Reflect 3 Attempt 02

Skill: Credit-Risk / Lending-Committee Answer Generation

Reusable workflow rules distilled from a 3-round reflect loop on 5 credit-risk train tasks (branch rating migration, Q1 allocation package, credit-union segment posture, watch-list stress packet, competing CRE decision). These are WORKFLOW RULES, not answers. Apply them to any task in this family.

1. Remote API SOP

The environment is a REMOTE read-only HTTP API. Never look for local env/ source, DB files, or a local server.

• Base URL: <remote-env-url> (all endpoints under /api)
• Call with curl -s ... | jq .... No auth. GET only.
• Endpoints used:
  • /api/health, /api/manifest (versions, policy_version, record counts)
  • /api/policies (capacity/concentration, risk-rating, cdfi_factor_scores, cre_weighted_score, stress) — READ THIS FIRST, it drives every rule below.
  • /api/branches, /api/branches/{branch_id} (assets, lending_capacity_q1, sector_ceiling_pct, cre_policy_limit_pct, state_code, institution_type)
  • /api/branches/{branch_id}/metrics?quarter= (per-quarter: total_loans_outstanding, nonperforming_loans, delinquency_30_plus_pct, allowance, charge-offs, deposits)
  • /api/branches/{branch_id}/loans (filter ?loan_type=, ?payment_status=, ?min_current_rating=). Fields: outstanding_balance, dscr, ltv, debt_to_asset, liquidity_months, fico, current_rating, payment_status, guarantor_strength, sector.
  • /api/branches/{branch_id}/sector-exposures — fields are current_exposure, limit_pct, grandfathered (NOT "exposure"/"concentration_pct").
  • /api/branches/{branch_id}/applications (pending apps)
  • /api/benchmarks/fdic/q4-2024, /api/benchmarks/ncua/q1-2025?state_code=
  • /api/credit-union-segments/{segment_id}
• Fetch the policy JSON and both benchmark sets ONCE and reuse. Use the LATEST quarter metrics (e.g. 2025Q1) unless the task names a quarter.

2. Risk-Rating Re-derivation (rating-migration tasks)

Policy risk_rating. Re-derive the final rating for each loan in the regrade population and NEVER substitute a "sensible" override — apply the rule strictly.

• Regrade population = loans with current_rating >= target_current_rating_min (task-dependent, e.g. >= 3). Pass loans (rating <= 2) are EXCLUDED.
• Compute a factor rating from each AVAILABLE factor; the final rating is the WORST (max numeric) of the available factor ratings:
  • 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 minimums (a FLOOR that is itself a factor): Current→(no floor), 30 DPD→4, 60 DPD→5, 90+ DPD→7, Nonaccrual→8
• "Available" = factor is non-null. A loan whose ONLY available factor is the delinquency floor STILL takes that floor as its final rating (do NOT keep the current rating to avoid an upgrade). Confirmed: overruling this dropped the score. Only when NO factor is available at all do you carry the current rating.
• Material downgrade = final_rating - current_rating >= 2 notches (risk_rating.material_downgrade_notches).
• final_rating_exposure_totals: ALL regrade-pop loans grouped by final_rating (ascending), no loan_ids.
• migration_from_current_rating_3: the rating-3 stratum — loans whose CURRENT rating == 3, grouped by final_rating, WITH loan_ids (ascending). (If the field appears elsewhere with a threshold suffix, treat "_3" as referring to the current-rating-3 stratum, not the whole population.)

3. Watch-list Action Coverage & Workout Actions

• Watch-list coverage population = loans with final_rating >= 5 (Watch and below). Pass loans (final 3-4) are NOT covered. Confirmed by feedback.
• Recommended-action ladder by final rating (ascending severity):
  • 5 → watchlist
  • 6 → special-assets
  • 7 → workout
  • 8 → partial_chargeoff_review (Nonaccrual / underwater-collateral cases)
  • (3-4 → monitor, only if a pass loan must be listed)
• top_problem_credit = the worst final rating (highest), break ties by exposure; for ties use the Nonaccrual/underwater loan. Its recommended_action must match the coverage action for the same loan.
• Adverse/watch-list population for stress tasks = current_rating >= 6 ("6 or worse"). severe_bucket_counts summarises payment-status counts by rating bucket — include the ratings the task calls "severe" (verify 6-vs-7,8 scope; the natural read is the same adverse population, but "severe" may mean 7-8 only — check against the task wording).
• monitoring_cadence: monthly for severe/workout lists; quarterly for a watch-list review packet. Pick the one matching the packet's severity.

4. CDFI Factor Scores & Risk Classes (watch-list / CRE scoring)

Policy cdfi_factor_scores. Factor tables (fico, ltv, debt_to_asset, liquidity_months) each map a value to a 0-6 score; sum them into a factor_score.

• MISSING factor → assign the WORST score for that factor (penalise missing data): fico null→5, ltv null→6, debt_to_asset null→6, liquidity_months null→5. Confirmed: switching null→0 to null→worst raised the score materially.
• 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 ranges/scores as ltv
• liquidity_months: >12→0, 6-12→1, 3-6→3, <3→5
• Risk class from factor_score: Prime 0-5, Desirable 6-9, Satisfactory 10-13, Watch 14-18, Doubtful >=19, Projected Loss = >=19 AND ltv>1.0.
• projected_loss boolean (workout queue) = risk_class == "Projected Loss".

