Self Attempt 03

Private Wealth Advisory Structured Output Skill

SOP for producing the JSON planning object for an unseen advisory client. Base API URL is in the staged environment_access.md (hereafter API_BASE). All amounts are USD rounded to cents (2 decimals). All dates are ISO YYYY-MM-DD. Return ONLY the JSON object; no prose.

0. Determine analysis type from the engagement text

Map the memo "Engagement:" line (or prompt title) to analysis_type:

• contains "Roth conversion" / "RMD" / "near-RMD Roth" -> roth_conversion_rmd
• contains "ILIT" / "Crummey" -> ilit_crummey_implementation
• contains "GRAT" and "CRAT" / "GRAT versus CRAT" -> trust_comparison
• contains "estate liquidity action plan" -> estate_liquidity_action_plan Set task_id = "test_NNN" where NNN is the zero-padded task folder number. Set client_id = the CLT-XXXX from the memo.

1. Remote API workflow (endpoint order)

1. GET {API_BASE}/api/health — confirm {"ok": true}.
2. GET {API_BASE}/api/policies/tax — cache the planning constants (gift exclusion by year, estate exemption by year, estate_tax_rate, conversion_bracket_targets by filing status, max_crat_term_years, charitable_deduction_rate).
3. GET {API_BASE}/api/rmd-factors — cache the age->divisor map (keys 73..99).
4. GET {API_BASE}/api/clients/{client_id} — base record (age, filing_status, planning_year, estate_value, liquid_assets).
5. GET {API_BASE}/api/source-documents?client_id={client_id} — the CONFLICTING sources (CRM_NOTE, ATTORNEY_MEMO, SIGNED_PROFILE, possibly CUSTODIAN_EXPORT / STALE_MARKETING_INTAKE). Each carries source_type, effective_date, facts{}.
6. Optionally fetch the data endpoints the analysis needs:
   • roth: GET /api/retirement-accounts?client_id=
   • ilit / estate-liquidity: GET /api/life-insurance?client_id=
   • trust-comparison / estate-liquidity: GET /api/trust-candidates?client_id=
7. Read the local request_memo.md for the planning horizon year and engagement framing.
8. Read answer_template.json for the required keys / enums for THIS task.

2. Source-resolution precedence (infer the controlling source)

Each source has an effective_date and source_type. Resolve every fact from the conflicting sources using this precedence (highest first):

1. SIGNED_PROFILE — most recent, client-signed. Controls PROFILE facts: annual_non_ira_income, marginal_tax_rate, beneficiary_count, age, filing_status, marital_status, planning_year, liquid_assets, estate_value, and GOAL facts (philanthropic_intent, family_transfer_priority).
2. ATTORNEY_MEMO — controls ASSET identification for trust transfers (which appreciating asset funds the GRAT/CRAT). Also confirms estate_value.
3. CUSTODIAN_EXPORT — controls ACCOUNT balances (traditional_balance, roth_balance, expected_return on retirement accounts). Always wins for account balances even if a profile states a different number.
4. CRM_NOTE — older import; superseded by any of the above.
5. STALE_MARKETING_INTAKE — lowest; never controls any scored fact.

Resulting source_resolution output fields by analysis type:

• roth_conversion_rmd: controlling_profile_source = SIGNED_PROFILE; controlling_account_source = CUSTODIAN_EXPORT.
• ilit_crummey_implementation: controlling_beneficiary_source = SIGNED_PROFILE; controlling_policy_source = SIGNED_PROFILE (the signed plan documents the policy).
• trust_comparison: controlling_goal_source = SIGNED_PROFILE; controlling_asset_source = ATTORNEY_MEMO.
• estate_liquidity_action_plan: controlling_goal_source = SIGNED_PROFILE; controlling_policy_source = SIGNED_PROFILE.

When SIGNED_PROFILE and ATTORNEY_MEMO agree on a goal fact, emit SIGNED_PROFILE. Only emit a different enum when the signed profile is genuinely absent for that fact category.

