Task Group 009 Fewshot Attempt 03

Task Group 009 Finance Ops SOP

Use this skill to solve new Finance Ops reporting tasks from their active input payloads. Do not use train answers as a lookup table. Always derive results from the request memo, answer template, and the public endpoints listed in the task input.

Access Protocol

1. Read only the task input artifacts: prompt.txt, payloads/request_memo.json, payloads/environment_access.json, and payloads/answer_template.json.
2. Get base_url and available_endpoints from environment_access.json.
3. Query only listed public endpoints under that base URL. Typical endpoints are:
   • Finance: /api/finance/branches, /api/finance/period-map, /api/finance/accounts, /api/finance/records
   • Compensation: /api/compensation/rate-book, /api/compensation/rosters, /api/compensation/scenarios
   • Payroll: /api/payroll/rate-book, /api/payroll/productions
4. Use request_memo.json to identify the target branch, region, ensemble, production, scenario, periods, and requested view.
5. Use answer_template.json as the output contract. Return one JSON object only, with exactly the requested keys and numeric JSON values.
6. Do not read test outputs, notes, eval files, evaluator code, or environment internals.

PowerShell note: Invoke-RestMethod often returns JSON arrays directly. Do not assume a .value wrapper exists just because ConvertTo-Json displayed one.

Rounding And Ordering

Use raw precision for all intermediate math. Round only values emitted in the final JSON.

from decimal import Decimal, ROUND_HALF_UP

def round_money(x):
    return float(Decimal(str(x)).quantize(Decimal("0.01"), rounding=ROUND_HALF_UP))

def round_ratio(x):
    return float(Decimal(str(x)).quantize(Decimal("0.0001"), rounding=ROUND_HALF_UP))

• Currency fields: round to 2 decimals.
• Percent and ratio fields: round to 4 decimals as decimal ratios, for example 0.0656.
• Aggregate totals from raw unrounded components, then round. Do not sum already rounded display rows.
• Sort stable ID lists ascending unless the template states a ranking order.
• For rankings, sort by metric descending, then stable ID ascending, and use the 1-based position after sorting.
• For ties in "largest" fields, use the rate-book pay type order or stable ID order unless the prompt says otherwise.

Finance Reporting

Data Model

• branches: branch metadata with branch_id, branch_name, region_id, and region_name.
• period-map: maps period labels such as M1 to fiscal years and month numbers.
• accounts: maps each account to a category such as revenue, cogs, sga, allocations, or operating.
• records: one row per branch/account with a values object keyed by period label.

Build fiscal-year period sets from period-map, not from calendar assumptions. If the request gives a close period, the current fiscal year is the fiscal year of that period; compare it with the prior fiscal year unless the memo specifies comparison years.

Account Rollups

For a branch, region, month, or fiscal-year period set:

• revenue = sum(accounts where category == "revenue")
• cogs = sum(accounts where category == "cogs")
• gross_margin = revenue - cogs
• sga = sum(accounts where category == "sga")
• allocations = sum(accounts where category == "allocations")
• ebitda = gross_margin - sga - allocations
• ebitda_margin = ebitda / revenue
• revenue_growth_pct = (current_revenue - prior_revenue) / prior_revenue
• ebitda_growth_pct = (current_ebitda - prior_ebitda) / prior_ebitda
• arpu = revenue / sum(active_customers over the same periods)
• sales_per_labor_headcount = revenue / sum(labor_headcount over the same periods)

Use the operating denominators as period sums. Do not use average active customers, average headcount, orders, or revenue units for these two ratios.

Branch Close View

For a target branch close package:

1. Resolve the target branch from /api/finance/branches.
2. Resolve current and prior periods from the memo.
3. Build period_convention from period-map, for example fiscal-year ranges and the current/prior period labels.
4. Build the current-month income statement for the close period.
5. Build month-over-month revenue variance as current period revenue minus prior period revenue, plus percent over prior period revenue.
6. Build current fiscal-year metrics and current-vs-prior fiscal-year growth.
7. Build regional context for the target branch's region:
   • branch_ids: all branches in the region, sorted ascending.
   • fy####_ebitda: sum the region's branch EBITDA for the current fiscal year.
   • ebitda_rank_desc: rank the region among all regions by current fiscal-year EBITDA descending.
8. Build branch rankings across all branches:
   • sales_growth_rank_desc: rank the target branch by current-vs-prior fiscal-year revenue growth.
   • top_sales_growth_branch_id: branch with the highest current-vs-prior fiscal-year revenue growth.
   • top_arpu_branch_id: branch with the highest current fiscal-year ARPU.

