| --- |
| viewer: false |
| license: other |
| license_name: errorbar-evaluation-only |
| pretty_name: errorbar — back-office RL environments (sample) |
| tags: |
| - rl-environments |
| - agents |
| - mcp |
| - finance |
| - accounting |
| - spreadsheets |
| - harbor |
| extra_gated_prompt: >- |
| This sample is shared for evaluation only. By requesting access you agree not to redistribute or publish it, |
| and not to use it to train or fine-tune models, without a written licence from errorbar. |
| extra_gated_fields: |
| Name: text |
| Work email: text |
| Company: text |
| Intended use: text |
| I agree to the terms above: checkbox |
| extra_gated_button_content: Request access |
| --- |
| |
| # errorbar — back-office RL environments (sample, 25 tasks) |
|
|
| Five real back-office accounting jobs, five months each, packaged as agentic RL environments. In every task the |
| agent is the desk's employee at month end: it gets the job as it was given, the real mailbox (500–2,300 |
| documents) and the real Excel workbook, both behind MCP tools. It has to find the month's facts in the mail and |
| finish the workbook. |
|
|
| - **Real records, simulated months:** the workbooks and mailboxes are the original documents; each month's |
| figures and planted problems are simulated, so the answers are in no public text. |
| - **Computed grading, no AI judge:** the workbook is recalculated and checked cell by cell against the answer |
| key. Any edit outside the answer cells fails the task. |
|
|
| ## Frontier baseline |
|
|
| GPT-6.1 Sol, jobs completed (strict pass) on each desk's held-out evaluation set (not included in this sample): |
|
|
| | desk | jobs completed | |
| |---|---| |
| | staab | 9/24 (38%) | |
| | ppep | 6/24 (25%) | |
| | griffith | 16/74 (22%) | |
| | oba | 2/12 (17%) | |
| | ces | 1/18 (6%) | |
| | **all five** | **34/152 (22%)** | |
|
|
| **Open-weight models**, jobs completed (partial runs on griffith, staab and ppep; not the held-out sets above): |
|
|
| | model | jobs completed | |
| |---|---| |
| | Qwen3.8-27B | 1/23 (4%) | |
| | Nemotron 3.5 Super VL (fp4) | 0/47 (0%) | |
|
|
| ## Desks |
|
|
| | desk | the job | tiers in this sample | |
| |---|---|---| |
| | staab | gas producer settlements | easy · medium · hard · dense · decoy | |
| | griffith | gas sales billing to a utility | easy · medium · hard · dense · decoy | |
| | ppep | power-plant fuel-management close | easy · medium · hard · dense · decoy | |
| | oba | pipeline imbalance close | hard ×2 · deep · dense · decoy | |
| | ces | agency invoicing | medium ×2 · hard · dense · decoy | |
|
|
| Tiers run from one planted problem (easy) to every problem type of the desk (dense). Decoy months have nothing |
| to fix, but contain documents that tempt a wrong "fix". |
|
|
| ## What it tests |
|
|
| | skill | what the agent must do | |
| |---|---| |
| | Search | find the month's few relevant facts in hundreds to thousands of real documents | |
| | Latest version | use the revised figure, not the first one | |
| | Rules and conventions | apply the desk's standing rules: units, signs, gross vs net, pricing conventions | |
| | Restraint | leave correct figures alone when a document suggests a "fix" | |
| | Inference | work out what no single document states | |
| | Precise editing | change only the right cells in a large, messy real workbook | |
|
|
| ## Grading |
|
|
| Deterministic: the submitted workbook and the answer key are recalculated (LibreOffice) and compared cell by |
| cell. |
|
|
| - **Strict pass:** every answer cell matches, and nothing else was damaged. |
| - **Tolerances:** per task, published with the task. |
| - **Alternative answers:** documented alternatives are accepted where the records leave a convention open. |
| - **Partial credit:** the share of answer cells right is reported as `errors_fixed`. |
|
|
| ## Verified |
|
|
| - The reference solution scores 1.0 and the untouched workbook 0.0 on all 25 tasks; each planted problem left |
| unfixed scores 0. |
| - The grader's recalculation matches Excel on all 32,129 formula cells of the source workbooks. |
| - In a review of 103 failed frontier-model attempts, none failed because of the grader. |
|
|
| ## What's in the download |
|
|
| - 25 tasks in Harbor format: instruction, world (workbook + mailbox behind MCP), grader, reference solution. |
| - A verified reference solution for every task: a complete step-by-step trajectory, built from the answer key, |
| replayed in the live environment and scoring 1.0. Also included as a chat-format SFT example. |
| - README, run guide (including the MDP description) and a spoilers file. |
|
|
| ## Intended use |
|
|
| - **Evaluate** agents on long-horizon, document-heavy back-office work. |
| - **Train** with RL: computed rewards, a partial-credit signal and difficulty tiers from easy to dense. |
| - **Domain:** real US energy-company back office: producer settlements, utility billing, fuel management, |
| pipeline imbalance and agency invoicing. |
|
|
| ## Available on request |
|
|
| - Graded written deliverables per desk. |
| - More tasks: any desk generates new months on demand, at every tier. |
|
|
| ## Privacy and access |
|
|
| - **Privacy:** contact details are pseudonymised throughout. |
| - **Access:** on request, for evaluation only (see the terms above the download). |
| - **Questions:** request access and we'll be in touch. |
|
|
