Transparent scoring example

See exactly how one task is scored.

This fictional normalized-EBITDA example shows the structure of a Kaelum Bench task, the required response, its ground truth, and deterministic partial credit.

Task type
Financial calculation
Required output
Structured JSON
Scoring
100 deterministic points
Tool access
None

01 · Case input

Income statement extract

USD · FY 2025

All figures needed for the answer are contained in the fictional case.

Line itemAmount
Revenue12,500,000
Cost of goods sold(7,400,000)
Personnel expense(2,150,000)
Other operating expenses(1,120,000)
Other operating income160,000
Depreciation & amortization(430,000)

Supplied facts

  • Plant closure cost: 180,000, completed and one-time.
  • Legal settlement: 95,000, one-time.
  • Recruitment campaign: 70,000, recurring annually.
  • Equipment disposal gain: 140,000.
  • Reported EBITDA already excludes depreciation and amortization.

02 · Question

What the model must do

Fixed before execution
Calculate reported EBITDA and normalized EBITDA. Identify every permitted normalization adjustment, calculate the normalized EBITDA margin, and return the result in the required structure. Do not infer adjustments beyond the supplied facts.
Allowed toolsNo web · no files · no external data

03 · Expected response

Structured output

Reference answer · not a model run
{
  "currency": "USD",
  "reported_ebitda": "1990000.00",
  "adjustments": [
    {
      "id": "plant_closure",
      "direction": "add_back",
      "amount": "180000.00"
    },
    {
      "id": "legal_settlement",
      "direction": "add_back",
      "amount": "95000.00"
    },
    {
      "id": "equipment_disposal_gain",
      "direction": "deduct",
      "amount": "140000.00"
    }
  ],
  "normalized_ebitda": "2125000.00",
  "normalized_ebitda_margin_percent": "17.0"
}

04 · Ground truth

Calculation bridge

Exact values
Reported EBITDA1,990,000
Net adjustments135,000
Normalized EBITDA2,125,000

Normalized EBITDA ÷ Revenue17.0%

05 · Reference checks

Why this answer earns full credit

Deterministic
  • Required structure and currencyPASS
  • Reported EBITDA exactPASS
  • 3 of 3 permitted adjustmentsPASS
  • Recurring item and D&A excludedPASS
  • Normalized EBITDA and margin exactPASS

This explains the reference answer. An actual model response can pass only some checks and receive partial credit.

100 points total

How deterministic partial credit is awarded

Every criterion is evaluated independently and remains connected to an exact expected field, value, or treatment.

5

Valid structure

Required JSON shape and USD currency.

15

Reported EBITDA

Exact value of 1,990,000.00.

36

Three adjustments

Correct ID, amount, and treatment for each item.

8

Correct exclusions

No adjustment for recruitment or depreciation and amortization.

26

Normalized EBITDA

Exact value of 2,125,000.00.

10

Normalized margin

Exact result of 17.0%.

Why partial credit matters

A parseable but incomplete response can still earn the points for exact fields and adjustments it gets right. This example requires no subjective review.

Reference answer100 / 100All criteria passed

How to read a score

What does a benchmark result actually mean?

The task score above is a teaching example. A published benchmark applies the same point principle across every task.

Each taskEarn predefined points
Whole benchmarkSum earned points ÷ sum available points
Observed resultPercentage for one recorded run

A published result describes one recorded model configuration on one benchmark version. It is not an intelligence percentage or a probability of success elsewhere.

Typical errors the scorer catches

These examples show what the scoring contract treats as materially wrong.

Recurring item added back

The annual recruitment campaign is normal operating activity and remains in EBITDA.

D&A adjusted twice

Reported EBITDA already excludes depreciation and amortization.

Gain handled in the wrong direction

The disposal gain is deducted from reported EBITDA, not added back.