Reflect 3 Attempt 01

Credit-Risk Committee Solver Skill (task_group_011)

A self-contained playbook for producing committee-ready JSON answers against the shared credit-office public API. A solver sees only: a task prompt, the target branch/segment identifier, an answer_template.json schema, this skill, and the public API. Apply the workflow rules below; emit ONLY valid JSON matching the task's template (no narrative).

0. Environment — Public REST API (read-only)

Base URL: <remote-env-url> (remote; do not look for a local server/db). All endpoints return JSON, no auth. Use curl (pipe through jq to shape).

branch_id values are uppercase (REDWOOD, LAKEVIEW, CIVIC-style segments via the segment endpoint). Query-param filters are the documented way to scope loan/application lists (e.g. ?min_current_rating=3, ?min_current_rating=6).

1. Numeric & ordering conventions (universal)

• Money/USD → 2 decimals.
• Ratios / percentages expressed as ratios → 4 decimals (e.g. 0.4695, not 46.95%).
• Basis points → 2 decimals, SIGNED (branch minus benchmark; positive = branch worse than benchmark / over policy limit).
• Counts and rating/score factors → integers (exactly as reported for NCUA state metrics).
• variance_bps = (branch_ratio - benchmark_ratio) * 10000, computed from FULL-PRECISION ratios then rounded to 2 dp. Do NOT recompute bps from already-rounded 4-dp ratios (rounding-then-multiplying loses precision and yields a different bps value). Always carry full precision through the subtraction and the *10000, then round only the final bps to 2 dp.
• Every list that has an ordering rule in its template MUST be sorted that way before output. Common orderings: ascending loan_id (string sort), ascending by final_rating, ascending by action, descending exposure then ascending loan_id, ascending current_rating then payment_status, ascending state code, ascending alphabetically (reason codes / conditions).
• loan_id / application_id are strings → sort lexicographically ("RED-LN-011" < "RED-LN-901").

2. Risk-rating re-derivation (from /api/policies risk_rating)

Re-derive a loan's rating from its objective factors, then take the WORST (max numeric) across available factors. Higher number = worse.

• DSCR thresholds: >=1.5→3, >=1.25→4, >=1.05→5, >=1.0→6, <1.0→7.
• LTV thresholds: <=0.65→3, <=0.75→4, <=0.85→5, <=1.0→6, >1.0→7.
• Delinquency floor (payment_status → minimum rating): 30 Days Past Due→4, 60 Days Past Due→5, 90+ Days Past Due→7, Nonaccrual→8, Current→none.
• final_rating = max( available dscr_rating, available ltv_rating, delinquency_floor ). "Available" = skip a factor only when it is genuinely null (e.g. consumer/HELOC with no DSCR/LTV). If NO factor is available, retain the current rating.
• downgrade_notches = final_rating - current_rating. material_downgrade_notches = 2 (from policy). A loan is a MATERIAL downgrade iff notches >= 2 (downgrades of exactly 1 notch are NOT material). Upgrades (negative notches) are not downgrades.
• Regrade population filter: /loans?min_current_rating=N returns loans with current_rating >= N. "Loans currently rated 3 or worse" ⇒ target_current_rating_min = 3, i.e. ALL rated loans EXCLUDING those rated better than 3 (rating 1/2 exist and are excluded). Do NOT assume the scale starts at 3.
• Severe-delinquency override on the recommended action: a Nonaccrual loan's action is partial_chargeoff_review when its final rating is 8. Confirmed via feedback: using legal_referral for a Nonaccrural/underwater credit LOWERED the score — partial_chargeoff_review is the correct action for rating-8 Nonaccrual. Reserve legal_referral for the most severe escalations only (do not trigger it merely from Nonaccrual). 90+ Days Past Due (floor 7) maps to workout.