3. Reading the planning horizon from the memo

• For roth_conversion_rmd: the memo states Planning horizon year: YYYY. Use that year for rmd_projection.horizon_year and project balances to that year.
• For the other three analysis types there is NO horizon field; the trust term is read from trust-candidates (grat_term_years / crat_term_years) and the ILIT dates from life-insurance.planned_contribution_date.

4. Type A — roth_conversion_rmd (tasks 1, 5 template)

Fetch retirement-accounts; you get traditional_balance, roth_balance, expected_return, rmd_start_age (=73), recommended_conversion_years. From SIGNED_PROFILE read annual_non_ira_income, marginal_tax_rate, filing_status, age, planning_year. From tax policy read conversion_bracket_targets[filing_status].

4a. conversion_plan

• first_conversion_year = planning_year.
• conversion_years = recommended_conversion_years.
• first_rmd_year = planning_year + (rmd_start_age - age) (year owner turns 73).
• annual_conversion_amount = conversion_bracket_targets[filing_status] MINUS annual_non_ira_income. (Fill the bracket: non-IRA income fills first, conversion fills the rest to the bracket ceiling.) Round to cents.
• total_converted = annual_conversion_amount * conversion_years (cents).
• total_conversion_tax = total_converted * marginal_tax_rate (cents).
• conversion_years_positive = number of conversion years that fall STRICTLY BEFORE first_rmd_year = max(0, min(conversion_years, first_rmd_year - first_conversion_year)).

4b. rmd_projection

• horizon_year = memo planning-horizon year.
• first_rmd_year = as above.

Baseline model (no conversion):

B = traditional_balance
for Y in [planning_year .. horizon_year]:
    ageY = age + (Y - planning_year)
    rmd = B / rmd_factors[ageY]          if Y >= first_rmd_year else 0
    tax = rmd * marginal_tax_rate         if Y >= first_rmd_year else 0
    B = (B - rmd) * (1 + expected_return)   # withdraw RMD, then grow remainder
baseline_rmd_tax_through_horizon = sum(tax over Y)

Conversion model (with conversion):

Bc = traditional_balance
Rc = roth_balance
for Y in [planning_year .. horizon_year]:
    ageY = age + (Y - planning_year)
    rmd = Bc / rmd_factors[ageY]          if Y >= first_rmd_year else 0
    conv = annual_conversion_amount       if Y is a conversion year else 0
    tax = rmd * marginal_tax_rate         if Y >= first_rmd_year else 0
    Bc = (Bc - rmd - conv) * (1 + expected_return)
    Rc = (Rc + conv) * (1 + expected_return)
conversion_rmd_tax_through_horizon = sum(tax over Y)

• rmd_tax_savings_through_horizon = baseline_rmd_tax_through_horizon MINUS conversion_rmd_tax_through_horizon (positive = savings; cents).

4c. legacy_projection

• projected_traditional_balance_horizon = Bc at horizon (conversion model).
• projected_roth_balance_horizon = Rc at horizon (conversion model).
• heir_tax_profile: ratio = Rc / (Rc + Bc).
  • ratio >= 0.60 -> MOSTLY_TAX_FREE
  • ratio <= 0.30 -> MOSTLY_TAXABLE
  • else -> MIXED_TAXABLE_AND_TAX_FREE

4d. recommendation

• If conversion_years_positive <= 1 (RMD imminent): primary_action=DEFER, suitability=DEFER, risk_flag=RMD_NEAR_TERM.
• Else if liquid_assets < total_conversion_tax: risk_flag=LIQUIDITY_CONSTRAINT, suitability=BORDERLINE, primary_action=DEFER.
• Else: primary_action=STAGED_ROTH_CONVERSION, suitability=SUITABLE, risk_flag=TAX_BRACKET_MANAGEMENT.

5. Type B — ilit_crummey_implementation (task 2 template)

Fetch life-insurance; you get death_benefit, annual_premium, planned_contribution_date, is_existing_policy_transfer. From SIGNED_PROFILE read beneficiary_count; from the client record read planning_year.

5a. gift_plan

• planning_year = client planning_year (2026 for train data).
• annual_exclusion_per_beneficiary = annual_gift_exclusion[planning_year].
• beneficiary_count = from SIGNED_PROFILE (resolves CRM conflict).
• annual_exclusion_capacity = beneficiary_count * annual_exclusion_per_beneficiary.
• annual_premium = from life-insurance record.
• premium_gap = max(0, annual_premium - annual_exclusion_capacity) (cents).

