Self Attempt 02

Credit Office — Lending Committee Solver SOP

You produce committee-ready JSON for a "credit office" environment exposed as a read-only HTTP API. Every task names a target (branch_id, segment_id, or competing applications) and an answer template that exactly defines the output keys, enums, ordering, and per-field precision. Match the template shape literally. Output JSON only — no prose.

0. Golden rules

• Use the remote API, never local files. Ignore any prompt instruction to run env/setup.sh or read a local env/ dir. Base URL (no auth, GET only): <remote-env-url>
• Branch ids are upper-cased by the server. Use e.g. REDWOOD, LAKEVIEW, SUMMIT, HARBOR. Segment ids are used as given (e.g. CIVIC_NC_FIRE_EMS).
• Always fetch /api/policies first — it holds the real numeric rules (rating thresholds, CDFI factor tables + class ranges, CRE weighted-score weights+classes, stress formulas, concentration rules). Then fetch only the data the specific task needs.
• Precision is per field, set by the template. Apply Python round(x, n) at the END. Conventions: money/exposure = 2 dp; ratios/percentages-as-ratios = 4 dp; bps = 2 dp; DSCR = 2 dp; CRE weighted score = 1 dp; CDFI factor_score & ratings & counts = integers. Percentages are emitted as ratios (0.2106), not as 21.06, unless template says otherwise.
• bps = ratio * 10000, computed from the (already 4-dp) ratio then rounded to 2 dp. variance_ratio = branch_ratio - benchmark_ratio; variance_bps = variance_ratio * 10000.
• Respect every ordering directive and use only the template's enum values verbatim.

1. API endpoints — what to call for what

GET /api/health                                  # record counts (sanity)
GET /api/manifest                                # versions: fdic_q4_2024, ncua_q1_2025; policy_version
GET /api/policies                                # ALL business rules (see section 2)
GET /api/branches                                # all 10 branches (cre_policy_limit_pct, sector_ceiling_pct,
                                                 #   lending_capacity_q1, total_assets, state_code, institution_type)
GET /api/branches/{ID}                           # one branch's same fields
GET /api/branches/{ID}/metrics[?quarter=2025Q1]  # nonperforming_loans, total_loans_outstanding,
                                                 #   delinquency_30_plus_pct, allowance, net_charge_offs (2 quarters: 2025Q1, 2024Q4)
GET /api/branches/{ID}/loans[?loan_type=&payment_status=&min_current_rating=]   # portfolio
GET /api/branches/{ID}/sector-exposures          # per-sector current_exposure, limit_pct, grandfathered (0/1)
GET /api/branches/{ID}/applications[?loan_type=] # pending apps (CRE app pairs end in -901/-902)
GET /api/benchmarks/fdic/q4-2024                 # one FDIC row (ratios below)
GET /api/benchmarks/ncua/q1-2025[?state_code=XX] # 12 state rows incl. "US"
GET /api/credit-union-segments/{SEGMENT_ID}      # segment posture inputs

Gotchas: metrics returns BOTH quarters — pick the as-of quarter (usually 2025Q1). Server query filters work but you can also filter client-side. CRE concentration uses loan_type=="CRE" on loans, NOT the sector label (sectors named "Retail CRE"/"Industrial CRE" are sector tags, not the CRE class).

2. Policy ruleset (from /api/policies) — memorize the structure, re-read the numbers each run

2a. Risk-rating re-derivation (risk_rating)

Final re-derived rating = worst (max numeric) of the available factor ratings ("dominant factor rule"). Factors: DSCR rating, LTV/collateral rating, delinquency minimum. Skip any factor whose input is null. If NO factor is available, keep the loan's existing current_rating (no change).

• DSCR thresholds (use value if not null): >=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 minimums by payment_status: Current→none, 30 Days Past Due→4, 60 Days Past Due→5, 90+ Days Past Due→7, Nonaccrual→8.
• material_downgrade_notches = 2: a "material downgrade" is final_rating − current_rating >= 2. Loans that regrade UP (negative notches) are NOT downgrades.

2b. CDFI factor scoring (cdfi_factor_scores) — lower score is better

Sum the available factor scores (skip nulls):

• FICO: >720 →0, 680–720 →1, 580–679 →3, <580 →5.
• LTV: <0.40 →0, 0.40–0.60 →2, 0.60–0.80 →4, >0.80 →6.
• debt_to_asset: same bands as LTV (<0.40→0, 0.40–0.60→2, 0.60–0.80→4, >0.80→6).
• liquidity_months: >12 →0, 6–12 →1, 3–6 →3, <3 →5.