3. Watch-list / action mapping (enum: monitor, watchlist, special-assets, workout, partial_chargeoff_review, legal_referral)

Map by the credit's RATING (re-derived final_rating for regrade tasks; current_rating for watch-list stress tasks), with payment_status overrides that coincide with the rating floor:

Payment-status override: 90+ Days Past Due → workout; Nonaccrual → partial_chargeoff_review. (These align with the delinquency floor so the action is consistent with the rating.) Watch-list action COVERAGE includes the FULL regrade/watch-list population (every regraded loan is assigned an action, including monitor for the cleanest) — confirmed: covering all regraded loans scored higher than covering only problem credits. covered_loan_count / covered_exposure = total population; by_action is grouped ascending by action with per-action loan_count, exposure, loan_ids (ascending loan_id). Note: the exact action for the mildest adverse bucket (rating 6 / Satisfactory-class) was not fully resolved by feedback — special-assets (by rating) and monitor/watchlist (by CDFI class) were both attempted; prefer the RATING-based mapping above for consistency with the regrade task.

4. NPA / FDIC benchmark variance

• For a BANK branch NPA review, choose the FDIC metric that matches the portfolio breadth. A mixed-portfolio branch (C&I + CRE + consumer) ⇒ total_loans_noncurrent_pct. The FDIC benchmark object is the single /api/benchmarks/fdic/q4-2024 record.
• branch_npa_exposure = branch nonperforming_loans from the metrics table for the review quarter (this equals the sum of 90+DPD + Nonaccrual loan balances — they reconcile). branch_total_loans = total_loans_outstanding (NOT total_assets — using total_assets is the classic mistake).
• branch_npa_ratio = branch_npa_exposure / branch_total_loans. variance_ratio = branch_npa_ratio - fdic_benchmark_ratio. variance_bps = variance_ratio * 10000, signed, full-precision-then-round.
• For CRE-specific tasks the FDIC metric enum is fixed to total_real_estate_30_89_pct (0.0051). The branch delinquency ratio to compare is the branch's REPORTED delinquency_30_plus_pct from the metrics table (confirmed: using the reported metric scored higher than recomputing a real-estate-30-89 ratio from loan records). fdic_variance_bps = (branch_delinquency_ratio - 0.0051) * 10000, signed.

5. NCUA / credit-union segment posture

• state_metrics = the target state's row from /api/benchmarks/ncua/q1-2025?state_code=.., integers EXACTLY as reported: delinquency_bps, loan_to_share_pct, roaa_bps, positive_net_income_pct. benchmark_version = "ncua_q1_2025".
• peer_states = the segment's peer_states list, sorted ascending state code.
• nc_vs_peer_median: median of the peer_states' NCUA rows (odd count ⇒ middle value after sorting). For each of the 4 metrics, direction higher/lower/equal = NC value vs comparison value (literal numeric compare, NOT better/worse).
• external_risk_status: compare NC vs US AND NC vs peer-median across all 4 metrics. If NC is worse on every metric vs both ⇒ weaker_than_national_and_peers; mixed ⇒ mixed_vs_national_and_peers; better on all ⇒ stronger_than_national_and_peers.
• posture: continue_with_tighter_conditions when capacity is available but external risk is weaker and/or operating-control issues exist (the common case); temporarily_pause only for a control breakdown severe enough to halt; continue_approving only when external risk is strong + controls clean.
• capacity_status: capacity_available when the segment notes/capacity field say so (do NOT flip to capacity_constrained merely from a staffing note — the notes are authoritative). risk_tolerance = the segment's risk_tolerance field (e.g. moderate). committee_message:capacity_available_but_external_risk_weaker pairs with capacity_available + weaker external; pause_until_state_metrics_recover with pause; routine_approval_path_supported with continue_approving + strong.
• required_checklist_gates = the segment's minimum_checklist (a subset of: 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 — map each segment risk signal to a control: insurance-binder control_issue ⇒ pre_close_insurance_binder_verification; "added closing controls" note ⇒ lien_perfection_prior_to_funding; staffing_constraint (single senior underwriter) ⇒ senior_underwriter_second_review; recent delinquency elevated ⇒ monthly_segment_delinquency_watch; state delinquency above national median ⇒ quarterly_state_benchmark_monitoring; throughput/capacity overrun risk ⇒ committee_exception_for_capacity_overrun. Confirmed: adding senior_underwriter_second_review for the staffing constraint improved the score.
• escalation_triggers (one per condition, ascending trigger_id) with owners by governance domain: missing_insurance_or_lien_exception→operations_control_manager; quarterly_capacity_exceeded_or_exception_requested→lending_committee_chair; segment_recent_delinquency_ge_90_bps→credit_risk_manager; state_delinquency_gap_widens_25_bps→operations_control_manager (external-benchmark monitoring). condition_choices / owner_choices are fixed enums.

