposure. Fields include borrower_name, payment_status, recommended_action.
8. watch_list_action_coverage: group regrade loans needing follow-up by action (§3.7). covered_loan_count, covered_exposure, by_action list (ascending by action name, loan_ids ascending).

Task type B: Allocation package (task 002 pattern)

Inputs needed: branch detail, branch metrics, sector-exposures, applications, policies.

Steps:

1. Fetch all applications for the branch. For each, evaluate DSCR/LTV/FICO/bankruptcy/years_in_business/documentation.
2. Assign decision per §3.8. For approve/conditional_approve, set approved_amount (= requested_amount, or reduced if mitigation) and bank_capacity_used.
3. allocation: lending_capacity_q1, gross_approved_amount (sum of approved_amounts), committed_capacity_amount (post-mitigation), remaining_capacity, priority_ranking (approved + conditional only, best credit first).
4. decisions: list sorted by application_id ascending. Fields: application_id, decision, approved_amount, bank_capacity_used, conditions.
5. concentration_flags: for each application in a potentially breaching sector, compute post_approval_pct vs limit_pct. Sort by sector then application_id.
6. decline_reasons: object mapping each declined application_id to a sorted (ascending alphabetically) list of reason codes.
7. post_approval_concentrations: list sorted by sector ascending. Fields: sector, exposure_after_approval, post_approval_pct (4dp), limit_pct, over_limit (bool). Denominator = total_loans_outstanding.

Task type C: Credit-union segment posture (task 003 pattern)

Inputs needed: segment data, NCUA benchmarks, policies.

Steps:

1. Fetch segment by segment_id. Extract state_code, peer_states, minimum_checklist, risk_tolerance, internal_context.
2. state_metrics: NCUA row for the segment's state_code. Report integer values exactly. Fields: state_code, benchmark_version ("ncua_q1_2025"), delinquency_bps, loan_to_share_pct, roaa_bps, positive_net_income_pct.
3. peer_comparison:
   • peer_states: ascending sorted list from segment data.
   • nc_vs_us: direction (higher/lower/equal) of state value vs US national value, for each of the 4 metrics.
   • nc_vs_peer_median: direction of state value vs median of peer_states' values, for each of the 4 metrics.
   • Higher delinquency = worse; higher roaa/positive_net_income = better; higher loan_to_share = neutral-to-risky.
4. posture: based on capacity + external risk:
   • capacity available + external risk weaker → continue_with_tighter_conditions
   • capacity available + external risk strong → continue_approving
   • no capacity or severe external weakness → temporarily_pause
5. controls:
   • required_checklist_gates: set from segment's minimum_checklist (intersect with allowed choices).
   • added_operating_controls: inferred from segment's internal_context (control_issue → matching control, staffing_constraint → second review, recent_delinquency → delinquency watch, external risk → benchmark monitoring).
6. escalation_triggers: list ascending by trigger_id. Assign owner based on trigger type (risk → credit_risk_manager, operations → operations_control_manager, capacity/exception → lending_committee_chair).
7. interpretation: capacity_status (from segment capacity), external_risk_status (from NC vs US + peers), risk_tolerance (from segment), committee_message (matching combination).

Task type D: Watch-list stress packet (task 004 pattern)

Inputs needed: branch loans (filtered ?min_current_rating=6), policies.

Steps:

1. watch_list_summary: adverse_rating_min = 6, adverse_loan_count, adverse_balance (sum outstanding_balance).
2. risk_classes: for each adverse loan, compute CDFI factor_score (§3.4) and risk_class. List ascending by loan_id. Fields: loan_id, risk_class, factor_score (integer).
3. stress_results: shock_label = "+200bp", breach_threshold = 1.00. For loans with DSCR available, compute stressed_dscr = dscr / 1.18 (2dp). breaches_threshold = stressed_dscr < 1.00. List ascending by loan_id. breach_loan_ids ascending.
4. workout_queue: all adverse loans needing active management. Order: descending exposure, then ascending loan_id. Fields: loan_id, exposure, risk_class, payment_status, recommended_action, projected_loss (bool = risk_class == "Projected Loss").
5. severe_bucket_counts: group adverse loans by (current_rating, payment_status). List ascending by current_rating, then payment_status. Fields: current_rating, payment_status, loan_count, exposure.
6. monitoring_cadence: monthly if any severe delinquencies (90+ DPD/Nonaccrual), else quarterly.