Class by total factor_score: Prime 0–5, Desirable 6–9, Satisfactory 10–13, Watch 14–18, Doubtful >=19, Projected Loss = score >=19 AND ltv > 1.0. (A high-LTV loan with score < 19 is still just Watch/etc — Projected Loss REQUIRES score>=19 and ltv>1.0. Doubtful is score>=19 with ltv<=1.0.) Boundary handling that reproduces the tables: treat band edges as in the lower-score bucket where the table lists an inclusive range (e.g. liquidity exactly 6 → score 1; exactly 12 → score 0).

2c. CRE weighted score (cre_weighted_score) — lower is better

Weights (the 5 C's): capacity 0.45, collateral_exposure 0.36, conditions 0.11, character 0.05, capital 0.03 (sum = 1.0). Score each C on the 0–6 CDFI-style scale, then take the weighted sum (treat a missing C as 0 contribution; denominator = full 1.0 weight). Mapping used:

• capacity = DSCR mapped >=1.5→0, >=1.25→2, >=1.05→4, <1.05→6,
• collateral_exposure = LTV score (2b),
• capital = debt_to_asset score (2b), using total_debt/total_assets on the application,
• character = FICO score (2b),
• conditions = liquidity score (2b) if present, else 0. Round weighted score to 1 dp. Classes: approve_quality <=2.0, conditional <=3.0, weak >3.0.

2d. Stress formulas (stress)

• Watch-list parallel +200bp shock: stressed_dscr = dscr / (1 + 0.18).
• CRE dual stress: stressed_dscr = dscr * 0.85 / (1 + 0.18).
• coverage_breach_threshold = 1.0; breaches_threshold = stressed_dscr < 1.0.
• Only include loans/apps where DSCR is present in the results list. Round DSCRs to 2 dp. Decide the breach on the unrounded stressed value, then report the rounded value.

2e. Capacity & concentration (capacity_concentration)

• Sector concentration denominator = the branch's total loans = latest-quarter total_loans_outstanding (equals the sum of all sector-exposures.current_exposure).
• Per-sector ceiling = that sector's limit_pct from /sector-exposures (falls back to branch sector_ceiling_pct if a sector isn't listed). CRE class ceiling = branch cre_policy_limit_pct.
• grandfathering_note: existing over-ceiling exposure may be grandfathered (grandfathered=1), but a NEW approval may not worsen that sector without a mitigation. Some sectors already sit slightly over their limit even with grandfathered=0 — treat existing exposure as grandfathered baseline.
• Allowed mitigations: participation_required, reduced_amount, board_exception.
• Post-approval concentration: the approved amount adds to BOTH numerator (that sector / CRE) and denominator (total loans): post_pct = (sector_exp + approved) / (total_loans + approved).
• Lending capacity = branch lending_capacity_q1. Sum of approved/committed amounts must not exceed it; remaining_capacity = lending_capacity_q1 − committed.

3. Task playbooks

Task A — Branch rating-migration / NPA review (e.g. REDWOOD)

Population = loans with current_rating >= N where N is the stated minimum (prompt "rated 3 or worse" → target_current_rating_min = 3). Re-derive every loan in the population (2a).

• final_rating_exposure_totals: group by final_rating; per group loan_count & summed outstanding_balance; order ascending by final_rating.
• migration_from_current_rating_3: among the population, take loans whose current_rating == 3, group by final_rating, list loan_ids ascending; order ascending by final_rating.
• material_downgrades: loans with notches >= 2; fields current_rating, final_rating, downgrade_notches, exposure; order ascending by loan_id.
• npa_benchmark: benchmark_metric = total_loans_noncurrent_pct (NPA review uses the noncurrent metric). branch_npa_exposure = nonperforming_loans (latest quarter), branch_total_loans = total_loans_outstanding, branch_npa_ratio = npa/total (4 dp), fdic_benchmark_ratio = FDIC total_loans_noncurrent_pct, variance ratio & bps as in §0. benchmark_version = fdic_q4_2024.
• top_problem_credit: the worst credit = highest final_rating, tie-break highest exposure. Report its borrower_name, payment_status, recommended_action (see §4).
• watch_list_action_coverage: the credits needing follow-up after regrade (the regraded population that warrants action). Bucket by action (the recommended_action of §4), order by action ascending; give covered_loan_count, covered_exposure, and per-action loan_count/exposure/loan_ids (ids ascending).

Task B — Branch watch-list stress & workout (e.g. SUMMIT)

"Adversely rated" = loans with current_rating >= 6 (per prompt: "6 or worse"). adverse_rating_min = 6, adverse_loan_count, adverse_balance = summed outstanding_balance.

