mercor/apex-accounting-harbor

APEX-Accounting — Harbor edition (10-task dev set)

APEX-Accounting is a benchmark built by Mercor in partnership with Ramp to assess whether frontier models can do the real work of accountants: reconciling accounts, accruing expenses, posting transactions, and producing reports. Tasks run inside a self-contained synthetic company world loaded with an accounting system's exports plus spreadsheets, PDFs, and other documents. Every task was authored and solved by practicing accountants and bookkeepers, who also wrote its grading rubric.

This repository is the runnable Harbor rendering of the public 10-task dev set (World 9). It contains the same tasks as mercor/apex-accounting, packaged as Harbor tasks with the three shared runtime images on disk. The scored benchmark (160 tasks across 10 held-out worlds) is closed and not published.

• Tasks: 10 (Reconciliation · Data Entry · Variance Analysis · Schedules & Accruals)
• Worlds: 1 — Sterling, Marsh & Associates LLP, a Philadelphia boutique law firm at its December 2024 month-end close
• Rubric criteria: 89 binary criteria; mean 8.9 per task
• Gold outputs: an expert text answer for every task; all 10 tasks are graded on the agent's final console message
• World assets: 90 shared files (83 workpapers, statements, registers, and invoices plus 7 accounting-system exports) and 16 task-specific files across 8 tasks
• Runtime: Harbor 0.20.0 with the same three shared linux/amd64 images as APEX-Agents 1.1
• License: CC-BY 4.0
• Intended use: task-format inspection, agent development, and evaluation. Dev-set results are not comparable to leaderboard results and should not be reported as APEX-Accounting scores.

Distribution

Channel	Contents
Harbor Hub	10 task packages. World, agent, and verifier images are pinned by digest and pulled from public ECR on first run.
Hugging Face	The same 10 tasks as buildable Harbor tasks, with the three image archives and the world seed on disk (this repository).
Hugging Face (source)	The original tabular release: prompts, rubrics, golden answers, and browsable world and task files.

This repository is a runnable Harbor delivery rather than a standard tabular datasets package.

environment/    Shared world, agent, and verifier image archives
tasks/          10 Harbor task directories
worlds/         1 read-only world seed
manifest.json   Delivery and image inventory
registry.json   Task slug to task-directory mapping
run_task.sh     Validated Harbor task launcher

Each task directory contains its instruction, metadata, agent configuration, environment overlay, Oracle solution, and grader. The world contains filesystem.tar.gz, apps_data.tar.gz, MCP configuration, snapshot hooks, and world metadata.

What differs from the source release

• Prompts, rubrics, and golden answers are exported from the Studio source of record. Three rubric criteria were tightened since the tabular release (tasks 4, 5, and 7); everything else is byte-identical.
• Tool layer. At run time the agent reaches the world only over MCP: Filesystem, Spreadsheets, PDFs, Mail, and Code Execution servers rooted at /filesystem. The accounting-software interface used on the leaderboard is not part of this release. The QuickBooks data ships as static exports in the world's apps_data.tar.gz (seeded to /.apps_data/quickbooks), exactly as in the source release.
• Grading. Each criterion is graded independently by an LLM judge from the task prompt and the agent's final output. The packaged default judge is vertex_ai/gemini-3-flash-preview (the Studio world default); set GRADING_MODEL to override it. The leaderboard judge (DeepSeek-v4-Flash with an unreleased grading template) is not included.

Get Started

Install Docker and uv, then install Harbor:

uv tool install harbor==0.20.0

Model and grader credentials are supplied through environment variables at runtime and are not included in the dataset.

Create a .env file with the API key for your model provider (for example ANTHROPIC_API_KEY, OPENAI_API_KEY, or GOOGLE_API_KEY) and any grader credentials (GRADING_MODEL, GOOGLE_APPLICATION_CREDENTIALS, VERTEXAI_PROJECT, VERTEXAI_LOCATION). The task's configured grader is used when GRADING_MODEL is blank.

Download the delivery and load the three shared images:

hf download mercor/apex-accounting-harbor \
  --repo-type dataset \
  --local-dir apex-accounting-harbor

bash apex-accounting-harbor/environment/load_images.sh

Run a task:

harbor run \
  --env-file .env \
  -p apex-accounting-harbor/tasks/world-9-task-4-baa13987 \
  -a <agent> \
  -m anthropic/claude-opus-5 \
  -y

Run the dev set:

harbor run \
  --env-file .env \
  -p apex-accounting-harbor/tasks \
  -a <agent> \
  -m anthropic/claude-opus-5

Replace <agent> with the Harbor agent you want to evaluate and -m with the model it should use.

Oracle validation

Every task includes a Harbor Oracle solution that replays the expert's golden answer and runs the same isolated verifier:

bash apex-accounting-harbor/run_task.sh world-9-task-4-baa13987 oracle

Citation

@misc{mercor2026apexaccounting,
  title        = {APEX-Accounting},
  author       = {Mercor},
  year         = {2026},
  howpublished = {arXiv},
  url          = {https://arxiv.org/abs/2607.27189}
}

Contact

apex@mercor.com

Legal disclaimer on the content of worlds

This material is provided for research, educational, and informational purposes only. It consists of a hypothetical, simulated company, its books, and illustrative accounting scenarios. No representation is made that any scenario described here reflects an actual entity, transaction, or event. This material does not constitute, and should not be construed as, financial, investment, legal, tax, accounting, or other professional advice. It is provided on an "as-is" and "as-available" basis, and to the maximum extent permitted by applicable law Mercor disclaims liability for losses or damages arising from or related to its use.