6. CDFI-style risk classes (factor_score)

From /api/policies cdfi_factor_scores. Four objective factor tables, each scored (higher = worse); MISSING factor ⇒ score its WORST tier (confirmed: skip-missing scored far worse than missing→worst):

• debt_to_asset: <0.40→0, 0.40-0.60→2, 0.60-0.80→4, >0.80→6; missing → 6.
• fico: >720→0, 680-720→1, 580-679→3, <580→5; missing → 5.
• liquidity_months: >12→0, 6-12→1, 3-6→3, <3→5; missing → 5.
• ltv: <0.40→0, 0.40-0.60→2, 0.60-0.80→4, >0.80→6; missing → 6.
• factor_score = SUM of the four sub-scores (worst-tier for missing).
• risk_class from total: Prime 0-5, Desirable 6-9, Satisfactory 10-13, Watch 14-18, Doubtful >=19, Projected Loss >=19 AND ltv>1.0. (The missing→worst rule is what lets an underwater Nonaccrual credit reach >=19 and qualify as Projected Loss; with skip-missing it can never reach 19 — that is why missing→worst is required.)

7. DSCR stress (from /api/policies stress)

Use the policy formula LITERALLY (the policy string is the source of truth, not the +200bp label):

• Watch-list parallel shock +200bp: stressed_dscr = dscr / (1 + 0.18) (= dscr / 1.18). coverage_breach_threshold = 1.0; breach = stressed_dscr < 1.0.
• CRE dual stress: stressed_dscr = dscr * 0.85 / (1 + 0.18) (= dscr * 0.72034). Same threshold 1.0.
• Apply stress ONLY to loans/applications that HAVE a DSCR (dscr non-null). Exclude consumer/residential/HELOC with null DSCR from stress_results (the template says "loans with DSCR available").
• breach_loan_ids (where the schema has it) = ascending loan_id list of loans with stressed_dscr < threshold.
• shock_label = +200bp. formula field = the policy formula string verbatim.

8. Lending capacity & sector concentration (allocation tasks)

• lending_capacity_q1 = branch field. Approve (and conditionally-approved) applications consume bank_capacity_used = their approved amount (full retention unless participation truly reduces retained exposure). Sum cannot exceed capacity → lower-priority approvable applications are declined for capacity_limit (or deferred) once capacity is exhausted. gross_approved_amount = sum of approved_amount over approve+conditional_approve; committed_capacity_amount = sum of bank_capacity_used (same set); remaining_capacity = capacity - committed.
• priority_ranking = ordered application_ids, highest priority first, INCLUDING approved AND conditionally approved only (not declines/defers).
• Sector concentration per application = (sector_exposure + requested_amount) / total_loans_outstanding (denominator = total_loans_outstanding, NOT total_assets). limit_pct comes from the sector-exposure row (may override sector_ceiling_pct, e.g. Healthcare 0.19). Flag (breach) when post > limit. handling = the decision applied to that application (conditional_approve / decline / participation_required / approve / none).
• Post-approval FINAL sector concentrations: exposure_after_approval = existing_sector_exposure + sum(approved amounts in that sector); post_approval_pct = exposure_after_approval / total_loans_outstanding (denominator FIXED at the current total_loans_outstanding for the per-sector final view — confirmed for concentration flags; for CRE-policy-concentration in the CRE task the denominator GROWS by the selected amount, see §9). over_limit = post_approval_pct > limit_pct. Include ALL sectors that have any exposure (existing or newly approved), sorted ascending by sector.
• Sector-breach handling per policy: "Existing over-ceiling exposure may be grandfathered, but new approvals may not worsen that sector WITHOUT mitigation." Allowed mitigations: participation_required, reduced_amount, board_exception. So a sector-breaching STRONG credit is mitigated (conditional_approve / participation_required) rather than auto-declined; an already-breached sector with a weak credit may be declined. (The exact mitigation-vs-decline boundary was not fully pinned by feedback — when capacity is ample, prefer mitigating strong credits; reserve sector_breach decline for weak credits that breach.)
• CRE policy concentration limit = branch cre_policy_limit_pct (e.g. 0.29/0.31). Compare existing + selected CRE exposure to this limit (see §9).

