Fewshot Attempt 01

Lending-Committee Credit-Risk Skill (task_group_011)

Executable experience for credit-risk / lending-committee committee packets built from the shared credit-office public REST API. A solver sees only a task prompt + an answer_template.json + this file + the live API. The tasks in this family always ask you to pull branch / loan / application / policy / benchmark data, apply a small set of deterministic credit rules, and emit a JSON object whose shape is fixed by the template.

This is a method document, not a per-task answer key. Read the section map, apply the rules, and let the template's enums/ordering/precision drive the output.

0. Environment SOP (READ FIRST)

• Base URL: <remote-env-url> — all endpoints under /api/, JSON, no auth.
• GET only. The environment is read-only for you. Never POST/PUT. Never call /api/judge (no test-time judge is available; it is out of scope and not part of the public surface).
• Always GET /api/health once to confirm the service is up. Use curl -s … | jq to shape JSON.
• Quote any URL containing ?/& in zsh (e.g. curl -s "http://…/loans?min_current_rating=3"), otherwise the shell tries to glob the ? and the call returns nothing.
• branch_id values are uppercase (REDWOOD, LAKEVIEW, SUMMIT, HARBOR, … (uppercase; a task names its own target branch)). Segment ids look like CIVIC_NC_FIRE_EMS.

Endpoint inventory and what each feeds

Query-param semantics: ?min_current_rating=N returns loans with current_rating >= N (ratings are worse-as-they-grow: 1 best … 8 worst). ?loan_type=CRE and ?payment_status= are exact-match filters. ?state_code=NC returns a single-row rows array.

Metric-minding rules

• metrics is a list keyed by quarter (e.g. 2025Q1, 2024Q4). Always pick the latest (highest) quarter for the as-of date in the prompt.
• sector-exposures may carry a per-sector limit_pct override (e.g. Healthcare 0.19 where the branch default sector_ceiling_pct is 0.21). Use the per-row limit_pct, falling back to branch.sector_ceiling_pct for sectors not listed.
• loans carry nullable factors: dscr, ltv, fico, debt_to_asset, liquidity_months, collateral_value can all be null. Handle nulls explicitly (see §3).

1. Numeric & ordering conventions (apply everywhere)

• Currency: 2 decimals (USD), round half-up.
• Ratios / percentages / concentrations: 4 decimals.
• Basis points (variance_bps, delinquency_bps): 2 decimals … except NCUA state_metrics values which are reported as bare integers exactly as the benchmark table gives them (see §6).
• variance_bps is computed from the UNROUNDED variance_ratio, then rounded to 2dp. Do not compute bps from the already-4dp-rounded ratio (that loses precision — e.g. NPA 1037.49 vs the 1037.0 you would get from the rounded 0.1037). Pattern: variance_ratio = branch_ratio - benchmark_ratio (unrounded); variance_bps = variance_ratio * 10000 (round 2dp); then publish variance_ratio rounded 4dp.
• signed variance: branch ratio minus benchmark ratio. A branch worse than benchmark yields a positive variance_bps (more delinquency / more NPA = bad). Keep the sign.
• List ordering is always explicit in the template — read it. Common orderings: ascending loan_id; ascending final_rating; ascending action/sector/payment_status (lexicographic on the enum string); ascending trigger_id; ascending current_rating then payment_status; descending exposure then ascending loan_id (workout queue); ascending application_id; ascending alphabetic for reason-code lists.
• All loan_ids arrays are sorted ascending loan_id (string sort).

2. Risk-rating re-derivation (regrade) — /api/policies → risk_rating

The regrade target population is loans with current_rating >= target_current_rating_min (default min = 3, i.e. "rated 3 or worse"). Pull them with GET /branches/{id}/loans?min_current_rating=3.

Factor → rating tables (from policy risk_rating)

Apply each available factor; null factors are skipped (they do not contribute a rating).

• DSCR (dscr): >=1.5→3, >=1.25→4, >=1.05→5, >=1.0→6, <1.0→7.
• LTV (ltv): <=0.65→3, <=0.75→4, <=0.85→5, <=1.0→6, >1.0→7.
• Delinquency floor (payment_status → minimum rating, the "severe-delinquency override"): Current→null, 30 Days Past Due→4, 60 Days Past Due→5, 90+ Days Past Due→7, Nonaccrual→8.