5. Stress Formulas

Policy stress. Two formulas — use the one the task names:

• Watch-list +200bp parallel shock (watch-list stress packets): stressed_dscr = dscr / (1 + 0.18); shock_label = "+200bp"; coverage_breach_threshold = 1.0; breach = stressed_dscr < 1.0. Include only loans WITH a DSCR (skip null DSCR).
• CRE dual-stress (competing-CRE decisions): stressed_dscr = dscr * 0.85 / (1 + 0.18); threshold 1.0. Put the formula string verbatim from the policy into the formula field when the template has one.

6. CRE Concentration & FDIC Variance (competing-CRE / concentration tasks)

• cre_policy_limit_pct and sector_ceiling_pct come from the branch record; per-sector limit_pct comes from /sector-exposures.
• DENOMINATOR RULE: every concentration/delinquency ratio divides by total_loans_outstanding (from metrics) — NOT total_assets, NOT total_deposits. This is a frequent error.
• existing_cre_concentration = sum(CRE loan balances) / total_loans_outstanding.
• post_approval concentration (per sector) = (existing_sector_exposure + sum of approved amounts in that sector) / total_loans_outstanding.
• selected_post_approval_cre_concentration = (existing_cre + selected_app.requested_amount) / total_loans_outstanding.
• FDIC/NCUA variance:
  • benchmark_version: fdic_q4_2024 / ncua_q1_2025 (from manifest).
  • For a bank NPA review use benchmark_metric = total_loans_noncurrent_pct (FDIC 0.0098); for a CRE real-estate delinquency compare use total_real_estate_30_89_pct (FDIC 0.0051) — match the template's allowed enum.
  • branch_*_ratio = branch exposure / total_loans_outstanding.
  • variance_ratio = branch_ratio - benchmark_ratio (SIGNED; positive = branch worse than benchmark).
  • variance_bps = (branch_ratio_RAW - benchmark_ratio_RAW) * 10000, rounded to 2dp. ALWAYS compute bps from the RAW (unrounded) variance, not from the 4dp-rounded ratio. Confirmed: this fix alone raised a score.
• NCUA segment posture: state_metrics are the integer values exactly as the NCUA row reports (delinquency_bps, loan_to_share_pct, roaa_bps, positive_net_income_pct). peer_comparison directions (higher/lower/equal) are NC-vs-US and NC-vs-peer-MEDIAN (median of the segment's peer_states).

7. Credit-Union Segment Posture (segment tasks)

• posture: continue_with_tighter_conditions when capacity is available but external state risk is weaker and there are control issues (the typical case when notes say "capacity remains available when added closing controls are used"). temporarily_pause only if metrics must recover first; continue_approving only for a strong segment.
• controls.required_checklist_gates = the segment's minimum_checklist set (from the segment endpoint). Do NOT add gates that aren't in the segment's list just because they appear in the template's universe (e.g. fleet/payer gates belong to other segments unless the segment lists them).
• controls.added_operating_controls: add controls justified by the segment's internal_context and notes — e.g. pre_close_insurance_binder_verification (missed insurance binder), lien_perfection_prior_to_funding (lien closing control), quarterly_state_benchmark_monitoring + monthly_segment_delinquency_watch (state/segment delinquency above median), senior_underwriter_second_review (oversight). Do NOT add committee_exception_for_capacity_overrun unless a capacity overrun actually exists. (Removing two confirmed-correct controls broke the score; each added control is checked individually.)
• escalation_triggers: one per condition in the template's condition_choices, ordered by ascending trigger_id; owner mapping: delinquency/state-benchmark triggers → credit_risk_manager; insurance/lien operational triggers → operations_control_manager; capacity/exception triggers → lending_committee_chair.
• interpretation: capacity_status from the segment's available capacity; external_risk_status from NC-vs-national-and-peers; risk_tolerance mirrors the segment's stated risk_tolerance; committee_message matches the capacity×external-risk combination.

8. Q1 Allocation Package (capacity tasks) — best-effort, structurally fragile

This task type was the hardest to score (feedback returned 0.0 across rounds; treat the rules below as UNVERIFIED hypotheses and double-check field shapes):