9. CRE competing-application decision

• existing_cre_exposure = sum of outstanding_balance over loans with loan_type == "CRE". existing_cre_concentration = existing_cre_exposure / total_loans_outstanding (4 dp).
• selected_post_approval_cre_concentration = (existing_cre_exposure + selected_amount) / (total_loans_outstanding + selected_amount) — denominator GROWS by the selected amount (confirmed: growing denominator scored higher than fixed). selected_policy_variance_bps = (selected_post_approval_cre_concentration - cre_policy_limit_pct) * 10000, signed.
• CRE weighted score (/api/policies cre_weighted_score): weights capacity 0.45, capital 0.03, character 0.05, collateral_exposure 0.36, conditions 0.11 (sum 1.0). Score each of the 5 Cs 0..5 (0 best, lower is better) from the application's objective factors (capacity≈DSCR, capital≈debt_to_asset=total_debt/total_assets, character≈fico or guarantor-strength when fico null, collateral_exposure≈LTV, conditions≈sector/conditions). weighted_cdfi_score = Σ weight·factor_score (1 dp). score_class: approve_quality if score <= 2.0, conditional if <= 3.0, weak if > 3.0. Select the application with the LOWER score (better credit). (The exact per-factor 0..5 thresholds are not in the policy; derive them consistently so the stronger credit scores lower and lands in the better class. Tolerate residual uncertainty on the exact 1-dp value; the class + selection is the load-bearing output.)
• recommended_path: selected_application_id = stronger (lower-score) application; path = conditional_approve when CRE is already/policy-over-limit (committee exception + conditions) or approve when within policy; unselected_application_id = the other. unselected_disposition = "defer" (NOT decline) — the unselected competing credit is DEFERRED, and unselected_reason_codes are the DEFERRAL reasons drawn from {sector_breach, weak_dscr, high_ltv, fdic_adverse_variance} (sorted alphabetically). Confirmed: switching the unselected disposition from decline to defer raised the score.
• applications_compared[].decision mirrors the path decision for the selected and defer for the unselected. reason_codes per application use the full reason-code enum (sorted alphabetically); the selected (clean) application typically has [].
• conditions (selected credit, ascending alphabetically) from the enum: 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. For a CRE credit over policy limit: include committee_cre_exception and no_additional_cre_without_committee_review at minimum; add minimum_dscr_covenant_1_25, updated_appraisal_before_close, tenant_roll_and_lease_review, quarterly_financial_reporting for a full CRE condition set.

