Credit Committee Packets

Credit Committee Packets

You author one JSON object that exactly matches the task's answer_template.json, computed from a live read-only HTTP API. The API is the single source of truth; the policy endpoint holds the authoritative thresholds. Do not invent numbers, do not read local files, and emit only the JSON (no prose) unless the prompt says otherwise.

This skill covers five recurring task types. They share one toolkit (the same API, the same policy, the same risk-rating / stress / scoring math). Read this body, then open the matching section of references/task_playbooks.md for the exact output shape and a worked, validated procedure.

0. Always do these first

1. Read the answer template. It is the contract. Match every required key, every enum value, every ordering rule, and every rounding precision exactly. The template's precision, ordering, and allowed_values annotations are not decoration — graders check them.
2. Identify the target. branch_id for branch tasks, segment_id for credit-union segment tasks. IDs are matched case-insensitively but echo them back uppercased exactly as the standard answers do (e.g. "REDWOOD", "CIVIC_NC_FIRE_EMS").
3. Fetch the policy once and keep it: GET /api/policies. Apply its thresholds verbatim; never hardcode numbers you remember from an example.
4. Identify which of the five task types you are in from the template's top-level keys (see the router below), then follow that playbook.

API quick reference

Base URL: http://127.0.0.1:8003. All JSON; money/ratios are plain numbers.

Key loan fields: loan_id, borrower_name, current_rating, payment_status, outstanding_balance (this is the exposure), dscr, ltv, fico, debt_to_asset, liquidity_months, collateral_value, loan_type, sector, days_past_due. Any of dscr/ltv/fico/etc. may be null — treat null as "factor not available" and skip it, never as zero.

Task-type router (match on the template's top-level keys):

• portfolio_regrade, npa_benchmark, material_downgrades, top_problem_credit → Rating Migration Review
• watch_list_summary, stress_results, workout_queue, severe_bucket_counts → Watch-List Stress Packet
• allocation, decisions, concentration_flags, decline_reasons, post_approval_concentrations → Allocation Package
• applications_compared, recommended_path, stress, concentration, conditions → Competing CRE Decision
• posture, state_metrics, peer_comparison, controls, escalation_triggers, interpretation → Segment Posture Page

Core business rules (shared across tasks)

These are derived from /api/policies and verified against the standard answers. The exact band values live in the policy JSON — read them at runtime; the rules below tell you how to apply them.

Re-deriving a risk rating (risk_rating block)

The final re-derived rating is the worst (numerically highest) rating produced by the available factors — this is the dominant_factor_rule. Compute a candidate rating from each factor that is present, then take the max:

• DSCR → walk dscr_thresholds from best to worst: the first band whose min the DSCR meets gives the rating; if DSCR is below every min (the max_below: 1.0 band), it is the worst rating (7).
• LTV → walk ltv_thresholds: the first band whose max the LTV is <= gives the rating; if LTV exceeds the top max (the min_above: 1.0 band) it is the worst rating (7).
• Delinquency → delinquency_minimums[payment_status] is a floor: Current→null (no floor), 30 Days Past Due→4, 60 Days Past Due→5, 90+ Days Past Due→7, Nonaccrual→8.

final_rating = max(candidate ratings that are not null). If no factor is available at all, keep the loan's existing current_rating. Exposure is always outstanding_balance.

Material downgrade: downgrade_notches = final_rating - current_rating. A downgrade is material when downgrade_notches >= risk_rating.material_downgrade_notches (currently 2). Upgrades and 1-notch moves are not material.

Recommended / watch-list action by governing rating

Map the governing rating (the final re-derived rating in regrade tasks; the current rating in watch-list tasks where the population is defined by current_rating) to an action enum:

• rating <= 5 → monitor (and such loans are excluded from watch-list action coverage entirely)
• rating 6 → watchlist
• rating 7 → special_assets
• rating 8 → partial_chargeoff_review

workout and legal_referral exist in the enum but were not exercised by the train set; only use them if a prompt explicitly directs a workout/legal path. The top/most-severe problem credit is the loan with the worst final rating (tie-break by largest exposure); for a rating-8 Nonaccrual credit the action is partial_chargeoff_review.

Stress formulas (stress block)

• Watch-list +200bp parallel shock: stressed_dscr = dscr / (1 + 0.18) (i.e. dscr / 1.18). shock_label = "+200bp".
• CRE dual stress: stressed_dscr = dscr * 0.85 / (1 + 0.18) (i.e. dscr * 0.85 / 1.18).
• Breach: stressed_dscr < coverage_breach_threshold (1.0). Round stressed/base DSCR to 2 decimals. Only include loans/applications where DSCR is present.

