Self Attempt 02

Private Wealth Advisory — Structured Planning Output SOP

Use this for personal-financial-advisory engagement tasks that ask for a single JSON object conforming to a per-task answer_template.json. Domain covers Roth conversion + RMD projection, ILIT Crummey implementation, GRAT-vs-CRAT trust comparison, and integrated estate-liquidity action planning. The advisory API serves client records plus CONFLICTING source documents (CRM note, attorney memo, custodian export, signed profile, stale marketing intake). You must resolve which source controls each fact before computing.

0. Output contract (all task types)

• Return ONLY a JSON object matching the task's answer_template.json required top-level keys. No prose, no markdown fences.
• All USD amounts are JSON numbers rounded to cents (2 decimals), e.g. 469504.00.
• Dates are ISO YYYY-MM-DD.
• task_id = the task identifier supplied by the harness, formatted train_NNN / test_NNN (derive from the task folder name if not explicit, e.g. folder 005 + harness phase train -> train_005).
• client_id = the stable client id from the request memo (see section 2).
• analysis_type = the enum string from the answer template's analysis_type field.

1. Read engagement inputs (memo vs prompt)

1. Read input/payloads/request_memo.md FIRST. It is the authoritative engagement record.
2. Extract from the memo:
   • Client ID: line -> client_id (USE THIS, not the client name in prompt.txt).
   • Engagement: line -> confirms which analysis type.
   • Planning horizon year: NNNN line -> the RMD projection horizon (Roth tasks only). If absent, the task is point-in-time / term-based (ILIT, GRAT/CRAT, estate-liquidity) and needs no multi-year horizon.
3. input/prompt.txt is generic boilerplate; its client name line can be STALE and may disagree with the memo. Always trust the memo's Client ID: over the prompt prose.
4. input/payloads/answer_template.json defines the exact required keys, enums, and any ordering constraint (e.g. action_set sorted alphabetically). Build the JSON to match.

2. Remote API workflow + endpoint order

Base URL is in environment_access.md (e.g. http://HOST:PORT). GET only.

Fetch in this order per engagement:

1. GET /api/health -> confirm {"ok": true}.
2. GET /api/policies/tax -> planning constants (cache once, reuse for every client):
   • annual_gift_exclusion (by year, e.g. {"2025":19000,"2026":20000})
   • estate_tax_exemption (by year, e.g. {"2025":13990000,"2026":13610000})
   • estate_tax_rate (e.g. 0.4)
   • conversion_bracket_targets by filing status (MFJ/SINGLE/HOH)
   • max_crat_term_years (e.g. 20)
   • charitable_deduction_rate (e.g. 0.35)
3. GET /api/rmd-factors -> map of age (string) -> IRS Uniform Lifetime Table divisor (73->26.5, 74->25.5, ...). Used only for Roth RMD projection.
4. GET /api/clients/<client_id> -> base client record (age, filing_status, marital_status, planning_year, estate_value, liquid_assets).
5. GET /api/source-documents?client_id=<id> -> array of conflicting source docs. Each has source_type, effective_date, and a facts object. THIS IS THE CONFLICT SURFACE.
6. Analysis-type-specific records:
   • Roth: GET /api/retirement-accounts?client_id=<id>
   • ILIT: GET /api/life-insurance?client_id=<id>
   • GRAT/CRAT or estate-liquidity: GET /api/trust-candidates?client_id=<id> AND GET /api/life-insurance?client_id=<id> (estate-liquidity needs both).

3. Source-resolution precedence (inferred rule)

For any fact, resolve conflicts using this authority ladder, HIGHEST to LOWEST:

1. SIGNED_PROFILE — client-signed, newest effective_date. Controls personal/demographic profile AND stated client goals: age, annual_non_ira_income, marginal_tax_rate, filing_status, marital_status, liquid_assets, estate_value, beneficiary_count, planning_year, philanthropic_intent, family_transfer_priority.
2. ATTORNEY_MEMO — attorney planning-call notes. Controls LEGAL trust/estate STRUCTURE characterization: ILIT trust design, GRAT/CRAT vehicle choice, and which asset is placed in trust. Wins for controlling_asset_source (trust assets are legal structures, not custodian holdings).
3. CUSTODIAN_EXPORT — account/policy holdings of record. Controls money-on-hand facts: traditional_balance, roth_balance, expected_return, rmd_start_age, recommended_conversion_years, and policy death_benefit/annual_premium (the retirement-accounts and life-insurance records are custodian-type data).
4. CRM_NOTE — older CRM import. Fallback only; never controls when a higher source states the same fact.
5. STALE_MARKETING_INTAKE — lowest; marketing-only. Never controls a material fact.