Dominant-factor (worst-notch) rule

final_rating = max(available factor ratings) — the worst (highest numeric) rating produced by DSCR, LTV, and the delinquency floor wins.

• The delinquency floor acts as a hard floor: a Nonaccrual loan is 8 no matter how strong DSCR/LTV are; a 90+ Days Past Due loan is ≥7.
• If all factor-derived ratings are null/absent (e.g. an unsecured consumer loan with no DSCR, no LTV, and Current), retain the current_rating (no re-derivation possible). You can only worsen or hold — never improve — during a regrade.
• material_downgrade_notches = 2 (policy). A loan whose final_rating - current_rating >= 2 is a material downgrade.

Regrade outputs (per template, e.g. train_001 shape)

• target_current_rating_min, target_loan_count, target_exposure = sum of outstanding_balance over the target population.
• final_rating_exposure_totals: group all target loans by final_rating; one row per final_rating present, ordered ascending final_rating, with loan_count and exposure.
• migration_from_current_rating_3: among target loans whose current_rating == 3 (the threshold entry point), group those that actually moved (final_rating > 3) by final_rating, with loan_ids. Loans that stayed at 3 are shown in the totals but omitted from this migration list (it is a migration — only movers).
• material_downgrades: every target loan with downgrade_notches >= 2, fields {loan_id, current_rating, final_rating, downgrade_notches, exposure}, ordered ascending loan_id. (downgrade_notches = final - current.)
• top_problem_credit: the single loan with the worst final_rating (ties → largest exposure, then lowest loan_id). Include borrower_name, payment_status, and a recommended_action per §10.

Watch-list action coverage (regrade follow-up)

Subset of regraded loans needing follow-up = those with final_rating >= 6. Group by recommended_action (§10 mapping). covered_loan_count/covered_exposure = totals over that subset; by_action ordered ascending by action enum string, loan_ids ascending.

Regrade vs watch-list exclusion: the regrade population is current_rating >= 3 and uses final_rating for action mapping; the watch-list packet (§9) uses current_rating >= 6 (adverse) with no regrade (actions keyed off current_rating). Don't mix the two rating bases.

3. NPA / FDIC benchmark variance

NPA review (train_001-style) uses benchmark metric total_loans_noncurrent_pct (the template also allows total_real_estate_noncurrent_pct, construction_development_noncurrent_pct).

benchmark_version      = "fdic_q4_2024"                       # from branch.fdic_benchmark_set / /api/manifest
benchmark_metric       = one of the template's allowed values (context: NPA → total_loans_noncurrent_pct)
branch_npa_exposure    = metrics[latest].nonperforming_loans
branch_total_loans     = metrics[latest].total_loans_outstanding
branch_npa_ratio       = branch_npa_exposure / branch_total_loans          # 4dp
fdic_benchmark_ratio   = FDIC[benchmark_metric]                            # 4dp
variance_ratio         = branch_npa_ratio - fdic_benchmark_ratio           # 4dp (unrounded for bps)
variance_bps           = variance_ratio * 10000                            # 2dp, from unrounded ratio

Metric → branch-ratio source mapping (pick by benchmark_metric)

• *_noncurrent_pct metrics → branch ratio = nonperforming_loans / total_loans_outstanding.
• *_30_89_pct metrics → branch ratio = metrics[latest].delinquency_30_plus_pct (use the field directly, already a ratio).

A CRE-concentration packet (train_005-style) instead pairs fdic_benchmark_metric = total_real_estate_30_89_pct with branch_delinquency_ratio = metrics[latest].delinquency_30_plus_pct. Same variance math; the branch ratio field name in that template is branch_delinquency_ratio / fdic_variance_*.

4. Lending capacity & allocation (train_002-style)

For a branch's pending-applications allocation packet:

lending_capacity_q1     = branch.lending_capacity_q1          (from /branches/{id})
gross_approved_amount   = sum of approved_amount over APPROVED + CONDITIONAL_APPROVE apps
committed_capacity_amount = sum of bank_capacity_used over those same apps
remaining_capacity      = lending_capacity_q1 - committed_capacity_amount

Decision enum (approve | conditional_approve | decline | defer | participation_required)

Conditions enum: participation_required | reduced_amount | board_exception | sba_guaranty_required | startup_monitoring | none.