Regional View

For a regional management report:

1. Resolve target_region_id and collect included branches sorted by branch_id.
2. For each requested comparison year, sum branch metrics across the region.
3. Emit the requested fiscal-year blocks, usually fy2024 and fy2025.
4. Compute current-year ebitda_margin and sales_per_labor_headcount from region totals.
5. Compute revenue_growth_pct between the requested years.
6. Find top_ebitda_branch_id and bottom_ebitda_branch_id within the target region using current-year branch EBITDA.
7. Compute region_reconciliation_variance = region_ebitda_from_region_total - sum(branch_ebitda_for_included_branches). When both sides come from the same records, this should round to 0.0.

Pitfall: in the branch close view, region_context.ebitda_rank_desc is the region's rank among all regions, not the target branch's rank within its region. In the regional view, top and bottom branch IDs are within the selected region.

Compensation Summaries

Rate Book And Roster Fields

Use /api/compensation/rate-book for:

• current_year
• minimum_weekly_scale
• pay_types order
• quarter_weeks
• title_premium_pct
• seniority_weekly bands
• business rules

Use /api/compensation/rosters for employees with:

• ensemble_id, employee_id, title
• years_of_service
• overscale_weekly
• combined_overscale_includes_title
• weeks_by_quarter
• notes

Select only the requested ensemble_id.

Current-Year Compensation

For each employee and quarter:

weeks = employee.weeks_by_quarter[quarter]
mws = minimum_weekly_scale * weeks
title = minimum_weekly_scale * title_premium_pct[title] * weeks
seniority = seniority_weekly_for(years_of_service) * weeks
overscale = overscale_weekly * weeks

If title is null, title premium is zero. If combined_overscale_includes_title is true, do not add a separate title premium for that employee, but still include the overscale_weekly amount in Overscale.

Build:

• roster_count: count selected roster rows.
• pay_types: exact order from the rate book.
• quarter_totals: sum all pay types by quarter.
• annual_pay_type_totals: sum all quarters by pay type.
• annual_total: sum annual pay type totals.
• largest_pay_type: pay type with the largest annual total.
• combined_overscale_employee_count: count employees where combined_overscale_includes_title is true. Do not count employees merely because overscale_weekly > 0.
• partial_quarter_employee_count: count employees where any weeks_by_quarter[Q] differs from rate_book.quarter_weeks[Q].

Compensation Forecasts

Use /api/compensation/scenarios only when the request includes a scenario_id.

Compute three annual layers when requested:

• current: current rate book and current years of service.
• year_plus_1: apply Year + 1 scenario growth; add 1 service year before seniority band lookup.
• year_plus_2: compound Year + 1 and Year + 2 scenario growth; add 2 service years before seniority band lookup.

For Year + 2, multiply growth factors cumulatively:

mws_multiplier_y2 = (1 + y1.mws_growth) * (1 + y2.mws_growth)
seniority_multiplier_y2 = (1 + y1.seniority_growth) * (1 + y2.seniority_growth)
overscale_multiplier_y2 = (1 + y1.overscale_growth) * (1 + y2.overscale_growth)
title_pct_multiplier_y2 = y1.title_pct_multiplier * y2.title_pct_multiplier

Forecast pay formulas:

scale_rate = minimum_weekly_scale * mws_multiplier
mws = scale_rate * weeks
title = scale_rate * title_premium_pct[title] * title_pct_multiplier * weeks
seniority = seniority_weekly_for(years_of_service + service_year_add) * seniority_multiplier * weeks
overscale = overscale_weekly * overscale_multiplier * weeks

Apply the same combined-overscale rule as current-year compensation.