Per-task source_resolution output mapping

Note the enum composition is itself a hint: controlling_asset_source never includes CUSTODIAN_EXPORT (a trust asset is not a brokerage holding), while controlling_policy_source does include it (an insurance policy is a financial product).

4. Shared formulas

• first_rmd_year = planning_year + max(0, rmd_start_age - age) (RMD begins at age 73).
• taxable_estate = max(0, estate_value - estate_tax_exemption[planning_year]) (estate_value from controlling source, typically signed profile).
• estate_tax_exposure = taxable_estate * estate_tax_rate.
• liquidity_gap_before_planning = max(0, estate_tax_exposure - liquid_assets).
• annual_gift_exclusion for the plan = annual_gift_exclusion[planning_year].

5. Analysis type A — roth_conversion_rmd

Resolved inputs: age, annual_non_ira_income, marginal_tax_rate, filing_status (from SIGNED_PROFILE); traditional_balance (T0), roth_balance (R0), expected_return (g), rmd_start_age, recommended_conversion_years (from CUSTODIAN_EXPORT); conversion_bracket_targets[filing_status] (bt); RMD factor table; horizon_year (memo).

conversion_plan (BRACKET-FILL method — conversion_bracket_targets is load-bearing)

• first_conversion_year = planning_year.
• conversion_years = recommended_conversion_years.
• annual_conversion_amount = max(0, bt - annual_non_ira_income) (fill remaining bracket room each year), capped in the final year at the remaining traditional balance.
• total_converted = sum of per-year conversions across the window (<= T0). Round cents.
• total_conversion_tax = total_converted * marginal_tax_rate. Round cents.
• conversion_years_positive = count of conversion years falling BEFORE first_rmd_year (the pre-RMD window), capped at conversion_years: max(0, min(conversion_years, first_rmd_year - first_conversion_year)). These are the years conversion is unambiguously beneficial.

rmd_projection (year-by-year simulation, planning_year..horizon_year)

Define factor(age_in_year) from the RMD table (age attained that year). BASELINE (no conversion): For each year Y, beginning balance B_Y:

• if Y >= first_rmd_year: rmd_Y = B_Y / factor(age_Y); tax_Y = rmd_Y * mtr; B_{Y+1} = (B_Y - rmd_Y) * (1+g).
• else: rmd_Y = 0; B_{Y+1} = B_Y * (1+g). baseline_rmd_tax_through_horizon = sum of tax_Y for Y in [first_rmd_year, horizon]. CONVERSION scenario: same loop, but in each conversion year reduce the traditional balance by that year's conversion amount (after taking any RMD first in overlap years), and credit the conversion to the Roth balance which then grows at g. Take RMDs on the reduced balance. conversion_rmd_tax_through_horizon = sum of tax_Y under conversion.
• horizon_year = memo horizon.
• first_rmd_year = as above.
• rmd_tax_savings_through_horizon = baseline_rmd_tax_through_horizon - conversion_rmd_tax_through_horizon. Round cents.

legacy_projection (end of horizon_year, conversion scenario)

• projected_roth_balance_horizon = R0 grown to horizon + each conversion grown the remaining years at g. Round cents.
• projected_traditional_balance_horizon = remaining traditional balance at end of horizon (after conversions and RMDs). Round cents.
• heir_tax_profile: let r = projected_roth_balance_horizon / (roth + trad) at horizon. r > 0.6 -> MOSTLY_TAX_FREE; r < 0.4 -> MOSTLY_TAXABLE; else MIXED_TAXABLE_AND_TAX_FREE.

recommendation

• primary_action: STAGED_ROTH_CONVERSION if T0 > 0 and pre-RMD window >= 1; DEFER if RMD imminent and bracket/liquidity poor; NO_CONVERSION if T0 == 0.
• suitability: SUITABLE if window >= ~3 yrs and liquid_assets cover conversion tax; BORDERLINE if window 1-2 yrs or tax exceeds liquid_assets; DEFER if no viable window.
• risk_flag (most salient wins, priority order): RMD_NEAR_TERM if first_rmd_year - planning_year <= 2; else LIQUIDITY_CONSTRAINT if total_conversion_tax > liquid_assets; else TAX_BRACKET_MANAGEMENT (bracket-fill inherently involves managing the bracket).

source_resolution

controlling_profile_source = SIGNED_PROFILE; controlling_account_source = CUSTODIAN_EXPORT.