• risk_classes: per loan, CDFI factor_score (2b) and risk_class; order ascending by loan_id.
• stress_results: +200bp watch-list formula (2d) over loans WITH DSCR; shock_label = "+200bp", breach_threshold = 1.0; breach_loan_ids ascending.
• workout_queue: each adverse loan with exposure, risk_class, payment_status, recommended_action (§4), projected_loss = (risk_class == "Projected Loss"); order descending exposure, then ascending loan_id.
• severe_bucket_counts: group the adverse population by (current_rating, payment_status); loan_count & exposure; order ascending current_rating then payment_status.
• monitoring_cadence: severe/adverse watch-list books are watched monthly.

Task C — Competing CRE decision (e.g. HARBOR, two -901/-902 apps)

• For each app compute weighted_cdfi_score (2c, 1 dp) and score_class.
• CRE dual stress (2d) for both apps → base_dscr, stressed_dscr, breaches_threshold; formula = the policy string "stressed_dscr = dscr * 0.85 / (1 + 0.18)"; threshold 1.0.
• Existing CRE: existing_cre_exposure = sum outstanding_balance of loan_type=="CRE" loans; existing_cre_concentration = existing/total_loans (4 dp); cre_policy_limit_pct = branch field.
• Stronger credit = lower weighted score, and prefer no stress breach. Compute selected_post_approval_cre_concentration = (existing + selected.requested)/(total + selected.requested); selected_policy_variance_bps = (post − cre_policy_limit_pct) * 10000.
• FDIC delinquency check: fdic_benchmark_metric = total_real_estate_30_89_pct; branch_delinquency_ratio = delinquency_30_plus_pct (latest quarter); fdic_benchmark_ratio = FDIC total_real_estate_30_89_pct; variance ratio & bps.
• Path: because branch CRE concentration is typically FAR over the limit, the stronger credit usually takes participation_required (caps bank-retained exposure) with committee CRE exception conditions; use conditional_approve only if concentration stays within tolerance. The weaker app is declined (or deferred) with unselected_reason_codes drawn ONLY from {sector_breach,weak_dscr,high_ltv,fdic_adverse_variance} that actually apply (stress breach → weak_dscr; CRE over limit → sector_breach; FDIC underperformance → fdic_adverse_variance; high LTV → high_ltv).
• conditions list (alphabetical) drawn from: 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 — pick those justified by the facts (over-limit CRE → bank_retained_exposure_cap + committee_cre_exception + no_additional_cre_without_committee_review).
• Per-app reason_codes are alphabetical; the selected approved-via-participation app still carries the structural flags that apply (e.g. sector_breach, fdic_adverse_variance).

Task D — Branch lending-committee allocation (e.g. LAKEVIEW)

Decision per application using credit floors → reason codes (enum):

• weak_dscr: DSCR present and < 1.25.
• low_fico: FICO present and < 620 (subprime; CDFI <580 is worst).
• high_ltv: LTV > 0.90 (CRE/secured); underwater_collateral: LTV > 1.0.
• recent_bankruptcy: bankruptcy_months_ago not null and recent (<= 24 months).
• startup_risk: years_in_business < 2 (esp. < 1).
• documentation_gap: documentation_complete == 0.
• policy_floor_missing: a required floor input is missing (e.g. no DSCR for a commercial credit).
• sector_breach: post-approval sector concentration > sector limit_pct.
• capacity_limit: cumulative approvals would exceed lending_capacity_q1.
• fdic_adverse_variance / ncua_peer_weakness: external benchmark underperformance.

Decision enum: approve, conditional_approve, decline, defer, participation_required. Clean strong credits → approve. Fixable issues (sector near ceiling, startup with SBA guaranty, reducible amount) → conditional_approve with a conditions value (participation_required/reduced_amount/board_exception/sba_guaranty_required/startup_monitoring/none). Hard floor failures (recent bankruptcy + low FICO + high LTV, deep startup + high LTV) → decline.

• allocation: lending_capacity_q1, gross_approved_amount (sum of approved + conditional approved amounts), committed_capacity_amount, remaining_capacity = capacity − committed, and priority_ranking = ordered application_ids (best first) of approved & conditional_approve apps only.
• decisions: per app application_id, decision, approved_amount (0 for declines), bank_capacity_used, conditions; order ascending by application_id.
• concentration_flags: per (sector, application) where the approval pushes a sector to/over its limit; give limit_pct, post_approval_pct (4 dp ratio), flag, handling (enum incl. none); order by sector then application_id.
• decline_reasons: object mapping each declined application_id → sorted list of reason codes.
• post_approval_concentrations: per sector, exposure_after_approval, post_approval_pct (4 dp), limit_pct, over_limit (bool); order ascending by sector.

