01 · Task
Objective and constraints
Choose the answer whose budget calculation is correct and easiest to audit.
- Use only the figures in the task packet.
- Apply contingency after fixed costs and keep cents through the final step.
- Evaluate one captured run per model; do not infer a general model ranking.
02 · Reproduction input
Complete prompt
A launch budget is $1,200. Fixed costs are $249 for software and $180 for research. Set aside a contingency equal to 15% of the money remaining after fixed costs. How much remains for ads? Show the calculation and state the final dollar amount.
03 · Evaluation scope
Criteria used for this task
The remaining amount after fixed costs is $771. Contingency is $115.65, leaving exactly $655.35 for ads. Strong answers show the calculation without rounding early.
- Check the fixed-cost subtotal.
- Apply 15% only after fixed costs.
- Reject early rounding or an unsupported total.
04 · Immutable snapshots
Captured outputs and review
Start with the comparison summary. Open a snapshot only when you need the full output, run metadata, and editorial review.
OpenAI
openai/gpt-oss-20b
- Strength
- Complete, labeled calculation
- Review boundary
- Longer than necessary for a small calculation
Qwen
qwen/qwen3-30b-a3b-fp8
- Strength
- Concise and fully checkable
- Review boundary
- Offers less explanation of why contingency is applied after fixed costs
Mistral AI
mistralai/mistral-small-3.1-24b-v2
- Strength
- Correct sequence and result
- Review boundary
- The final unlabeled amount is less scannable
Output 1Raw captured output · collapsed by defaultOpenAI · openai/gpt-oss-20b
Raw output
Calculation
- Total budget: $1,200
- Fixed costs:
- Software: $249
- Research: $180
- Total fixed costs = $249 + $180 = $429
- Money remaining after fixed costs:
$1,200 – $429 = $771 - Contingency (15 % of $771):
0.15 × $771 = $115.65 - Amount left for ads:
$771 – $115.65 = $655.35
Final dollar amount available for ads: $655.35
meets
Strengths: Complete, labeled calculation Keeps cents through the final step
Risks: Longer than necessary for a small calculation
Best use: When an auditable calculation matters most.
Output 2Raw captured output · collapsed by defaultQwen · qwen/qwen3-30b-a3b-fp8
Raw output
Total budget: $1,200
Fixed costs: $249 + $180 = $429
Remaining after fixed costs: $1,200 - $429 = $771
Contingency: 15% of $771 = 0.15 × $771 = $115.65
Amount remaining for ads: $771 - $115.65 = $655.35
Final answer: $655.35
meets
Strengths: Concise and fully checkable States the correct $655.35 result
Risks: Offers less explanation of why contingency is applied after fixed costs
Best use: When speed and traceable arithmetic both matter.
Output 3Raw captured output · collapsed by defaultMistral AI · mistralai/mistral-small-3.1-24b-v2
Raw output
Fixed costs: $249 + $180 = $429 Remaining after fixed costs: $1,200 - $429 = $771 Contingency: 15% of $771 = 0.15 * $771 = $115.65 Remaining for ads: $771 - $115.65 = $655.35
$655.35
meets
Strengths: Correct sequence and result No unnecessary claims
Risks: The final unlabeled amount is less scannable
Best use: When a compact calculation is sufficient.
05 · Claim ledger
Claims, sources, evidence, and boundaries
Claim 1
The available advertising budget is $655.35.
- Source
- Published task packet and deterministic arithmetic
- Support
- $249 + $180 = $429; $1,200 − $429 = $771; 15% of $771 = $115.65; $771 − $115.65 = $655.35.
- Boundary
- This verifies the supplied arithmetic only; it does not validate the business assumptions.
Claim 2
All three captured outputs reach the same final amount.
- Source
- Immutable output snapshots in this case
- Support
- Each displayed raw output states $655.35 after following the same fixed prompt.
- Boundary
- One output per model does not establish a stable model-level performance rate.
06 · Review conclusion
What this evidence supports
Prefer the answer whose intermediate values make the final number easiest to verify—not a model name.
Known limitations
- The task is deterministic and does not represent open-ended financial analysis.
- The review compares one captured output per model under one prompt.
- Formatting preference is evaluated only as it affects auditability.
Reusable takeaway
Use the intermediate-value audit checklist before accepting a generated calculation.
Use the supporting method