---
license: bsd-3-clause
pretty_name: Eval-rules desk (OpenEnv)
tags:
- openenv
- reinforcement-learning
- agents
- tool-use
configs:
- config_name: tasks
  data_files:
  - split: train
    path: data/tasks.jsonl
---

# Eval-rules desk (Workbench v9.1)

An OpenEnv environment for [OpenEnv Arena](https://openenvarena-arena.hf.space/). It trains the **survival habits** that, in our local proxy of the hidden test, decided whether an episode ever reached an answer:
- one clean JSON object per turn
- small observations
- short scripts
- write the output early, then submit

The tasks are served under the hidden test's published rules.

- **Source:** https://github.com/moarshy/workbench (`server/evalrules_environment.py`)
- **Image:** `ghcr.io/moarshy/workbench@sha256:1ac8bdb927b007bed9bb8db1c4979ee115c6d23f5bd7435ca3d31915406830a1`
- **Design, calibration and pre-registration:** https://github.com/moarshy/openenv-arena-notes/blob/main/versions/v9-eval-rules-desk.md
- **Day-0 proxy:** 23 public SkillsBench tasks, untrained Qwen3.8-27B. 0/91 episodes reached submit; 48 died on invalid replies and 35 on the budget. https://github.com/moarshy/openenv-arena-notes/blob/main/research/2026-10-18-day0-proxy-eval.md

## Interface
Field for field the organizers' PostTrain adapter: `{type: run_bash | submit | finalize, command, answer}`.
- Output is cut to the first 8,192 characters.
- The workspace is `/root`.
- The first observation is the instruction only.
- Budgets: 8,192 completion tokens (counting observations), 16,384 context, 128 tool calls.

## Tasks (9, v9.1)
loans-lite/easy/medium (finance), expenses-lite/medium (office), downtime-medium and reflow-hard (industrial), pyrepo-bug1 and repair-medium (software).
- Data arrives as real **.xlsx** workbooks and **.pdf** documents where a SkillsBench task would use them.
- v9.1 swaps the three tasks that showed no reward spread in v9 training (loans-hard, expenses-easy, pyrepo-implement) for smaller lite levels of the finance and office tasks.

## Reward
```
task_score = share of checks passed, 0 if a protected check fails (schema, inputs unchanged, regressions)
reward = 0.85*task_score + 0.15*(submitted, if task_score >= 0.5) + 0.02*valid actions (max 5) - 0.05*repeated turns (max 0.2)
```
The reward is clipped to [0, 1].

## Licences
Vendored libraries keep their own licences: toolz and boltons are BSD-3-Clause, tabulate and inflection are MIT.