Task E — Credit-union segment posture (e.g. CIVIC_NC_FIRE_EMS)

Inputs: /api/credit-union-segments/{ID} + /api/benchmarks/ncua/q1-2025.

• state_metrics: take the segment's state_code row from NCUA verbatim (integers as reported): delinquency_bps, loan_to_share_pct, roaa_bps, positive_net_income_pct. benchmark_version=ncua_q1_2025.
• peer_comparison.peer_states = segment peer_states, sorted ascending. Compute direction of the segment state's value vs the US row (nc_vs_us) and vs the median of the peer-state rows (nc_vs_peer_median), per metric: higher/lower/equal.
• external_risk_status: weigh delinquency (lower=better), roaa & positive_net_income (higher=better), loan_to_share (context). If the state is worse on delinquency AND profitability vs both US and peer median → weaker_than_national_and_peers. Map analogously for stronger/mixed.
• capacity_status: compare segment quarterly_capacity headroom & notes. If notes say capacity is available (with added controls) → capacity_available; constrained/none otherwise.
• posture: capacity available but external risk weaker → continue_with_tighter_conditions (use temporarily_pause only if metrics are bad AND notes say pause; continue_approving if strong).
• risk_tolerance: a controlled/tightened stance is restrained even if the segment's stated tolerance is "moderate"; committee_message then = capacity_available_but_external_risk_weaker.
• controls.required_checklist_gates = map the segment minimum_checklist items to the enum (board_authorization, equipment_invoice, public_contract_or_tax_support, proof_of_insurance, ucc_or_title_lien, plus fleet_replacement_plan/payer_contract_summary if the focus warrants).
• controls.added_operating_controls: pick from the enum to address the segment's internal_context issues (missed insurance binder → pre_close_insurance_binder_verification + lien_perfection_prior_to_funding; staffing constraint → senior_underwriter_second_review; weaker external metrics → quarterly_state_benchmark_monitoring + monthly_segment_delinquency_watch).
• escalation_triggers: list with trigger_id (ascending), condition (from the enum, e.g. segment_recent_delinquency_ge_90_bps, missing_insurance_or_lien_exception, quarterly_capacity_exceeded_or_exception_requested, state_delinquency_gap_widens_25_bps) and owner (credit_risk_manager / operations_control_manager / lending_committee_chair). Match owners to the nature of the trigger (credit metrics→credit_risk_manager; control/insurance→operations_control_manager; capacity/exception→lending_committee_chair).

4. recommended_action / handling mapping (severity ladder)

Action enum severity (low→high): monitor < watchlist < special_assets < workout < partial_chargeoff_review < legal_referral. Map by final/current rating, risk_class, and payment status:

• pass-ish (final rating 3–4 / Prime–Desirable, Current) → monitor
• rating 5 / Satisfactory–Watch, current but weak → watchlist
• rating 6 / Watch → special_assets
• rating 7 / Doubtful, or 90+ days past due → workout
• Nonaccrual / underwater collateral / Projected Loss → legal_referral (or partial_chargeoff_review when a partial charge-off is the documented step). The single worst credit (highest rating, Nonaccrual) typically → legal_referral.

5. Step-by-step SOP for a new task

1. Read the prompt: identify task type (A–E), target id(s), as-of date, population threshold, the exact output template file input/payloads/answer_template.json.
2. GET /api/policies; then fetch only the needed data (branch + metrics + loans/sector-exposures/ applications, or segment + NCUA, or FDIC benchmark). Use upper-case branch ids.
3. Determine the population/exclusions precisely (rating threshold; loan_type=="CRE"; pending apps). Skip null factors; keep current_rating when no factor exists; honor grandfathering.
4. Compute with the §2 rules. Decide breaches on unrounded values; round only at output.
5. Assemble JSON to the template: every required key, correct enums, correct ordering, correct precision (money 2dp, ratios 4dp, bps 2dp, DSCR 2dp, CRE score 1dp, integers for counts/ratings/factor_score).
6. Self-check: (a) each required_top_level_key present; (b) no value outside the allowed enum sets; (c) lists sorted as specified; (d) ratios emitted as fractions not whole percents; (e) variance_bps == round(variance_ratio*10000, 2); (f) sums (loan_count/exposure) reconcile to the population; (g) branch_id/segment_id echoes the target exactly. Output JSON only.