Per-application decision logic (apply in priority order):

1. Hard decline reasons first (§8): if any reason code like low_fico, recent_bankruptcy, underwater_collateral, policy_floor_missing, documentation_gap fires → decline.
2. Credit weakness decline: high_ltv AND weak_dscr together, or high_ltv AND startup_risk, generally → decline. A single borderline weakness with strong offsets (e.g. high FICO) may still approve.
3. Capacity: after higher-priority apps are funded, if remaining capacity can't absorb the app and no other weakness exists → decline with reason capacity_limit.
4. Sector/CRE concentration approach (post_approval pct near or over the sector limit_pct) → do not decline outright; mitigate via conditional_approve with condition participation_required (sell the excess to participants) or reduced_amount (approve a smaller amount) or board_exception (committee override).
5. SBA-guaranteed apps (sba_guaranty_pct present) → conditional_approve with sba_guaranty_required; add startup_monitoring when years_in_business is short (< ~2y).
6. Otherwise → approve (condition none).

approved_amount vs bank_capacity_used

• plain approve: approved_amount = requested_amount; bank_capacity_used = approved_amount.
• sba_guaranty_required: approved_amount = requested_amount; bank_capacity_used = approved_amount * (1 - sba_guaranty_pct) (the SBA-guaranteed share is not bank risk/capacity).
• participation_required: approved_amount = requested_amount (the full loan is originated); bank_capacity_used = approved_amount − participated_portion (the sold portion relieves bank capacity). The participated portion is the amount needed to bring the relevant sector/large-exposure back under its ceiling. When LTV>1 / underwater, the bank may also retain only a capped share.
• reduced_amount: approved_amount = min(requested, capacity/headroom); bank_capacity_used = approved_amount.
• decline/defer: approved_amount = 0.0, bank_capacity_used = 0.0, conditions = ["none"].

priority_ranking

Ordered list of application_id, highest priority first, approved + conditional_approve only (declines/defers excluded). Priority is committee priority — strategic relationship/quality first, then core approves, then conditional — not simply by amount. Reverse it for tie-breaks if the template asks ascending.

5. Concentration (sector ceilings & CRE limit)

Sector concentration

• Per-sector limit_pct comes from /sector-exposures; default to branch.sector_ceiling_pct for sectors absent from that table.
• grandfathered=1 sectors may sit over ceiling but new approvals may not worsen them without mitigation (policy capacity_concentration.grandfathering_note).
• Allowed mitigations: participation_required, reduced_amount, board_exception.

Post-approval concentration denominator (IMPORTANT — do not use total_assets)

• "existing" concentration (e.g. CRE): exposure / total_loans_outstanding.
• "post_approval" concentration (sector view after funding approved apps): exposure_after_approval = existing_sector_exposure + approved_amount(s) in that sector; post_approval_pct = exposure_after_approval / (total_loans_outstanding + gross_approved_amount). i.e. the post-approval loan book = current book plus the gross new approvals.
• For a single selected CRE app (train_005): selected_post_approval_cre_concentration = (existing_cre_exposure + selected.requested_amount) / (total_loans_outstanding + selected.requested_amount) — the full requested amount is added (participation is a decision/condition, not a concentration reducer; the bank still originates the full loan).

over_limit = post_approval_pct > limit_pct (strict >; a value exactly at the limit is not over).

existing_cre_exposure derivation

Sum of outstanding_balance over GET /branches/{id}/loans?loan_type=CRE. (This is the loan_type=CRE aggregation, not the sector-exposures table — the two differ for grandfathered/ mixed sectors.)

concentration_flags (per approved app touching a sector)

One row per approved/conditional app whose sector post-approval pct approaches or exceeds the sector limit_pct (within a thin headroom, e.g. ~0.01, or over). Fields: sector, application_id, limit_pct, post_approval_pct (4dp), flag (bool), handling (enum: approve | conditional_approve | decline | participation_required | none). Sort by sector then application_id.

