Credit Committee Packets

Credit Committee Packets

You produce a single JSON object that conforms to the task's answer_template.json. The underlying data lives behind a read-only HTTP API; the policy endpoint is the single source of truth for every threshold and formula. Never hardcode thresholds from memory or from a worked example — re-read them from the API each run, because branch limits and policy bands vary per task.

Golden rules (apply to every task)

1. Output is JSON only. No prose, no markdown fences, nothing outside the object. Emit exactly the keys the template lists, in the template's structure. Extra keys are risk; missing required keys fail.
2. The API is the only data source. Use curl/urllib against http://127.0.0.1:8003. Ignore any mention of env/setup.sh, local DBs, or files — the service is already up. Treat every number as authoritative; do not re-estimate it.
3. Re-read /api/policies every run and drive all bands/weights/formulas from it. Branch objects carry their own cre_policy_limit_pct, sector_ceiling_pct, and lending_capacity_q1; sector rows carry their own limit_pct. Use the per-entity value, not a global default.
4. Use only enum values the template allows. Reason codes, decisions, actions, postures, risk classes, conditions, owners — all are closed vocabularies. Never invent a value.
5. Respect rounding and ordering exactly (see "Precision & ordering" below). The grader compares numbers and list order literally.
6. Compute the helper script once. scripts/credit_office.py wraps the API and the policy math (risk re-derivation, CDFI scoring, stress, ratios). Read it and reuse it instead of re-deriving formulas by hand.

How to approach any task in this family

1. Read the prompt and the staged input/payloads/answer_template.json. The template's required_top_level_keys, item_required_keys, allowed_values/choices, ordering, and precision/numeric_precision notes ARE the spec. Build your output shape from it.
2. Identify the variant from the requested keys and verbs (see references/playbooks.md):
   • regrade / rating migration → portfolio_regrade + npa_benchmark + material_downgrades + top_problem_credit
   • watch-list stress / workout → watch_list_summary + stress_results + workout_queue + severe_bucket_counts
   • allocation / pending applications → allocation + decisions + concentration_flags + decline_reasons + post_approval_concentrations
   • competing CRE decision → applications_compared + recommended_path + stress + concentration + conditions
   • credit-union segment posture → posture + state_metrics + peer_comparison + controls + escalation_triggers + interpretation
3. Pull /api/policies and the entity data the prompt names (branch_id or segment_id). Fetch exactly the surfaces the prompt lists; extra fetches are fine but do not invent data.
4. Apply the playbook for that variant (references/playbooks.md) using policy thresholds.
5. Assemble JSON, apply precision/ordering, validate enums, print only the JSON.

Core business rules (driven by /api/policies, schema in references/policy.md)

These are stable across tasks; the exact numbers come from the live policy each run.

Risk-rating re-derivation (dominant-factor rule). For a loan, compute a candidate rating from each available objective factor and take the worst (highest) numeric rating:

• DSCR → risk_rating.dscr_thresholds (first band whose min is met, descending; if DSCR < the lowest min, use the max_below band, i.e. rating 7).
• LTV → risk_rating.ltv_thresholds (first band whose max is satisfied, ascending; if LTV exceeds the top max, use the min_above band, rating 7).
• Delinquency → risk_rating.delinquency_minimums[payment_status] is a floor (e.g. Nonaccrual ≥ 8, 90+ ≥ 7, 60 ≥ 5, 30 ≥ 4; "Current" is null = no floor).
• Final rating = max of the available candidates. If no factor is available (no DSCR, no LTV, payment_status Current → null floor), keep the loan's existing current_rating.
• A downgrade is material when final_rating − current_rating ≥ risk_rating.material_downgrade_notches.

CDFI factor scoring (additive, lower is better). Sum the four factor scores from cdfi_factor_scores for FICO, LTV, debt_to_asset, liquidity_months. A factor that is missing contributes 0 (it cannot be scored). Map the total to a class via the score bands (Prime 0-5, Desirable 6-9, Satisfactory 10-13, Watch 14-18, Doubtful ≥19). Projected Loss override: a credit already in the Watch band or worse whose LTV > 1.0 (underwater collateral) is reported as Projected Loss rather than its score-band class.

CRE weighted score (lower is better). cre_weighted_score.weights (capacity 0.45, collateral_exposure 0.36, conditions 0.11, character 0.05, capital 0.03) weight CDFI-style factor scores. Class from cre_weighted_score.classes: approve_quality ≤ 2.0, conditional ≤ 3.0, weak > 3.0. The class and the resulting decision are the gradeable outputs; weight the per-C factor scores as described in references/playbooks.md.