6. Analysis type B — ilit_crummey_implementation

Resolved inputs: beneficiary_count (SIGNED_PROFILE); planning_year (client record); policy death_benefit, annual_premium, planned_contribution_date, is_existing_policy_transfer (life-insurance record); annual_gift_exclusion[planning_year].

gift_plan

• planning_year = client.planning_year.
• annual_exclusion_per_beneficiary = annual_gift_exclusion[planning_year] (e.g. 20000 for 2026).
• beneficiary_count = from SIGNED_PROFILE.
• annual_exclusion_capacity = annual_exclusion_per_beneficiary * beneficiary_count.
• annual_premium = policy.annual_premium.
• premium_gap = annual_premium - annual_exclusion_capacity (raw; positive => shortfall).

administration (Crummey notice formalities)

• notices_required = beneficiary_count (one notice per beneficiary).
• contribution_date = policy.planned_contribution_date (ISO).
• notice_due_date = contribution_date (notice delivered at/with the contribution).
• withdrawal_window_end = contribution_date + 30 days (30-day Crummey withdrawal right). ISO date arithmetic: add 30 days to the contribution date.
• earliest_premium_payment_date = withdrawal_window_end (pay premium after the withdrawal right lapses so funds are clearly trust property).
• dedicated_bank_account_required = true (ILIT must hold gifts in a dedicated trust account).

estate_result

• death_benefit = policy.death_benefit.
• estate_inclusion_risk = same value as recommendation.risk_flag.
• projected_outside_estate_if_implemented = death_benefit when no 3-year lookback (is_existing_policy_transfer == false); 0 when lookback applies within 3 years.
• tax_liquidity_support = death_benefit (liquidity the ILIT provides to the estate).

recommendation

• risk_flag: if is_existing_policy_transfer AND premium_gap > 0 -> THREE_YEAR_LOOKBACK_AND_EXCLUSION_SHORTFALL; elif lookback -> THREE_YEAR_LOOKBACK; elif premium_gap > 0 -> EXCLUSION_SHORTFALL; else LOW_IF_FORMALITIES_MET.
• primary_action: if lookback and shortfall -> DISCLOSE_LOOKBACK_AND_USE_EXEMPTION; elif lookback -> USE_NEW_POLICY_OR_ACCEPT_LOOKBACK; elif shortfall -> USE_LIFETIME_EXEMPTION_FOR_SHORTFALL; else FUND_WITH_CRUMMEY_NOTICES.
• suitability: SUITABLE_WITH_ADMINISTRATION (no lookback, no shortfall); BORDERLINE (shortfall only); NOT_SUITABLE (lookback present).

source_resolution

controlling_beneficiary_source = SIGNED_PROFILE; controlling_policy_source = CUSTODIAN_EXPORT (default; override to a higher source only if that source explicitly documents the policy).

7. Analysis type C — trust_comparison (GRAT vs CRAT)

Resolved inputs: goals philanthropic_intent, family_transfer_priority (SIGNED_PROFILE); estate_value (SIGNED_PROFILE); trust candidate asset_value (A), expected_growth_rate (g), grat_term_years (nG), grat_annuity_rate (aG), crat_term_years (= max_crat_term_years, e.g. 20), crat_payout_rate (aC); estate_tax_rate, charitable_deduction_rate.

estate_context

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

grat (fixed-amount annuity on initial value)

• term_years = grat_term_years.
• annuity payment P = A * aG per year.
• FV_asset = A * (1+g)^nG.
• FV_annuity = P * ((1+g)^nG - 1) / g.
• projected_remainder_to_heirs = max(0, FV_asset - FV_annuity). Round cents.
• estimated_estate_tax_reduction = projected_remainder_to_heirs * estate_tax_rate. (remainder passes to heirs free of estate tax via the GRAT).
• mortality_inclusion_risk = TERM_SURVIVAL_REQUIRED (always; grantor must survive term).

crat

• term_years = crat_term_years (== max_crat_term_years).
• charity payout C = A * aC per year.
• projected_charitable_remainder = C * ((1+g)^term - 1) / g (FV of charitable payouts). Round cents.
• estimated_income_tax_deduction = A * charitable_deduction_rate. Round cents.
• family_transfer_fit = LOW (CRAT is charity-first; family transfer is incidental).