Task type E: Competing CRE decision (task 005 pattern)

Inputs needed: branch detail, branch metrics, branch loans, sector-exposures, applications (the two named IDs), policies, FDIC benchmark.

Steps:

1. Fetch the two named applications (e.g. HAR-APP-901, HAR-APP-902).
2. For each: compute weighted_cdfi_score (§3.5, 1dp) and score_class (§3.5).
3. applications_compared: list ascending by application_id. Fields: application_id, weighted_cdfi_score, score_class, decision, reason_codes (ascending alphabetically).
4. stress: formula = "stressed_dscr = dscr * 0.85 / (1 + 0.18)", coverage_breach_threshold = 1.00. Results ascending by application_id: application_id, base_dscr (2dp), stressed_dscr (2dp), breaches_threshold (bool).
5. concentration (§3.3): compute existing CRE exposure from loan_type == "CRE" loans, existing + post-approval concentrations, policy variance bps, and FDIC variance (forced metric total_real_estate_30_89_pct).
6. recommended_path: select the application with the BETTER (lower) weighted_cdfi_score and that doesn't breach the stress threshold. path = approve/conditional_approve/decline/defer/participation_required. unselected_application_id, unselected_disposition (decline or defer), unselected_reason_codes (ascending alphabetically, from subset: sector_breach, weak_dscr, high_ltv, fdic_adverse_variance).
7. conditions: ascending alphabetically, from the allowed set. Infer from selected app's risk factors (CRE concentration elevated → committee_cre_exception, no_additional_cre_without_committee_review; DSCR covenant → minimum_dscr_covenant_1_25; CRE collateral → updated_appraisal_before_close, tenant_roll_and_lease_review; monitoring → quarterly_financial_reporting; exposure management → bank_retained_exposure_cap).

5. OUTPUT CONVENTIONS

Precision

Ordering rules (CRITICAL — evaluators check ordering)

• loan_ids in any list: ascending string sort.
• migration_from_current_rating_3: ascending by final_rating.
• final_rating_exposure_totals: ascending by final_rating.
• material_downgrades: ascending by loan_id.
• decisions (task 002): ascending by application_id.
• concentration_flags (task 002): sort by sector, then application_id.
• post_approval_concentrations (task 002): ascending by sector.
• applications_compared (task 005): ascending by application_id.
• stress.results (task 004, 005): ascending by loan_id / application_id.
• workout_queue (task 004): descending exposure, then ascending loan_id (this one is DIFFERENT).
• severe_bucket_counts (task 004): ascending by current_rating, then payment_status.
• by_action (task 001): ascending by action name.
• risk_classes (task 004): ascending by loan_id.
• escalation_triggers (task 003): ascending by trigger_id.
• peer_states (task 003): ascending state code.
• reason_codes, conditions: ascending alphabetically.
• priority_ranking (task 002): highest priority first (NOT alphabetical).

Complete enum reference

payment_status: Current, 30 Days Past Due, 60 Days Past Due, 90+ Days Past Due, Nonaccrual

decision: approve, conditional_approve, decline, defer, participation_required

conditions (002): participation_required, reduced_amount, board_exception, sba_guaranty_required, startup_monitoring, none

reason_codes (002/005): 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

watch-list action: monitor, watchlist, special-assets, workout, partial_chargeoff_review, legal_referral