Stress formulas (from stress):

• Watch-list +200bp parallel shock: stressed_dscr = dscr / (1 + 0.18).
• CRE dual stress: stressed_dscr = dscr * 0.85 / (1 + 0.18).
• Breach when stressed_dscr < coverage_breach_threshold (1.0). Only loans/apps with a DSCR appear in stress results; skip rows where DSCR is null.

Recommended-action / workout mapping keyed off the effective severity rating (re-derived final rating, or current_rating where the variant specifies): rating 8 / Nonaccrual / Projected Loss → partial_chargeoff_review; rating 7 → special_assets; rating 6 → watchlist. Lower/healthier → monitor. Reserve legal_referral/workout for cases the prompt explicitly calls out.

Capacity & concentration.

• lending_capacity_q1 is the branch quarterly capacity. bank_capacity_used for an approved app = approved amount, reduced by participation (bank retains a portion) and by SBA guaranty (approved × (1 − sba_guaranty_pct)).
• committed_capacity_amount = Σ bank_capacity_used; remaining_capacity = lending_capacity_q1 − committed_capacity_amount; gross_approved_amount = Σ approved amounts (full, before participation/SBA reduction) for approved + conditional apps.
• Concentration % = sector exposure ÷ post-approval total portfolio (Σ all sector current_exposures + all approved additions), rounded to 4 decimals. Compare to the sector's limit_pct (or branch sector_ceiling_pct). Existing over-ceiling exposure may be grandfathered, but a new approval may not worsen a breached sector without a mitigation (participation_required, reduced_amount, board_exception).

Benchmark variance. Branch ratio − benchmark ratio. variance_bps is computed from the UNROUNDED ratios (branch_ratio − benchmark_ratio) × 10000 then rounded to 2 dp — do NOT multiply the already-rounded 4-dp variance_ratio, which loses precision. Pick the benchmark metric the template's enum specifies (e.g. total_loans_noncurrent_pct for NPA, total_real_estate_30_89_pct for CRE real-estate delinquency).

Precision & ordering (grader compares literally)

• Currency (USD) fields → 2 decimals. Ratios/percentages-as-ratios → 4 decimals. bps → 2 decimals (from unrounded inputs, per above). CRE weighted score → 1 decimal. DSCR base/ stressed → 2 decimals. Counts/ratings → integers.
• Emit ratios as decimals (0.1897), never as "18.97%".
• Apply each list's stated ordering exactly. Common ones: ascending loan_id / application_id (string sort), ascending final_rating, reason codes ascending alphabetically, sectors ascending, peer states ascending, workout queue descending exposure then ascending loan_id, severe buckets ascending current_rating then ascending payment_status (plain string sort, so "90+ Days Past Due" sorts before "Current").
• NCUA state-benchmark integers (delinquency_bps, loan_to_share_pct, roaa_bps, positive_net_income_pct) are reported exactly as the table gives them — no recomputation.

Common misjudgments to avoid

• Don't drop the no-factor loans. A loan with no DSCR/LTV and a Current status stays at its current_rating and still counts in totals — it does not vanish or become rating 0.
• Population filters are inclusive and literal. "rated 3 or worse" = current_rating ≥ 3; "current_rating 6 or worse" = ≥ 6; "adversely rated" uses the rating the prompt names.
• Stress only covers DSCR-available rows. Missing-DSCR loans are excluded from stress_results/results but may still appear in other sections.
• bps from unrounded ratios (see above) — the single most common numeric error.
• Concentration denominator is the portfolio post-approval total, not metrics total_loans_outstanding and not the lending capacity.
• SBA guaranty and participation reduce capacity used, not the approved/gross amount.
• Enums only. A plausible-sounding label that isn't in the template's list is wrong.
• score_class/decision, not the raw weighted score, are usually graded — when a weighted decomposition is ambiguous, get the class band and the decision right and keep the score consistent with the band.
• NPA/CRE benchmark metric must match the template enum, not whatever metric seems closest.

References

• references/policy.md — annotated structure of /api/policies and every band/weight, plus the exact endpoint list and the fields that matter on each surface.
• references/playbooks.md — step-by-step SOP for each of the five known variants, including field-by-field derivation and the data surfaces each variant needs.
• scripts/credit_office.py — runnable helper: API client + policy-driven functions for risk re-derivation, CDFI scoring, stress, ratios, and rounding/ordering. Import or copy it.