README.md

metadata

title: CORP-ENV
emoji: 💼
colorFrom: gray
colorTo: blue
sdk: docker
pinned: false
license: mit
app_port: 7860

CORP-ENV — Shared Workspace Governance for Corporate Planning Agents

CORP-ENV is a highly ambitious yet realistic reinforcement-learning environment designed for long-horizon planning in an enterprise. The agent steps into a Master role (PM, CFO, CEO) and maintains a Shared Workspace Document (SWD) : a structured JSON state while delegating tasks to heavily specialized frozen worker models (dev_agent, hr_agent, finance_agent). The environment boasts a rich, composable reward signal that is intentionally hard to game.

📖 Read our story: The Blog.MD dives deep into our philosophy. Watch it on YouTube.

Training & Models

We've focused purely on training Qwen 2.5 through a comprehensive Base → SFT → RLVR pipeline to demonstrate verifiable improvement in our environment.

• Primary Architecture: The Qwen 2.5-7B Instruct model was trained on H100 servers using the robust scripts provided in the training/ directory. All training logs are safely preserved in training_logs/.
• Google Colab Notebook (T4 Friendly): Check out notebooks/training.ipynb. This notebook has been specifically tested on a T4 instance. To respect memory bounds (OOM issues), it utilizes the smaller Qwen 2.5 3B instruct model.
  • NOTE: other notebooks for differnet models such as deepseek-14B, nemotron-30B in the repository are provided for reference but are not guaranteed or tested to function reliably on Google Colab free tiers.

Qwen 2.5-7B Results (Average over 5 episodes)

Our latest comprehensive evaluation highlights the leap from Base to SFT and robustly to RLVR:

Stage	E1 Reward	M1 Reward	H1 Reward	M1 Success
Base (Qwen 2.5-7B)	0.910	0.707	0.761	0%
SFT (Qwen 2.5-7B)	0.910	0.943	0.882	100%
RLVR (Qwen 2.5-7B)	0.910	0.932	0.779	80%

Environment Actions

Action	Meaning
delegate	Call a worker (agent_id + payload task text).
update_swd	RFC 6902 JSON Patch on the SWD.
query_swd	Read-only JSONPath over the SWD.
log_reasoning	Append a structured reasoning note to the SWD.
log_decision	Append a decision note to the SWD.
log_conflict	Append a conflict object to conflicts_identified.
log_resolution	Append a conflict-resolution object to conflict_resolutions.
advance_phase	Move the SWD phase through analysis, decision, or execution.
finalize	End episode; terminal reward from rich verifiers and OpenEnv rubrics.

Shared Workspace Document (SWD) Structure

The SWD is a rigorous JSON schema defining the exact state of the enterprise episode. Agents must issue valid JSON patches against this structure:

{
  "episode_id": "uuid",
  "scenario": "description of the problem",
  "phase": "discovery | analysis | decision | execution",
  "milestones": [
    {
      "id": "str",
      "label": "str",
      "due_by_turn": 10,
      "status": "pending",
      "owner": "agent_id",
      "output": null
    }
  ],
  "agent_reports": {
    "qa": null,
    "dev": null,
    "hr": null,
    "finance": null
  },
  "decisions": [],
  "conflicts_identified": [],
  "conflict_resolutions": [],
  "reasoning_log": [],
  "final_recommendation": null,
  "swd_version": 0
}

Tasks

ID	Difficulty	Summary
e1_launch_readiness	Easy	48h product launch readiness (QA stability gate).
m1_budget_reallocation	Medium	Budget conflict across dev / HR / finance.
h1_acquisition_defence	Hard	Acquisition defence with injected contradictory intel.

🏆 A Reward Signal That Actually Teaches

A great environment has a reward function that:

Requirement	How CORP-ENV Delivers
Provides a rich, informative signal	Blends Phase Transitions, Conflict Identification, Resolution Logging, and Iterative Validation rather than a 0/1 final score.
Captures something hard to measure	Evaluates how well the model organizes chaos. The strict structure of the SWD and documented reasoning phases provide dense intermediate signals.
Uses Rubric system thoughtfully	The final reward relies on programmatic validations of corporate rigor and is a composition of granular rubric items rather than monolithic success/failure.
Is hard to game	Attempting to skip to finalize, missing milestones, or submitting malformed JSON patches aggressively clamps the reward for agents trying to exploit it.

Reward Breakdown (Terminal, at finalize)

Component	Weight	Evaluation Method
Completion	35%	Verifier
SWD Coherence	25%	Structural
Milestones	20%	On-time
Reasoning	10%	Log entries
LLM Judge	10%	3 YES/NO Qs

Quick Start

# Using uv
uv venv && uv sync
uv run uvicorn server.app:app --host 0.0.0.0 --port 7860

# Or simply
uv run server

Baseline Inference

Run a deterministic E1 smoke test using stub workers. We ensure 100% determinism via CORP_STUB_WORKERS=1 and CORP_DISABLE_LLM_JUDGE=1:

uv run python inference.py
uv run python inference.py --tasks e1_launch_readiness --max-steps 25 --swd-trace logs/run.jsonl

OpenEnv Validation

uv run openenv validate

Docker

docker build -t corp-env .
docker run -p 7860:7860 --env-file .env.example corp-env

License

MIT.