5b. administration

• notices_required = beneficiary_count.
• contribution_date = planned_contribution_date (ISO).
• notice_due_date = contribution_date (Crummey notice served on contribution).
• withdrawal_window_end = contribution_date + 30 days (ISO).
• earliest_premium_payment_date = withdrawal_window_end + 1 day (trustee pays premium only after the 30-day withdrawal right closes).
• dedicated_bank_account_required = true.

Date arithmetic: add 30 calendar days to the ISO date; + 1 day after. E.g. 2026-03-10 -> window end 2026-04-09 -> earliest premium 2026-04-10.

5c. estate_result

• death_benefit = from life-insurance record.
• estate_inclusion_risk = same value as recommendation.risk_flag (see 5d).
• projected_outside_estate_if_implemented = death_benefit (ILIT-owned policy passes outside the taxable estate when formalities are met).
• tax_liquidity_support = death_benefit (liquidity the ILIT supplies).

5d. recommendation (matrix on two axes)

Axes: is_existing_policy_transfer (T/F) and premium_gap > 0 (T/F).

• F / F: FUND_WITH_CRUMMEY_NOTICES, SUITABLE_WITH_ADMINISTRATION, LOW_IF_FORMALITIES_MET.
• F / T: USE_LIFETIME_EXEMPTION_FOR_SHORTFALL, BORDERLINE, EXCLUSION_SHORTFALL.
• T / F: USE_NEW_POLICY_OR_ACCEPT_LOOKBACK, BORDERLINE, THREE_YEAR_LOOKBACK.
• T / T: DISCLOSE_LOOKBACK_AND_USE_EXEMPTION, NOT_SUITABLE, THREE_YEAR_LOOKBACK_AND_EXCLUSION_SHORTFALL. estate_result.estate_inclusion_risk takes the same risk_flag value.

6. Type C — trust_comparison (task 3 template)

Fetch trust-candidates; you get asset_value, expected_growth_rate, grat_term_years, grat_annuity_rate, crat_term_years, crat_payout_rate. From SIGNED_PROFILE read estate_value, liquid_assets, filing_status, planning_year, and goals (family_transfer_priority, philanthropic_intent).

6a. estate_context (shared formula for types C and D)

• taxable_estate = estate_value from SIGNED_PROFILE.
• estate_tax_exposure = max(0, taxable_estate - estate_tax_exemption[planning_year]) * estate_tax_rate.
• liquidity_gap_before_planning = max(0, estate_tax_exposure - liquid_assets).

6b. grat

• term_years = grat_term_years.
• annuity_payment = asset_value * grat_annuity_rate (fixed each year).
• projected_remainder_to_heirs = asset_value * (1+growth)^term - annuity_payment * (((1+growth)^term - 1) / growth) (cents). (Grow the asset each year, pay the fixed annuity at year-end; remainder passes to heirs.)
• estimated_estate_tax_reduction = projected_remainder_to_heirs * estate_tax_rate (cents).
• mortality_inclusion_risk = TERM_SURVIVAL_REQUIRED (always; grantor must survive the term).

6c. crat

• term_years = min(crat_term_years, max_crat_term_years).
• annuity_payment = asset_value * crat_payout_rate (fixed each year).
• projected_charitable_remainder = asset_value * (1+growth)^term - annuity_payment * (((1+growth)^term - 1) / growth) (cents).
• estimated_income_tax_deduction = asset_value * charitable_deduction_rate (cents).
• family_transfer_fit based on family_transfer_priority: high -> LOW; moderate -> MODERATE; low -> HIGH.

6d. recommendation

• preferred_strategy: GRAT if family_transfer_priority == "high" and philanthropic_intent != "high"; else CRAT if philanthropic_intent == "high"; else default GRAT.
• rationale_code: GRAT -> CHILDREN_TRANSFER_PRIORITY; CRAT -> PHILANTHROPIC_PRIORITY.
• alternate_role: GRAT preferred -> SECONDARY_CHARITABLE_TOOL; CRAT preferred -> SECONDARY_FAMILY_TRANSFER_TOOL.