CDFI risk_class: Prime, Desirable, Satisfactory, Watch, Doubtful, Projected Loss

CRE score_class: approve_quality, conditional, weak

handling (002): approve, conditional_approve, decline, participation_required, none

posture (003): continue_approving, continue_with_tighter_conditions, temporarily_pause

capacity_status (003): capacity_available, capacity_constrained, no_capacity

external_risk_status (003): stronger_than_national_and_peers, mixed_vs_national_and_peers, weaker_than_national_and_peers

risk_tolerance (003): restrained, moderate, expansive

committee_message (003): capacity_available_but_external_risk_weaker, pause_until_state_metrics_recover, routine_approval_path_supported

direction (003): higher, lower, equal

checklist_gates (003): board_authorization, equipment_invoice, fleet_replacement_plan, payer_contract_summary, public_contract_or_tax_support, proof_of_insurance, ucc_or_title_lien

operating_controls (003): 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 conditions (003): segment_recent_delinquency_ge_90_bps, missing_insurance_or_lien_exception, quarterly_capacity_exceeded_or_exception_requested, state_delinquency_gap_widens_25_bps

escalation owners (003): credit_risk_manager, operations_control_manager, lending_committee_chair

monitoring_cadence (004): monthly, quarterly, semiannual

CRE conditions (005): 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

unselected_disposition (005): decline, defer

unselected_reason_codes (005): sector_breach, weak_dscr, high_ltv, fdic_adverse_variance

FDIC benchmark_metric (005) — FORCED: total_real_estate_30_89_pct

FDIC benchmark_metric (001) — choose: total_loans_noncurrent_pct, total_real_estate_noncurrent_pct, construction_development_noncurrent_pct

6. PITFALLS & EDGE CASES

6.1 Regrade population vs watch-list

• Task 001 regrades loans rated 3 or worse (min_current_rating=3, i.e. rating >= 3).
• Task 004 watch-list is loans rated 6 or worse (min_current_rating=6, i.e. rating >= 6).
• These are DIFFERENT populations — do not mix them. The target_current_rating_min / adverse_rating_min field records which threshold was used.

6.2 Severe-delinquency override

• The delinquency minimum is a FLOOR, not a replacement. Compute DSCR rating and LTV rating independently, THEN take the max with the delinquency floor.
• A loan with DSCR=1.79 (rating 3) and payment_status="Nonaccrual" (floor 8) gets final_rating=8, NOT 3. The override always wins when delinquency is severe.
• 90+ DPD forces minimum rating 7; Nonaccrual forces minimum rating 8. These are the most common override cases.

6.3 Null factors in rating re-derivation

• When dscr is null (HELOC, some residential mortgages), skip the DSCR factor entirely.
• When ltv is null, skip the LTV factor.
• When BOTH are null but payment_status is "Current" (no delinquency floor), the loan retains its current_rating (cannot re-derive). It should NOT appear in migration tables as a downgrade.
• Always check for nulls before applying thresholds.

6.4 Concentration denominator

• Use total_loans_outstanding (from branch metrics), NOT total_assets. This is the most common error. The sector_ceiling_pct and cre_policy_limit_pct are fractions of the loan portfolio, not of total assets.
• total_loans_outstanding = sum of all loan outstanding_balances = branch metrics total_loans_outstanding (confirmed identical).

6.5 CRE exposure measurement

• Use loan_type == "CRE" from the loans endpoint as the authoritative CRE exposure. The sector-exposures table categorizes by sector name, and not all real-estate sectors carry the cre_policy_limit_pct as their limit_pct. Some CRE sectors (Construction, Multifamily) may show sector_ceiling_pct as their limit, creating an undercount if you filter by limit_pct == cre_policy_limit_pct.

6.6 Variance sign convention

• variance_ratio = branch_ratio - benchmark_ratio (branch minus benchmark).
• variance_bps = variance_ratio * 10000.
• Positive = branch is WORSE (higher NPA/delinquency than benchmark = adverse).
• Negative = branch is BETTER than benchmark.
• The sign matters — do not use absolute value.

6.7 NCUA direction interpretation

• For delinquency_bps: higher = WORSE (more delinquent).
• For loan_to_share_pct: higher = riskier (more leveraged), but direction is just "higher/lower" — report the factual direction.
• For roaa_bps: higher = BETTER (more profitable).
• For positive_net_income_pct: higher = BETTER.
• When determining external_risk_status, weigh: if NC is worse (higher delinquency, lower roaa, lower positive_net_income) than BOTH US and peers on most metrics → weaker_than_national_and_peers. If mixed → mixed_vs_national_and_peers.