10. Output field & enum reference (from each task's answer_template.json — always re-read it)

Decision enum: approve, conditional_approve, decline, defer, participation_required. Conditions enum (allocation): participation_required, reduced_amount, board_exception, sba_guaranty_required, startup_monitoring, none. (NOTE: conditions is NOT the place for decline reason codes — reason codes go ONLY in decline_reasons. Putting reason codes in conditions is a schema-breaking mistake.) Decline 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. decline_reasons maps each DECLINED application_id → its sorted list of reason codes (only declined apps appear). concentration_flags handling enum: approve, conditional_approve, decline, participation_required, none. Payment-status enum (used in loans/answers): Current, 30 Days Past Due, 60 Days Past Due, 90+ Days Past Due, Nonaccrual. Watch-list/Stress action enum: monitor, watchlist, special-assets, workout, partial_chargeoff_review, legal_referral. CDFI risk-class enum: Prime, Desirable, Satisfactory, Watch, Doubtful, Projected Loss. Posture enum: 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: capacity_available_but_external_risk_weaker, pause_until_state_metrics_recover, routine_approval_path_supported. monitoring_cadence: monthly, quarterly, semiannual.

11. Common misjudgments & exclusion rules (generalized from feedback)

1. Regrade population ≠ watch-list population. "Rated 3 or worse" ⇒ min_current_rating=3 ⇒ ALL rated loans except those rated 1/2. "Adversely rated 6 or worse" (watch-list stress) ⇒ min_current_rating=6. Do not conflate the two filters.
2. Watch-list action COVERAGE = full population. Cover every regraded loan with an action (incl. monitor), not just problem credits. covered_loan_count = the whole regrade/watch-list count.
3. Severe-delinquency action override: Nonaccrual → partial_chargeoff_review (rating 8), NOT legal_referral. Using legal_referral for a Nonaccrual/underwater credit was penalized. 90+DPD → workout.
4. conditions field uses the conditions enum only; decline REASON codes live in decline_reasons. Never cross the two.
5. Concentration denominator = total_loans_outstanding, NEVER total_assets. For per-application concentration flags and existing concentration it is the CURRENT total_loans_outstanding. For the CRE policy variance the post-approval denominator GROWS by the selected amount.
6. variance_bps from full precision, then round. Do not round ratios to 4 dp first and then multiply by 10000.
7. CDFI factor_score: missing factor ⇒ worst tier (dta→6, fico→5, liq→5, ltv→6), NOT skip. Skip-missing was strongly penalized and makes Projected Loss (>=19) unreachable for fico-missing underwater credits.
8. Stress = policy formula literal (dscr/1.18 watch-list; dscr*0.85/1.18 CRE dual). The +200bp is the label, not a per-loan rate add-on. Apply only where DSCR exists.
9. NPA branch ratio uses the metrics table (nonperforming_loans, delinquency_30_plus_pct); do not hand-recompute a different denominator unless the metric definition demands it.
10. Unselected competing credit ⇒ defer with deferral reason codes, not decline.
11. material_downgrades = notches >= 2 only (policy material_downgrade_notches=2); 1-notch downgrades and upgrades are excluded from that list.
12. Always sort lists per the template's ordering before emitting; ascending loan_id is lexicographic string sort.
13. Operating controls are warranted by specific segment signals (control_issue→insurance-binder verification; staffing_constraint→senior-underwriter second review; recent delinquency→monthly delinquency watch; state delinquency above median→quarterly state benchmark monitoring; throughput overrun risk→committee exception for capacity overrun).
14. NCUA directions are literal numeric (higher/lower/equal), not better/worse — a higher delinquency_bps is "higher" even though it is worse.

12. Solving procedure per task

1. GET /api/policies first; load the risk-rating, CDFI, stress, concentration rules into memory.
2. GET /api/branches/{branch_id} + /metrics (latest quarter matching the review date) + /sector-exposures + /loans (with the documented min_current_rating/loan_type/payment_status filter for the task) + /applications. For credit-union tasks GET /api/credit-union-segments/{segment_id} + /api/benchmarks/ncua/q1-2025. For FDIC tasks GET /api/benchmarks/fdic/q4-2024.
3. Re-derive ratings / CDFI classes / stress / concentrations per §§2–9, computing from full precision and rounding outputs to the template's precision (§1).
4. Build the JSON object matching the task's answer_template.json exactly (required top-level keys, required sub-keys, enums from §10, orderings from the template).
5. Validate every list ordering and every enum value before emitting. Emit ONLY the JSON.