Fewshot Attempt 02

Crescent Finance Ops — Transferable Skill

Environment

All tasks use a single remote Finance Ops API. The base URL is provided in the task's environment_access.json (or overridden by environment_access.md). Always use HTTP, never HTTPS. Do not use localhost unless the override explicitly points there.

API Reference

Finance Domain

Period Convention:

• M1 – M12 → earlier fiscal year (typically FY2024)
• M13 – M24 → later fiscal year (typically FY2025)
• Current year's mapping is always available from the period-map endpoint; query it, do not hardcode.

Account Categories:

Compensation Domain

Payroll Domain

Calculation Formulas

1. Branch Income Statement (per period or per FY sum)

Revenue        = sum(product_revenue + service_revenue)
COGS           = sum(direct_materials_cogs + direct_labor_cogs)
Gross Margin   = Revenue − COGS
SG&A           = sum(sales_sga + admin_sga + occupancy_sga)
Allocations    = sum(shared_service_allocations)
EBITDA         = Gross Margin − SG&A − Allocations

Per-period: use values for a single period key (e.g. M24). Per-FY: sum values across all 12 period keys in that fiscal year (e.g. M13–M24 for FY2025).

2. Branch Ratios and Derived Metrics (FY level only)

EBITDA Margin             = EBITDA / Revenue

ARPU (Avg Revenue Per User/Active Customer)
  = FY Revenue / Σ active_customers across all 12 FY periods
  Use the sum of the count metric, not an average.

Sales per Labor Headcount
  = FY Revenue / Σ labor_headcount across all 12 FY periods
  Use the sum of the count metric, not an average.

3. MoM Revenue Variance (Branch)

amount = revenue(current_period) − revenue(prior_period)
pct    = amount / revenue(prior_period)

4. Period-over-Period Growth Rates (Branch or Region)

revenue_growth_pct = (FY{later}_revenue − FY{earlier}_revenue) / FY{earlier}_revenue
ebitda_growth_pct  = (FY{later}_ebitda  − FY{earlier}_ebitda)  / FY{earlier}_ebitda

5. Regional Aggregation

Sum each account across all branches in the region, then apply the income statement formulas above per FY. Branch membership is determined from /api/finance/branches — filter by region_id.

Region-level revenue, sga, allocations, and ebitda are reported for each FY. Additionally FY2025 needs ebitda_margin and sales_per_labor_headcount computed at the region level (aggregate revenue ÷ aggregate metric sum).

region_reconciliation_variance should always be 0.0 — regional totals are the direct sum of branch-level data.

6. Branch Rankings

Sales Growth Rank Descending: compute revenue_growth_pct for every branch, sort descending (highest = rank 1), report the target branch's rank.

Top ARPU Branch: compute ARPU for every branch, pick the branch with the highest value.

Region EBITDA Rank Descending: compute FY2025 EBITDA for each branch in the target region, sort descending, report the target branch's rank (1 = highest EBITDA in region).

Top/Bottom EBITDA Branch in Region: the branch with highest and lowest FY2025 EBITDA in the region.

7. Compensation — Current Year Summary

For each employee in the target ensemble, per quarter:

Base                = weeks_in_quarter × minimum_weekly_scale

Title Premium       = weeks_in_quarter × minimum_weekly_scale × title_premium_pct[title]
                      SKIP if combined_overscale_includes_title is true
                      SKIP if title is null/None

Seniority           = weeks_in_quarter × seniority_weekly_amount
                      Find the band where years_of_service is between
                      min_years and max_years (max_years may be null = no upper bound)

Overscale           = weeks_in_quarter × overscale_weekly

Quarter total for an employee = Base + Title Premium + Seniority + Overscale. Quarter total for the ensemble = sum of all employee quarter totals. Annual total = sum(Q1 + Q2 + Q3 + Q4).

Annual pay type totals = sum of each pay type component across all employees and all quarters.

largest_pay_type: the pay type with the highest annual total (enum string, not the dollar amount).

roster_count: count of employees in the ensemble roster.

combined_overscale_employee_count: count of employees where combined_overscale_includes_title is true.

partial_quarter_employee_count: count of employees where any quarter has weeks < standard_quarter_weeks[that_quarter]. The standard is from the rate book's quarter_weeks.

pay_types: ordered list matching the rate book's pay_types array.

8. Compensation — Forecast (Current, Year+1, Year+2)

Current year: same as current year summary (Section 7).

Year+1: apply the scenario's year_plus_1 growth multipliers to the rate book base values, then add 1 year to each employee's years_of_service before looking up seniority bands. Keep the roster structure identical.

Year+1 MWS          = current MWS × (1 + mws_growth)
Year+1 seniority    = current seniority_weekly_amount × (1 + seniority_growth)
Year+1 overscale    = employee.overscale_weekly × (1 + overscale_growth)
Year+1 title pct    = title_premium_pct × title_pct_multiplier
Year+1 yos          = employee.years_of_service + 1

Then recompute per-employee quarterly totals using Year+1 parameters. Annual total is the sum.