• allocation: lending_capacity_q1 (from branch) = gross capacity; gross_approved_amount = sum of approved_amount over approve+conditional apps; committed_capacity_amount = sum of bank_capacity_used (bank-retained, after participations); remaining_capacity = lending_capacity_q1 - committed. priority_ranking = application_ids of approved+conditionally-approved apps, strongest credit first.
• Hard DECLINE triggers (reason codes): recent_bankruptcy (months_ago < 24), low_fico (<580), underwater_collateral (ltv>1.0), documentation_gap (documentation_complete==0), weak_dscr (dscr<1.0, or consumer DTI>0.50).
• Mitigable → conditional_approve with conditions: sector_breach (post-approval sector > limit_pct) → participation_required; high_ltv (ltv>0.85) / weak_dscr (1.0-1.25) → reduced_amount; startup_risk (years_in_business<2) → startup_monitoring (and sba_guaranty_required if SBA guaranty present).
• participation_required appears in BOTH the decision enum and the conditions enum. Prefer using it as a CONDITION on a conditional_approve decision (the decision stays approve/conditional_approve/decline/defer). Bank-retained capacity for a participation loan = retain only up to the sector headroom (limit×total_loans − existing sector exposure), sell the rest.
• For a sector ALREADY over its ceiling, a new approval worsens it → decline (sector_breach) unless the credit is strong enough to mitigate via full participation.
• concentration_flags: one entry per app whose post-approval sector > limit (flag boolean true), sorted by sector then application_id; handling = how the breach is treated (decline / participation_required / conditional_approve).
• decline_reasons: object mapping each DECLINED application_id → its sorted list of reason codes (only declined apps are keys).
• post_approval_concentrations: per sector with approved apps, sorted by sector.
• Conditions list for every app: ["none"] for approve/decline/defer; the specific mitigations for conditional_approve.

9. Decision / Reason / Condition Enums (exact strings)

• Decisions: approve, conditional_approve, decline, defer, participation_required.
• Conditions: participation_required, reduced_amount, board_exception, sba_guaranty_required, startup_monitoring, 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.
• Recommended actions (workout/watch-list): monitor, watchlist, special-assets, workout, partial_chargeoff_review, legal_referral.
• Payment status: Current, 30 Days Past Due, 60 Days Past Due, 90+ Days Past Due, Nonaccrual.
• CDFI risk classes: Prime, Desirable, Satisfactory, Watch, Doubtful, Projected Loss.
• CRE score classes: approve_quality (<=2.0), conditional (<=3.0), weak (>3.0). Lower weighted score is better.

10. CRE Weighted Score (competing-CRE) — UNRESOLVED, treat with care

Policy cre_weighted_score: weights capacity 0.45, capital 0.03, character 0.05, collateral_exposure 0.36, conditions 0.11 (sum 1.0); classes approve_quality<=2.0, conditional<=3.0, weak>3.0. The 5-C sub-score inputs are NOT documented in the policy and this was the single biggest unresolved gap (score stuck at ~0.21 regardless of null treatment). Best-guess mapping to investigate: capacity→DSCR (score 0-6), capital→debt_to_asset, character→fico, collateral_exposure→ltv, conditions→liquidity, then weighted sum rounded to 1dp. NULL treatment (null→0 vs null→worst) did NOT change the score here, so the mapping itself is likely wrong — re-derive carefully (consider character→guarantor_strength, conditions→sector/purpose, or a DSCR score table derived from the risk_rating.dscr_thresholds). Compare the two apps; the lower score is the stronger credit and becomes the selected path.

11. Numeric Conventions & Output Discipline

• Money: 2 decimals (USD). Ratios/percentages: 4 decimals. Basis points: 2 decimals, SIGNED (positive = branch/adverse over benchmark).
• Compute bps from RAW values, round only the final bps (never from a pre-rounded ratio). Compute ratios to 4dp from raw, independently.
• All list orderings are explicit in each template — follow them exactly: ascending loan_id / application_id / sector / trigger_id / action; "descending exposure then ascending loan_id" for workout queues; sort reason-code lists alphabetically.
• Output ONLY the JSON object matching answer_template.json — no narrative text outside the JSON.
• Re-derive every value from live API data; do not hard-code.

12. Common Misjudgments to Avoid

• bps from rounded ratio → recompute from raw variance. (corrected, +score)
• "Sensible" override of the strict re-derivation rule (e.g. keeping current rating for a delinquency-only loan to avoid an upgrade) → WRONG. Apply max-of-available-factors strictly. (overruling dropped the score)
• CDFI null→0 → WRONG. Use worst-tier score for missing factors. (corrected, +score)
• Concentration denominator = total_assets → WRONG. Use total_loans_outstanding.
• Regrade population vs watch-list coverage: regrade population is current_rating >= min (e.g. >=3); watch-list action coverage is final_rating >= 5. Different sets.
• Severe-delinquency override: Nonaccrual/90+ delinquency is a factor equal to its floor (8/7) even when DSCR/LTV are absent — it must be the max.
• Including pass loans in coverage (final 3-4) → exclude; only final>=5.
• Adding template-universe checklist gates not in the segment's minimum_checklist → only use the segment's actual gates.
• Global string-replace when patching a build script → it can hit multiple loops and silently corrupt other sections; patch the specific code block.
• participation_required as a standalone decision vs as a condition → prefer conditional_approve + participation_required condition unless the template clearly wants the decision value.

13. Iterative Verification Discipline

When iterating on a candidate answer, change ONE hypothesis per round where possible so the score delta isolates the fix; keep the best-scoring candidate. When a candidate scores near-zero across several unrelated content changes, suspect a structural/conformance issue (field type, enum value, missing key, wrong list ordering) rather than content, and re-read answer_template.json field-by-field before retrying. Verify each list's required ordering and each field's required type/precision against the template before submitting.