Build:

• annual_totals.current, annual_totals.year_plus_1, annual_totals.year_plus_2.
• growth_rates.year_plus_1_vs_current = (year_plus_1_total - current_total) / current_total.
• growth_rates.year_plus_2_vs_year_plus_1 = (year_plus_2_total - year_plus_1_total) / year_plus_1_total.
• year_plus_2_quarter_totals and year_plus_2_pay_type_totals from Year + 2 calculations.
• largest_growth_pay_type: pay type with the largest percent increase from current annual pay-type total to Year + 2 annual pay-type total. This is a percentage-growth driver classification, not the largest absolute dollar increase.
• Roster treatment counts using the selected current roster.

Pitfalls:

• Use roster quarter weeks, not a fixed 13 weeks.
• Recalculate seniority bands after adding forecast service years.
• Compound scenario growth for Year + 2.
• Skip separate title premium when combined_overscale_includes_title is true.

Theatre Payroll Review

Rate Book And Production Fields

Use /api/payroll/rate-book for:

• service_rates
• premium_pct
• weekly_guarantee
• service_time_limits
• conflict_thresholds
• business rules

Use /api/payroll/productions to select the requested production_id. A production has schedule services and a roster with each musician's assigned service IDs and flags.

Service Counts

Count services from the production schedule, not from roster assignments. Emit service type keys sorted alphabetically unless the template says otherwise.

Per-Musician Pay

For each musician:

1. Build base service categories from assigned services:
   • Rehearsal: service_rates["Rehearsal"] * max(duration_hours, 3.0).
   • Performance: flat service_rates["Performance"].
   • Audit: flat service_rates["Audit"].
   • Any sound check: flat service rate for that exact sound-check service type, category sound_check.
2. If substitute is true:
   • Add 2 * service_rates["Performance"] to the performance category.
   • Include that same amount in base_service_pay for premium and doubles calculations.
   • Add 2 * service_rates["Performance"] as a separate substitute_adjustment category.
   • Do not apply the weekly guarantee.
3. Compute role premium on base_service_pay using the sum of applicable role percentages:
   • principal_or_lead once when either principal or lead is true.
   • quartet when true.
   • electronic when true.
   • concertmaster only if a roster field exposes that flag.
4. Compute doubles on base_service_pay:
   • First extra instrument: first_double.
   • Each additional extra instrument: additional_double.
   • Formula: base_service_pay * (first_double + max(0, doubles - 1) * additional_double).
5. Compute guarantee adjustment for non-substitute regular players:
   • max(0, weekly_guarantee - base_service_pay).
   • Exclude premiums, doubles, vacation, and adjustment categories from this comparison.
6. Compute vacation only when vacation_eligible is true:
   • vacation_pct * (base_service_pay + premium + doubles).
   • Exclude guarantee adjustment and the separate substitute adjustment line.
7. Omit zero categories from each musician's categories object.

Aggregate payroll category totals from raw per-musician category values, then round. Do not aggregate from rounded musician display values.

Payroll Output Fields

• category_totals: sum raw category amounts by category, sorted by category name when order is not specified.
• weekly_total: sum all raw category totals.
• per_musician: order by musician_id; include musician_id, name, rounded total, and rounded nonzero categories.
• top_paid_musician_id: highest raw musician total, tie by musician_id ascending.

Conflict Flags

Build a set of flags from the production schedule, then sort alphabetically:

• REHEARSAL_EARLY_START: any rehearsal starts before conflict_thresholds.rehearsal_earliest_start.
• REHEARSAL_LATE_END: any rehearsal ends after conflict_thresholds.rehearsal_latest_end.
• SERVICE_OVER_TIME_LIMIT: any service has duration_hours > service_time_limits[service_type].
• SOUND_CHECK_DURATION_MISMATCH: any sound-check service has duration_hours != service_time_limits[service_type].

Sound-check mismatch and over-time are independent. A long sound check can trigger both if its duration exceeds the limit.

Final Validation Checklist

Before finalizing:

1. Confirm every required key from answer_template.json is present and no extra top-level keys are added.
2. Confirm all currency, percent, and ratio fields use the template's required rounding.
3. Confirm lists and per-person rows use required stable ordering.
4. Confirm rankings are computed across the correct population: all branches, all regions, or selected region branches as requested.
5. Confirm compensation counts distinguish partial-quarter service from combined overscale side-letter treatment.
6. Confirm payroll totals aggregate raw values, not rounded display rows.
7. Return only the final JSON object unless the user explicitly asks for explanation.