CRE policy variance (train_005)

cre_policy_limit_pct                       = branch.cre_policy_limit_pct
existing_cre_exposure                      = sum(CRE loan balances)
existing_cre_concentration                 = existing_cre_exposure / total_loans_outstanding   # 4dp
selected_post_approval_cre_concentration   = (existing + requested) / (total_loans + requested) # 4dp
selected_policy_variance_bps               = (selected_post - cre_policy_limit_pct) * 10000     # 2dp, unrounded

6. Credit-union segment posture (train_003-style)

Inputs: GET /credit-union-segments/{segment_id} + GET /api/benchmarks/ncua/q1-2025 (+ optional ?state_code=).

state_metrics

Take the segment's state_code row from the NCUA benchmark. Report integer values exactly as in the table (no rounding, no /100): delinquency_bps, loan_to_share_pct, roaa_bps, positive_net_income_pct. benchmark_version = "ncua_q1_2025". (Contrast: ratio fields elsewhere are 4dp — these NCUA fields are integers.)

peer_comparison

• peer_states = segment.peer_states (use verbatim, then sort ascending state code).
• nc_vs_us and nc_vs_peer_median: for each of the four metrics, direction is "higher" | "lower" | "equal" comparing the segment state's value to (a) the US aggregate row and (b) the median of the peer_states rows.
• Note: for delinquency_bps higher = worse; for roaa_bps/positive_net_income_pct higher = better; for loan_to_share_pct higher = more leveraged. The direction is a pure numeric comparison (higher/lower), the interpretation is separate. Compute medians with the usual middle-of-three for the 3 peer rows.

posture (enum: continue_approving | continue_with_tighter_conditions | temporarily_pause)

Decision matrix driven by segment.notes, risk_tolerance, internal_context, and the direction matrix:

• Capacity available and external risk no worse than national/peers → continue_approving.
• Capacity available but external risk weaker (NC higher delinquency, lower ROAA, lower positive-net-income vs both US and peer median) → continue_with_tighter_conditions (add operating controls, don't pause).
• Capacity exhausted or external risk severely adverse / control breakdown → temporarily_pause.

controls

• required_checklist_gates = segment.minimum_checklist (verbatim set from the endpoint), drawn from enum: board_authorization, equipment_invoice, fleet_replacement_plan, payer_contract_summary, public_contract_or_tax_support, proof_of_insurance, ucc_or_title_lien.
• added_operating_controls derive from internal_context:
  • insurance/lien control issue → pre_close_insurance_binder_verification, lien_perfection_prior_to_funding;
  • staffing/senior-underwriter constraint → senior_underwriter_second_review;
  • recent delinquency watch → monthly_segment_delinquency_watch;
  • external state monitoring → quarterly_state_benchmark_monitoring;
  • capacity overrun possibility → committee_exception_for_capacity_overrun. 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. Sort the set.

escalation_triggers

List of {trigger_id, condition, owner}, ordered ascending trigger_id (ET001, ET002, …). Select conditions relevant to the segment's risk profile from: segment_recent_delinquency_ge_ 90_bps, missing_insurance_or_lien_exception, quarterly_capacity_exceeded_or_exception_requested, state_delinquency_gap_widens_25_bps. Owner mapping:

• 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

(Include a trigger as a standing escalation gate for the segment even if the current value is just below its threshold — they are armed gates, not only currently-breached ones. Drop a trigger only when the segment's profile makes it irrelevant — e.g. omit state_delinquency_gap_widens_25_bps when the state gap is already the dominant, not a widening-delta, risk.)

interpretation

• capacity_status: capacity_available | capacity_constrained | no_capacity — based on quarterly_capacity headroom vs current origination flow (capacity_available when the quarterly budget is not exhausted; do not compare current_outstanding stock against quarterly_ capacity flow).
• external_risk_status: stronger_than_national_and_peers | mixed_vs_national_and_peers | weaker_than_national_and_peers — from the nc_vs_us / nc_vs_peer_median direction matrix.
• risk_tolerance: take segment.risk_tolerance (restrained | moderate | expansive).
• committee_message: pick from capacity_available_but_external_risk_weaker | pause_until_state_metrics_recover | routine_approval_path_supported to match the posture.