6.8 Application ID verification

• When a prompt names specific application IDs (e.g. "HAR-APP-901 and HAR-APP-902"), fetch ALL applications for the branch and verify those IDs exist. The -9xx scenario IDs always exist but sort after normal IDs — never slice a list and assume you have all records.
• If named IDs are missing (should not happen with well-formed tasks), fall back to the applications matching the described criteria (e.g. the two CRE applications with the highest requested amounts).

6.9 priority_ranking includes conditional approvals

• The priority_ranking list (task 002) includes BOTH approve AND conditional_approve decisions — NOT just clean approvals. Declined and deferred applications are excluded.

6.10 workout_queue ordering is DIFFERENT

• Unlike most lists (ascending ID), the workout_queue is ordered by descending exposure first, then ascending loan_id as a tiebreaker. This is the one list where the largest exposures come first.

6.11 Factor score edge cases for CDFI

• fico = null: assign max penalty score 5 (missing credit history = risk).
• liquidity_months = null (applications): assign max penalty 5 or use existing_relationship_years as a partial proxy.
• When computing debt_to_asset for applications: debt_to_asset = total_debt / total_assets. If total_assets is 0 or null, assign max penalty 6.
• A loan can have risk_class = "Projected Loss" ONLY if factor_score >= 19 AND ltv > 1.0. If factor_score >= 19 but ltv <= 1.0, the class is Doubtful.

6.12 Quarter selection

• Branch metrics returns TWO quarters (2025Q1 and 2024Q4). Always use 2025Q1 (the latest) unless the task explicitly specifies a prior quarter. Using the wrong quarter produces wrong NPA ratios, wrong total_loans_outstanding, and wrong concentrations.

6.13 SBA guaranty as a mitigant

• If an application has sba_guaranty_pct > 0, this is a strong mitigant. A loan that would otherwise decline for weak_dscr or high_ltv may become conditional_approve with condition sba_guaranty_required. The SBA guaranty reduces the bank's effective risk.

6.14 Grandfathered sectors

• If a sector in sector-exposures has grandfathered == 1, the existing over-ceiling exposure is allowed to remain. BUT new approvals that would worsen (increase) that sector's concentration require a mitigation (participation_required, reduced_amount, board_exception) or must be declined with sector_breach.

6.15 NCUA peer median calculation

• Peer median = median of the NCUA benchmark values for the segment's peer_states. For 3 peer states, this is the middle value when sorted. For an even number of peers, average the two middle values.
• Compare the segment's state value to this median for the nc_vs_peer_median direction fields.

6.16 Escalation trigger threshold checks

• segment_recent_delinquency_ge_90_bps: fires when internal_context.recent_delinquency_bps >= 90. Check the actual value (e.g. 86 < 90 → NOT currently triggered, but should still be listed as an escalation trigger to monitor).
• state_delinquency_gap_widens_25_bps: compare segment state delinquency_bps vs US delinquency_bps. If the gap >= 25 bps → triggered. NC (79) vs US (58) = 21 bps gap → NOT yet triggered, but monitor.

7. SOLVING WORKFLOW (step-by-step)

1. Read the prompt — identify the task type (A-E), target branch_id or segment_id, and any named IDs.
2. Read the answer_template.json — catalog every required field, its type, precision, enum constraints, and ordering rule. The template is the contract.
3. Fetch policies (/api/policies) — load all scoring tables, thresholds, and formulas.
4. Fetch the branch/segment detail and branch metrics (latest quarter).
5. Fetch loans (with appropriate ?min_current_rating= filter) or applications (with ?loan_type= filter). Get COMPLETE lists.
6. Fetch sector-exposures and benchmarks (FDIC and/or NCUA as needed).
7. Compute each answer field following the playbooks above. Double-check precision and ordering.
8. Assemble the JSON output matching the template's required_top_level_keys exactly. Use only allowed enum values.
9. Verify: all lists are correctly ordered, all numbers have correct precision, all enums are from the allowed set, no extra/missing top-level keys.
10. Output only valid JSON — no narrative text outside the JSON.