recommendation

• If family_transfer_priority == "high" -> preferred_strategy = GRAT, rationale_code = CHILDREN_TRANSFER_PRIORITY, alternate_role = SECONDARY_CHARITABLE_TOOL.
• Elif philanthropic_intent == "high" -> preferred_strategy = CRAT, rationale_code = PHILANTHROPIC_PRIORITY, alternate_role = SECONDARY_FAMILY_TRANSFER_TOOL.
• Default (no "high") -> GRAT (family transfer is the harder wealth-transfer problem).

source_resolution

controlling_goal_source = SIGNED_PROFILE; controlling_asset_source = ATTORNEY_MEMO.

8. Analysis type D — estate_liquidity_action_plan

Combines ILIT + trust transfer. Resolved inputs as in B and C, plus estate context.

estate_context

Same as section 7 (taxable_estate, estate_tax_exposure, liquidity_gap_before_planning).

ilit

• annual_exclusion_capacity, premium_gap as in section 6.
• estate_inclusion_risk = the ILIT risk_flag (section 6 logic).
• projected_outside_estate_if_implemented = death_benefit (no lookback) or 0.

trust_transfer

• preferred_strategy = GRAT or CRAT (section 7 goal logic).
• projected_remainder_to_heirs = GRAT remainder (section 7) if GRAT chosen.
• estimated_estate_tax_reduction = projected_remainder_to_heirs * estate_tax_rate.
• projected_charitable_remainder = CRAT charitable FV (section 7) — the alternate charitable tool's figure, reported for the combined plan.

action_set (sorted alphabetically; include applicable subset)

• ATTORNEY_DRAFT_REVIEW — always (trust docs need drafting).
• ILIT_CRUMMEY_NOTICE_CYCLE — if ILIT funded via Crummey (no lookback-blocking issue).
• LIFETIME_EXEMPTION_ALLOCATION — if premium_gap > 0 (shortfall).
• GRAT_FOR_APPRECIATING_SHARES — if preferred_strategy == GRAT.
• CRAT_FOR_CHARITABLE_REMAINDER — if preferred_strategy == CRAT. Sort the included values alphabetically before emitting.

recommendation

• primary_action: if ILIT viable and GRAT preferred -> COMBINE_ILIT_AND_GRAT; if CRAT preferred -> CRAT_WITH_LIQUIDITY_REVIEW; if ILIT shortfall dominates -> 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 = ILIT risk_flag (section 6 logic).

source_resolution

controlling_goal_source = SIGNED_PROFILE; controlling_policy_source = CUSTODIAN_EXPORT.

9. Formatting & rounding rules

• Round every USD amount to cents (2 decimals). Apply rounding only at the final emitted value, not at intermediate simulation steps, to avoid drift.
• Years are integers; ISO dates are YYYY-MM-DD.
• action_set must be sorted alphabetically (estate-liquidity task).
• Numbers are JSON numbers, never quoted strings.
• Use the controlling-source values for every computation; never average conflicting sources.

10. Common mistakes / pitfalls

• Trusting prompt.txt client name over the memo's Client ID: line. The prompt can be stale; the memo is authoritative.
• Forgetting to read Planning horizon year: from the memo (Roth tasks) and defaulting to a wrong horizon.
• Using CRM_NOTE beneficiary_count / annual_non_ira_income instead of the SIGNED_PROFILE values (they disagree by design).
• Computing taxable_estate as gross estate_value instead of max(0, estate_value - exemption).
• Treating conversion_years_positive as equal to conversion_years by default; for a near-RMD client the pre-RMD window is much shorter than the recommended conversion window.
• Ignoring that bracket-fill makes conversion_bracket_targets the conversion-sizing driver (annual = bracket_target - non_ira_income), not an even split.
• Forgetting RMD-must-come-out-first in years that are both conversion and RMD years.
• Setting earliest_premium_payment_date to the contribution date instead of after the 30-day Crummey withdrawal window closes.
• Setting dedicated_bank_account_required to false (an ILIT always needs one).
• Not flagging THREE_YEAR_LOOKBACK when is_existing_policy_transfer == true.
• Putting CRAT/GRAT numbers from the non-preferred strategy in the preferred-strategy fields, or forgetting projected_charitable_remainder in the combined estate plan.
• Emitting action_set unsorted.
• Rounding intermediate values and propagating error; round only final emitted amounts.