7. CDFI-style risk classes & factor scoring (watch-list, train_004-style)

From policy.cdfi_factor_scores: score each available factor (null → skip), sum the scores, map the sum to a class. Null factors are skipped, not zeroed.

Factor score tables

(LTV/DTA bands are value ranges; FICO/liquidity are their own scales — use the table literally.)

Class mapping (policy cdfi_factor_scores.classes)

The policy text states Projected Loss = ">=19 and ltv>1.0", but the applied rule escalates any Watch-or-worse score (>=14) with ltv > 1.0 (underwater collateral) to Projected Loss. Apply this override. (ltv>1.0 is the "underwater collateral" trigger.)

watch_list_summary.risk_classes: list of {loan_id, risk_class, factor_score} for the adverse population, ordered ascending loan_id. monitoring_cadence = monthly for an adverse watch-list (rating >= 6); quarterly/semiannual only for healthier populations.

8. Decline reason codes (train_002/005) — derive from application/loan factors

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.

Trigger thresholds (apply to application fields):

• high_ltv — ltv > 0.80 (for business/CRE; consumer may tolerate slightly more with strong FICO).
• weak_dscr — dscr < 1.25 (below the "pass" pass rating 4 threshold). dscr is None on a product that doesn't use DSCR (consumer/mortgage) is not weak_dscr.
• low_fico — fico < 580.
• recent_bankruptcy — bankruptcy_months_ago present and < ~24 months.
• startup_risk (decline) vs startup_monitoring (condition): years_in_business < ~2 is startup; pair with strong credit → startup_monitoring condition on a conditional approve; pair with another weakness → startup_risk decline reason.
• underwater_collateral — ltv > 1.0 (collateral value below loan).
• capacity_limit — the only reason when an otherwise-acceptable app is declined because Q1 capacity is consumed by higher-priority apps.
• sector_breach — approving would push the sector (or CRE aggregate) over its limit_pct.
• fdic_adverse_variance — branch's FDIC benchmark variance is adverse (positive bps, branch worse than benchmark) for the relevant metric; an ambient reason code on CRE/NPA-sensitive apps.
• ncua_peer_weakness — credit-union segment where the state is weaker than national/peers.
• policy_floor_missing / documentation_gap — missing documentation_complete or required checklist gate.

decline_reasons maps each declined application_id → sorted (ascending alphabetic) list of reason codes. For "competing credit" tasks (train_005), unselected_reason_codes is restricted to sector_breach | weak_dscr | high_ltv | fdic_adverse_variance. The applications_compared[].reason_ codes on a defer decision mirrors the same list. Always sort reason-code lists ascending alphabetically.

9. Watch-list stress & workout (train_004-style)

Adverse population = current_rating >= 6 (parameter adverse_rating_min, default 6). Pull via GET /branches/{id}/loans?min_current_rating=6. No regrade here — use current_rating directly.

+200bp DSCR stress (policy stress)

shock_label          = "+200bp"
formula              = stressed_dscr = dscr / (1 + 0.18)          # watch_list_formula
breach_threshold     = 1.0                                          # policy stress.coverage_breach_threshold
stressed_dscr        = dscr / 1.18                                   # 2dp
breaches_threshold   = stressed_dscr < 1.0                          # strict <

stress_results.results lists only loans with dscr available (dscr is not None), ordered ascending loan_id. breach_loan_ids = those with breaches_threshold true, ascending loan_id.

CRE dual stress (train_005-style) — when comparing CRE applications

formula              = "dscr * 0.85 / 1.18"     # policy cre_dual_stress_formula
                              (= dscr * 0.85 / (1 + 0.18))
coverage_breach_threshold = 1.0
stressed_dscr        = dscr * 0.85 / 1.18        # 2dp
breaches_threshold   = stressed_dscr < 1.0

stress.results ordered ascending application_id.

workout_queue

All adverse loans, each {loan_id, exposure, risk_class, payment_status, recommended_action, projected_loss}. Order descending exposure, then ascending loan_id.

