pooria/benchmark-cost-per-task

Cost per Verified Task: Replay-Based Evaluation of Back-Office Agents

This release contains 30 public tasks and the paper's generated data for cost-per-task. Its scope is replay-based evaluation of back-office agents. The results come from scripted controls and simulations. No learned model or human participant is evaluated.

• Paper: 10.5281/zenodo.23752393
• Code and task generator: https://github.com/pooriaarab/benchmark-cost-per-task
• Grading rules: docs/grading.md
• Author and maintainer: Pooria Arab (pooriaarab)
• Questions: repository issues

The tasks use synthetic fixtures and public benchmark templates. Scores describe the included scripts and simulation assumptions. They do not establish learned-model ability, human performance, or deployment safety.

Files and structure

• data/tasks.jsonl: original public task records, copied without changes.
• data/PROVENANCE.md: original source descriptions, licenses, dates, and generation instructions.
• paper/data/: original generated results, summaries, and verification records.
• paper/README.md: experiment commands from the paper source checkout.
• viewer/tasks.jsonl: a lossless table for the dataset viewer.
• manifest.json: source commit, file sizes, and SHA-256 checksums.

The viewer has four string columns: id, kind, prompt, and task_json. task_json preserves the complete task, including nested fixtures and available reference information. Decode it with json.loads. The original JSONL retains its native structure. The test split identifies public evaluation fixtures. It is not a private holdout. Generated experiment files have separate schemas; they are not additional task splits.

Load the public tasks

import json
from datasets import load_dataset

rows = load_dataset("pooria/benchmark-cost-per-task", "public_tasks", split="test")
tasks = [json.loads(row["task_json"]) for row in rows]
assert len(tasks) == 30

Download all original results with:

from huggingface_hub import snapshot_download

folder = snapshot_download("pooria/benchmark-cost-per-task", repo_type="dataset")

Run experiments from the linked code repository, using its paper instructions. This dataset contains data and documentation; the benchmark repository contains the executable grader.

Limits and intended use

Use these files to inspect tasks and reproduce the paper's scoring analysis. Public prompts and reference information can enter training data. Scores on these fixtures cannot establish performance on unseen tasks. Scripted reference policies can access information unavailable to a normal agent. Simulation parameters describe chosen assumptions, not measured population behavior. The task templates and deterministic grading rules limit the conclusions. Read the paper for benchmark-specific limits and uncertainty methods.

Excluded from this release: Private and hidden task splits; data/prices.json: external provider snapshot with NOASSERTION source license. Original provenance files can mention sources excluded from this release. Source provenance labels describe fixture authorship, not evidence of a participant study.

License and attribution

This dataset release and card use CC-BY-4.0. Attribute Pooria Arab and cite the linked paper and code repository. Original source notices remain in data/PROVENANCE.md and SOURCE-LICENSE. Any MIT or CC0 permissions already attached to source material remain available. The public tasks and paper results are unchanged copies. Only the viewer table, card, and release manifest are added.

Citation

@misc{arab2026costpertask,
  author = {Pooria Arab},
  title = {Cost per Verified Task: Replay-Based Evaluation of Back-Office Agents},
  year = {2026},
  doi = {10.5281/zenodo.23752393},
  url = {https://doi.org/10.5281/zenodo.23752393}
}