ContractSim

ContractSim is the research code for studying LLM negotiation and contract performance in stochastic contracting environments.

Repository map

• negotiation/: runs the customer–supplier LLM negotiation protocol that produces contracts.
• performance/: executes agreed contracts with LLM and rational participants over the delivery weeks.
• domain-robustness/: reruns the fixed Environment 5 negotiation and performance games under Hotel Cleaning and AI Model Hosting terminology.
• algorithms/rc-solver/: exact JAX dynamic-programming solver for contract value under each contingency type.
• algorithms/pareto/: generates and merges F1/F2 Pareto frontiers over contract terms.
• algorithms/synth-contracts/: generates the synthetic performance-ready contract and policy suite.
• algorithms/concession-negotiation/: heuristic concession-negotiation baseline whose contracts are compared against the LLM contracts and Pareto frontiers.
• analysis/: parses, solves, and aggregates negotiation and performance run data into campaign statistics.
• envs/: the six resolved benchmark environments and their model-facing prompt contexts.
• llm_utils/: shared provider, agent, prompting, tracing, and process helpers used across the codebase.

Setup

Python 3.12 is used for development.

python3 -m venv .venv
.venv/bin/python -m pip install -r requirements.txt

Live runs use ANTHROPIC_API_KEY, OPENAI_API_KEY, and either GEMINI_API_KEY or GOOGLE_API_KEY, as applicable. File-backed credentials can instead be stored under the directory selected by KEYS_DIR.

Use --help on an entry point before launching a large campaign.

Citation

@misc{sajja2026evaluatingrationalcontractingnatural,
  title={Evaluating Rational Contracting in Natural Language},
  author={Bhavyesh Sajja and Max Kleiman-Weiner and Roger Zimmermann and Tan Zhi-Xuan},
  year={2026},
  eprint={2608.10475},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2608.10475},
}