• exposure = outstanding_balance.
• risk_class from §7.
• recommended_action from §10 (keyed off current_rating in watch-list packets).
• projected_loss = true iff risk_class == "Projected Loss", else false.

severe_bucket_counts

Group the adverse population by (current_rating, payment_status); one row per non-empty bucket. Fields {current_rating, payment_status, loan_count, exposure}. Order ascending current_rating, then payment_status (lexicographic on the enum string — "90+ Days Past Due" sorts before "Current" because '9'(0x39) < 'C'(0x43)).

10. recommended_action mapping (the action enum)

Enum: monitor | watchlist | special-assets | workout | partial_chargeoff_review | legal_referral.

The primary mapping is by the governing rating (which is final_rating after a regrade, or current_rating for a pure watch-list packet) combined with payment status:

(workout and legal_referral apply to active-restructuring / fraud-litigation situations; in the train data the 6/7/8 mapping above covers standard adverse credits. legal_referral is reserved for fraud/litigation indicators; workout for loans already in active restructuring.)

top_problem_credit.recommended_action and each by_action bucket use this enum, sorted ascending by action string in by_action.

11. Competing-CRE decision (train_005-style) synthesis

Given two competing CRE applications at one branch:

1. Compute each app's weighted CDFI score (policy cre_weighted_score):
   • Weights: capacity 0.45, collateral_exposure 0.36, conditions 0.11, character 0.05, capital 0.03 (capacity + collateral dominate at 0.81 combined).
   • Score each of the 5 Cs 1-5 from application attributes (capacityDSCR, collateralLTV, capitaldebt_to_asset = total_debt/total_assets, characterfico/ co_guarantor_strength/years_in_business/prior_delinquencies, conditions~proposed_rate/ sector/documentation/purpose). Lower = better.
   • weighted_cdfi_score = Σ weight_c × score_c (1 decimal).
   • score_class: <=2.0 → approve_quality, <=3.0 → conditional, >3.0 → weak.
2. CRE dual stress both apps (§9). Flag weak_dscr reason if stressed_dscr < 1.0.
3. Reason codes per app: fdic_adverse_variance (branch FDIC variance adverse — ambient on both), sector_breach if the app's sector/CRE would breach, weak_dscr if base or stressed DSCR weak, high_ltv if ltv > 0.80.
4. Decision per app: approve | conditional_approve | decline | defer | participation_required.
   • approve_quality score + no breach + stress passes → approve.
   • conditional score but CRE concentration breaches → participation_required (sell excess; keep the better credit).
   • weak score + stress breach → defer (collect more info / sponsor support), or decline if a hard reason (low_fico/recent_bankruptcy/underwater) fires.
5. recommended_path: selected_application_id = the stronger credit (lower weighted score, better stress, fewer reason codes). path = the selected app's decision enum. unselected_application_id, unselected_disposition ∈ {decline, defer} (defer when the credit is merely weak/not-yet-declinable; decline only on a hard reason). unselected_reason_ codes (restricted enum §8) sorted ascending alphabetic.
6. concentration block per §5. fdic_* fields use total_real_estate_30_89_pct + delinquency_30_plus_pct.
7. conditions set 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. Choose the set matching the selected path (participation → bank_retained_exposure_cap; existing CRE already over policy limit → committee_cre_exception + no_additional_cre_without_committee_review; CRE deal → updated_appraisal_before_close + tenant_roll_and_lease_review + minimum_dscr_covenant_1_25 + quarterly_financial_reporting). Sort ascending alphabetic.