Year+2: apply year_plus_2 growths on top of Year+1 values, add 2 years total to years_of_service.

Growth Rates (between scenarios):

year_plus_1_vs_current = (Y1_total − current_total) / current_total
year_plus_2_vs_year_plus_1 = (Y2_total − Y1_total) / Y1_total

year_plus_2_pay_type_totals: the breakdown by pay type for Year+2 only (same structure as annual_pay_type_totals but for the forecast year).

year_plus_2_quarter_totals: quarterly totals for Year+2 only (Q1–Q4).

largest_growth_pay_type: the pay type with the largest absolute dollar increase from current to Year+2 (current pay_type_total → year_plus_2 pay_type_total). Return the enum string.

combined_overscale_employee_count and partial_quarter_employee_count: use current roster data (same as the current year task).

9. Payroll — Weekly Package

For each musician, for each assigned service:

Base Service Pay:
  - Performance / Audit: service_rates[service_type] per service
  - Sound Check: service_rates[service_type] per service
  - Rehearsal: service_rates["Rehearsal"] × max(duration_hours, 3.0)
               (3-hour minimum call applies)

Premium multipliers (applied to Base Service Pay, additive):
  - principal_or_lead: 0.15  if principal=true or lead=true
  - concertmaster:    0.20  if the musician is designated concertmaster (check roster title or role)
  - electronic:       0.25  if electronic=true
  - quartet:          0.15  if quartet=true
  - first_double:     0.25  if doubles ≥ 1
  - additional_double:0.10  per each double beyond the first (i.e. (doubles-1) × 0.10)

Premiums = Base Service Pay × Σ(applicable premium pcts)

Vacation = 0.04 × (Base Service Pay + Premiums)  if vacation_eligible=true

Weekly Guarantee Adjustment:
  If substitute=false AND Base Service Pay < weekly_guarantee:
    guarantee_adjustment = weekly_guarantee − Base Service Pay
  Otherwise: 0

Substitute Adjustment:
  For substitute=true: substitute earns Base Service Pay + Premiums,
  but substitute_adjustment is the amount the regular guarantee would have
  been (i.e. weekly_guarantee − Base Service Pay if Base Service Pay < guarantee,
  else 0). This is a separate category.

Musician total = sum of all categories for that musician.

Service Counts: map each distinct service_type from the schedule to its count of occurrences. Use the schedule's own service_type strings.

Conflict Flags (check ALL services in the schedule, not just assigned ones):

Conflicts are flagged against the schedule globally (not per-musician, not just assigned services).

Top Paid Musician: the musician with the highest total pay. If tie, use the lower musician_id.

10. Substitute Payroll Details

Substitutes (substitute=true):

• Get base service pay + premiums (same rates as regular players)
• Do NOT get vacation pay
• Do NOT get guarantee adjustments
• Do get a substitute_adjustment (the amount they would have received as a guarantee adjustment if they were regular — computed as max(0, weekly_guarantee − base_service_pay))

Non-substitutes (substitute=false):

• Get base service pay + premiums + vacation (if eligible) + guarantee_adjustment (if applicable)
• Get a guarantee_adjustment equal to max(0, weekly_guarantee − base_service_pay)

Output Conventions

Rounding

• Currency values: round to 2 decimal places
• Percent and ratio values: round to 4 decimal places
• Use standard round() (round half to even / banker's rounding)

Ordering

• Lists of branch IDs: ascending string sort (e.g. ["BR-004","BR-005","BR-006"])
• per_musician: ascending by musician_id
• conflict_flags: alphabetical sort
• pay_types: the order from the compensation rate book's pay_types array
• All lists use ascending stable identifiers unless a rank order is explicitly requested

Field Naming

• JSON keys are always snake_case as shown in answer templates
• null is never used for numeric fields — use 0 or 0.0 instead
• Empty lists use [], not null

Common Pitfalls

1. Period convention is dynamic: always read /api/finance/period-map to determine which periods belong to which FY. Do not assume M13 always equals FY2025.
2. ARPU and sales_per_labor_headcount use summed counts, not averages: sum the count metric across all 12 periods in the FY, then divide.
3. Forecast seniority: add years of service BEFORE looking up the seniority band. Year+1 adds 1, Year+2 adds 2. Crossing a band boundary can cause disproportionate growth.
4. combined_overscale_includes_title: when true, skip the separate title premium. The overscale amount already bundles it.
5. Rehearsal minimum: rehearsal pay is per-hour with a 3-hour minimum per rehearsal service, not per day.
6. Premiums are additive: a musician can have principal, doubles, electronic, and quartet all at once. Each applicable premium pct is added together before multiplying base pay.
7. Vacation base: vacation is 4% of (base service pay + all premiums), not just base pay.
8. Conflict flags are schedule-level, not musician-level: check every service in the schedule, even if no musician is assigned to it.
9. Substitute adjustment is a positive category total: it represents pay the substitute actually receives.
10. Region reconciliation variance is always 0.0: the data is internally consistent; simple summation of branch data yields the regional totals.