CDFI factor scoring & risk class (cdfi_factor_scores block)

Sum the integer sub-scores for each factor that is present (skip nulls): fico band + ltv band + debt_to_asset band + liquidity_months band. Then map the total to a class using classes: Prime 0–5, Desirable 6–9, Satisfactory 10–13, Watch 14–18, Doubtful ≥19. Projected Loss override: classify as Projected Loss when collateral is underwater (ltv > 1.0) on a defaulted credit (Nonaccrual), or when score ≥19 and ltv > 1.0. The projected_loss boolean in a workout queue is true exactly for the Projected Loss loans.

Capacity, concentration & benchmark math

• Exposure of a loan = outstanding_balance. Sector / CRE existing exposure = sum of the relevant outstanding_balance (or the current_exposure rows from /sector-exposures).
• Branch total loans = metrics.total_loans_outstanding for the review quarter (2025Q1).
• Concentration ratio = exposure ÷ total book, rounded to 4 decimals. Post-approval ratios add the newly approved amount to both numerator (the affected sector / CRE) and the denominator (total book).
• CRE concentration = sum of loan_type == "CRE" outstanding balances ÷ total_loans_outstanding.
• Variance between any branch ratio and a benchmark ratio: variance_ratio = branch_ratio - benchmark_ratio (4 decimals), variance_bps = variance_ratio * 10000 (2 decimals). Same pattern for a post-approval ratio vs a policy limit (policy_variance_bps).
• NPA / noncurrent loans = payment_status in ("90+ Days Past Due", "Nonaccrual").

FDIC / NCUA benchmark metric selection

• FDIC noncurrent NPA work uses total_loans_noncurrent_pct. FDIC CRE delinquency work uses total_real_estate_30_89_pct (compared against the branch delinquency_30_plus_pct). The template's benchmark_metric / fdic_benchmark_metric enum tells you which one is expected — pick that one.
• NCUA state metrics are copied verbatim as integers from the matching state_code row. Compare the target state against the US row (nc_vs_us) and against the median of the segment's peer_states (nc_vs_peer_median); direction is higher/lower/equal of the target value vs the comparison value (note: for delinquency, "higher" means worse).

Output conventions (apply to every task)

• Rounding: currency/exposure → 2 decimals; ratios/concentration/variance_ratio → 4 decimals; bps → 2 decimals; DSCR (base & stressed) → 2 decimals; CDFI weighted score → 1 decimal. Round only the final emitted value; carry full precision through intermediate steps.
• Ordering: obey each list's stated ordering exactly. Common ones: ascending loan_id, ascending by final_rating, ascending by application_id, reason codes / condition lists ascending alphabetically, sectors ascending, workout_queue is descending exposure then ascending loan_id, severe_bucket_counts ascending current_rating then payment_status.
• Enums: every action / decision / reason-code / class / direction value must come from the template's allowed set. Never emit a free-text value where an enum is required.
• Numbers, not strings: money and ratios are JSON numbers. Echo IDs as uppercase strings.
• Whole-list selection: when a task says "rated 3 or worse" or "rating 6 or worse", fetch with min_current_rating=N; "noncurrent" means the two noncurrent payment statuses; population filters are by current_rating unless the task explicitly says to filter on the re-derived rating.

Procedure

1. Read the template; identify task type via the router.
2. GET /api/policies; fetch the branch/segment, metrics (2025Q1), loans/applications, sector-exposures, and the FDIC/NCUA benchmark the template names — only what that task needs.
3. Open the matching section of references/task_playbooks.md and follow the worked procedure. It lists the exact fields, the field-by-field computation, and the validated formulas for that task.
4. Compute with full precision; round and order at the end.
5. Emit only the JSON object. Re-check it against the template's required keys, enums, ordering, and precision before finishing.

Common misjudgments to avoid

• Null factors are not zero. A null DSCR/LTV/FICO is skipped, not scored 0. A loan with no rateable factor keeps its existing current_rating.
• Don't round mid-computation. Round once, at emit time, or downstream ratios/variances drift.
• Population vs governing rating. "Adverse" / "rated N or worse" selects by current_rating; the action and migration use the rating appropriate to the task (final for regrade, current for watch-list). Don't mix them.
• Exposure = outstanding_balance, not collateral_value or requested_amount (for booked loans).
• Post-approval denominators grow. Adding an approval increases the total book in the denominator too; don't divide by the pre-approval total.
• Pick the benchmark metric the template names, not the first FDIC field you see. NPA uses the noncurrent metric; CRE delinquency uses the real-estate 30–89 metric.
• Watch-list coverage excludes rating ≤5. Only 6/7/8 credits get a watch action.
• Echo the review_date / as-of date and branch/segment id exactly as given (uppercased id).
• Emit JSON only. No markdown fences, no commentary, when the prompt asks for the JSON answer.