12. Common misjudgments & exclusion rules (guardrails)

• Regrade population vs watch-list population are different. Regrade = current_rating >= 3 using final_rating for actions. Watch-list = current_rating >= 6 using current_rating for actions. Do not apply final_rating to a watch-list packet or current_rating to a regrade's action mapping.
• Severe-delinquency override: Nonaccrual → rating 8 and 90+ Days Past Due → rating ≥7 are floors, not suggestions. A Nonaccrual loan never gets a final_rating below 8 even with DSCR 2.0 and LTV 0.4.
• Null factors are skipped, never zeroed (regrade dominant-factor, CDFI factor_score). A loan with dscr=None, ltv=None, payment_status=Current keeps its current_rating.
• migration_from_current_rating_3 lists only movers (final_rating > 3) among loans whose current_rating == 3. Loans that stayed at 3 appear in final_rating_exposure_totals but not in the migration list.
• material_downgrades uses >= 2 notches (policy material_downgrade_notches), across the entire regrade target population (not just current_rating==3 loans).
• Concentration denominator is the loan book, never total_assets. Existing → total_loans_outstanding; post-approval → total_loans_outstanding + gross_approved_amount (or + selected.requested_amount for a single selected CRE app).
• existing_cre_exposure = sum of loan_type=CRE loan balances, not the sector-exposures table (they diverge for grandfathered/mixed sectors).
• variance_bps from the UNROUNDED ratio, then round 2dp. Computing from the 4dp-rounded ratio is a precision bug.
• Signed variance: branch worse-than-benchmark → positive bps (the common case for these branches). Keep the sign; don't absolute-value.
• post_approval_pct > limit_pct is over_limit (strict >). Exactly-at-limit is not over.
• NCUA state_metrics are bare integers, not 4dp ratios. delinquency_bps, loan_to_share_ pct, roaa_bps, positive_net_income_pct are reported as-is.
• bank_capacity_used ≠ approved_amount for SBA/participation: SBA → multiply by (1 - sba_guaranty_pct); participation → minus the sold portion. But concentration uses the full approved/requested amount (the loan is originated in full).
• severe_bucket_counts payment_status ordering is lexicographic on the enum string — 90+ Days Past Due comes before Current (digit before letter in ASCII), not by delinquency severity.
• priority_ranking excludes declines and defers — approved + conditional only.
• conditions: ["none"] for plain approve and for decline/defer (conditions attach to approval paths, not declines). Declines carry their reasons in decline_reasons, not in conditions.
• GET only / never /api/judge — the judge is not a public endpoint for you.
• Quote URLs with ?/& in zsh.

13. Output field & enum quick-reference (consolidated)

• payment_status: Current | 30 Days Past Due | 60 Days Past Due | 90+ Days Past Due | Nonaccrual.
• recommended_action: monitor | watchlist | special-assets | workout | partial_chargeoff_review | legal_referral.
• risk_class (CDFI): Prime | Desirable | Satisfactory | Watch | Doubtful | Projected Loss.
• decision: approve | conditional_approve | decline | defer | participation_required.
• conditions: participation_required | reduced_amount | board_exception | sba_guaranty_required | startup_monitoring | none.
• handling (concentration_flags): approve | conditional_approve | decline | participation_required | none.
• 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.
• CRE conditions: 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.
• posture: continue_approving | continue_with_tighter_conditions | temporarily_pause.
• benchmark_version strings: fdic_q4_2024, ncua_q1_2025.
• FDIC benchmark_metric values: total_loans_noncurrent_pct | total_real_estate_noncurrent_pct | construction_development_noncurrent_pct | total_real_estate_30_89_pct (use only those the template's enum allows).

14. Per-family execution checklist (transfer to unseen tasks)

1. Parse the prompt for: branch_id / segment_id, as-of date, target rating threshold, specific application_ids, and which section family (regrade / allocation / segment-posture / watch-list / competing-CRE).
2. GET /api/policies once and cache the rule tables (risk_rating, cdfi_factor_scores, cre_weighted_score, stress, capacity_concentration).
3. Pull the branch, latest-quarter metrics, loans (with the right min_current_rating / loan_type filter), applications, sector-exposures, and the relevant benchmark (fdic_q4_2024 / ncua_q1_2025) — or the segment — as the family requires.
4. Apply the rule sections above in order; compute every numeric field with the §1 precision rules (remember: bps from unrounded ratio).
5. Build the JSON object matching the template's required keys, ordering, enums, and precision exactly. Sort every list per its template ordering clause. Strip any key not in the template.
6. Re-check guardrails (§12): no total_assets denominator, no rounded-ratio bps, regrade-vs- watchlist rating basis, null-factor skip, severe-delinquency floor, over_limit strict >.
7. Emit only the JSON object (no narrative outside it) unless the prompt allows commentary.