7. Type D — estate_liquidity_action_plan (task 4 template)

Combines estate_context + ILIT + trust transfer. Fetch life-insurance AND trust-candidates. Use SIGNED_PROFILE for estate/liquid goals, beneficiary_count.

7a. estate_context — same formula as 6a.

7b. ilit

• annual_exclusion_capacity = beneficiary_count * annual_gift_exclusion[planning_year].
• premium_gap = max(0, annual_premium - annual_exclusion_capacity).
• estate_inclusion_risk = risk_flag from the 5d matrix.
• projected_outside_estate_if_implemented = death_benefit.

7c. trust_transfer

• preferred_strategy: GRAT/CRAT using the same goal rule as 6d.
• projected_remainder_to_heirs = GRAT remainder (formula 6b).
• estimated_estate_tax_reduction = projected_remainder_to_heirs * estate_tax_rate.
• projected_charitable_remainder = CRAT remainder (formula 6c, computed regardless of which strategy is preferred).

7d. recommendation

• primary_action:
  • GRAT preferred + ILIT funding (no shortfall) -> COMBINE_ILIT_AND_GRAT.
  • CRAT preferred (philanthropic_intent high) -> CRAT_WITH_LIQUIDITY_REVIEW.
  • ILIT has premium_gap > 0 -> ILIT_WITH_EXEMPTION_REVIEW.
• sequencing:
  • COMBINE_ILIT_AND_GRAT -> ILIT_FIRST_THEN_GRAT.
  • CRAT_WITH_LIQUIDITY_REVIEW -> TRUST_DECISION_FIRST.
  • ILIT_WITH_EXEMPTION_REVIEW -> ILIT_FIRST_THEN_ATTORNEY_REVIEW.
• risk_flag = the ILIT risk from the 5d matrix.

7e. action_set (sorted alphabetically, subset of the 5 enum values)

Always include: ATTORNEY_DRAFT_REVIEW. Include GRAT_FOR_APPRECIATING_SHARES when GRAT is preferred. Include CRAT_FOR_CHARITABLE_REMAINDER when CRAT is preferred. Include ILIT_CRUMMEY_NOTICE_CYCLE when the ILIT is being funded. Include LIFETIME_EXEMPTION_ALLOCATION when taxable_estate > estate_tax_exemption (always true for these high-net-worth clients) OR when premium_gap > 0. Sort the resulting list alphabetically and output as a JSON array of strings.

8. Pitfalls

• Do NOT use CRM_NOTE values for income, beneficiary_count, or goals when a SIGNED_PROFILE exists — the CRM is the stale import that disagrees.
• Do NOT use the signed-profile stated balance for retirement accounts; use the CUSTODIAN_EXPORT balance from /api/retirement-accounts.
• conversion_years is the RECOMMENDED count from the account record, NOT a value you derive.
• annual_conversion_amount is the bracket fill (target MINUS non-IRA income), not the balance divided by years.
• first_rmd_year = year the owner turns 73 = planning_year + (73 - age). Patel (age 72) starts RMD the very next year.
• RMD factor for year Y uses the age IN that year (age + (Y - planning_year)), not the current age. Guard against age > 99 (no factor) by clamping.
• premium_gap is max(0, ...), never negative.
• GRAT/CRAT annuity payment = asset_value * rate EACH YEAR (fixed), and the remainder formula grows the starting asset and subtracts the FV of the annuity stream at expected_growth_rate.
• estimated_income_tax_deduction for CRAT = asset_value * charitable_deduction_rate (a flat fraction of the contribution), NOT the future charitable remainder.
• All numeric outputs must be JSON numbers (not strings) rounded to 2 decimals.
• All date outputs must be ISO YYYY-MM-DD strings.
• action_set MUST be sorted alphabetically.
• Use the estate_tax_exemption and annual_gift_exclusion for the client's planning_year (2026 -> exemption 13,610,000; gift exclusion 20,000), not 2025.
• task_id must be the literal test_NNN for the current test task, not a train id.