Law in Silico: Simulating Legal Society with LLM-Based Agents

Repository at a glance

Path	Role
legal_simulation/auto_main.py	Micro labor/company command-line entrypoint.
legal_simulation/main.py	Labor/company micro simulation bootstrap.
legal_simulation/simulation.py	Labor/company micro simulation runtime.
legal_simulation/config.py	Labor/company micro configuration.
legal_simulation/auto_main_pollution.py	Micro pollution command-line entrypoint.
legal_simulation/simulation_pollution.py	Pollution micro simulation runtime.
legal_simulation/config_pollution.py	Pollution micro configuration.
legal_simulation/macro_simulation/main.py	Generic macro simulation runner.
legal_simulation/macro_simulation/run_theft_baseline.py	Specialized macro theft baseline experiment.
legal_simulation/macro_simulation/scenes/	Macro-format scene files.
legal_simulation/utils/micro_macro_bridge.py	Optional Micro-Macro Bridge API.
tests/	No-network release smoke tests.
requirements.txt	Dependencies for the smoke path and non-vLLM components.

Three code paths

Micro simulation

The micro path is for small-agent, multi-step interaction. It is not a macro-style one-shot multiple-choice workflow.

The active micro path currently covers two active micro scene families, not exactly two fixed micro scenes:

• Labor/company
• Pollution/factory-resident

Each micro family may have sub-scene or hyperparameter variants.

Labor/company

Entrypoint:

cd legal_simulation && python auto_main.py --WHICH_EXP basic_exp --has-judge --COURT_BIAS neutral --LABOR_TRUST_LAWS not_available --DETERRENCE_OF_LAWS not_available

Code path:

• legal_simulation/auto_main.py
• legal_simulation/main.py
• legal_simulation/simulation.py

Configuration:

• legal_simulation/config.py

Agents:

• Company
• Laborer

Decision format:

• Open-ended / XML-style action response, for example <response><think><action>...
• Not multiple-choice

This is a real-LLM entrypoint. It requires environment configuration and should not be treated as a no-network smoke command.

Pollution/factory-resident

Entrypoint:

cd legal_simulation && python auto_main_pollution.py --EXP_NAME pollution_exp_001

Code path:

• legal_simulation/auto_main_pollution.py
• legal_simulation/simulation_pollution.py

Configuration:

• legal_simulation/config_pollution.py

Agents:

• Factory
• Resident

Decision format:

• Flat JSON action response with action, param, and reason
• Not multiple-choice

This is also a real-LLM entrypoint. It requires environment configuration and should not be treated as a no-network smoke command.

Micro Experiment Matrix

Active micro readiness currently covers two scene families: labor/company and pollution/factory-resident. Paper-intended families and conditions are experiment design labels; this matrix lists selected core runtime defaults and CLI knobs used by readiness checks, not the full runtime surface.

Family	Runtime defaults	Runtime knobs
Labor/company	NUM_LABORERS=3, SIMULATION_MONTHS=4, NUM_ACTIONS_PER_MONTH=2, SEED=42, DETERRENCE_OF_LAWS=high, empty INITIAL_LAW_CODES	HAS_JUDGE via --has-judge / --no-judge, --COURT_BIAS, --LABOR_TRUST_LAWS, --DETERRENCE_OF_LAWS
Pollution/factory-resident	NUM_RESIDENTS=3, SIMULATION_MONTHS=7, NUM_ACTIONS_PER_MONTH=2, TEMPERATURE=1, INITIAL_FACTORY_CASH=20000.0, INITIAL_RESIDENT_CASH=800.0, BASE_REVENUE=14000.0 per turn	--NUM_RESIDENTS, --SIMULATION_MONTHS, --NUM_ACTIONS_PER_MONTH, --INITIAL_FACTORY_CASH, --INITIAL_RESIDENT_CASH, --BASE_REVENUE, --TEMPERATURE

Labor/company scale defaults such as laborer count, months, actions per month, seed, and initial law codes are config/in-process settings only; they are not exposed by auto_main.py CLI. The labor/company runtime default includes DETERRENCE_OF_LAWS=high; neutral/no-deterrence baseline commands must override it with --DETERRENCE_OF_LAWS not_available. Labor bias or trust variants must set DETERRENCE_OF_LAWS=not_available unless the run is testing deterrence itself, because the prompt configuration rejects combining deterrence variants with bias/trust variants.

Pollution NUM_ACTIONS_PER_MONTH can be changed through auto_main_pollution.py, but several derived per-turn constants are computed when config_pollution.py is imported. Keep the default for calibrated runs and lightweight smokes unless the code is updated to recompute those derived constants after CLI overrides.

Macro simulation

Macro simulation is an independent many-agent, single-step decision workflow. It is not the natural required continuation of micro. It does not require the bridge for all runs.

Generic macro entrypoint:

cd legal_simulation/macro_simulation && python main.py --scene_path scenes/theft.json --output_path ../../outputs/macro_results.json --model_path <local-vllm-model>

Optional law-code mode can add a macro scene law-code file, for example:

--law_codes_path scenes/micro_to_macro/scene_1/law_codes/laws_middle.json

Macro code path:

• legal_simulation/macro_simulation/main.py

Macro scene files include:

• legal_simulation/macro_simulation/scenes/theft.json
• legal_simulation/macro_simulation/scenes/assault.json
• legal_simulation/macro_simulation/scenes/sexual.json
• legal_simulation/macro_simulation/scenes/micro_to_macro/scene_1/*
• legal_simulation/macro_simulation/scenes/micro_to_macro/scene_2/*

The macro_simulation/scenes/micro_to_macro/* files are macro-format scenes for bridge-related experiments. They are not the active micro simulators.

Real macro runs require vLLM, GPU access, and a local model setup. requirements.txt does not install vLLM.

Macro theft baseline

The theft baseline is a specialized theft experiment, not the generic macro runner.

Entrypoint:

cd legal_simulation/macro_simulation && python run_theft_baseline.py --model_path <local-vllm-model> --output_dir ../outputs/theft_baseline

Micro-Macro Bridge

The Micro-Macro Bridge is a utility/API component for scale-transfer or mapping experiments. It is optional and lightly integrated. It is not the main micro runtime, and it is not the main macro runtime.

Bridge API:

• legal_simulation/utils/micro_macro_bridge.py::MicroMacroBridge

Known methods:

• export_micro_simulation
• create_macro_scene_from_micro
• load_micro_export
• apply_micro_laws_to_macro

There is no committed bridge CLI. Do not treat bridge use as required for ordinary macro runs.

Entrypoints table

Code path	Entrypoint	Runtime expectation
Micro labor/company	cd legal_simulation && python auto_main.py --WHICH_EXP basic_exp --has-judge --COURT_BIAS neutral --LABOR_TRUST_LAWS not_available --DETERRENCE_OF_LAWS not_available	Real LLM/API configuration required.
Micro pollution	cd legal_simulation && python auto_main_pollution.py --EXP_NAME pollution_exp_001	Real LLM/API configuration required.
Macro generic runner	cd legal_simulation/macro_simulation && python main.py --scene_path scenes/theft.json --output_path ../../outputs/macro_results.json --model_path <local-vllm-model>	Local vLLM/GPU/model setup required.
Macro theft baseline	cd legal_simulation/macro_simulation && python run_theft_baseline.py --model_path <local-vllm-model> --output_dir ../outputs/theft_baseline	Local vLLM/GPU/model setup required.
Bridge API	legal_simulation/utils/micro_macro_bridge.py::MicroMacroBridge	API-level utility only; no committed CLI.

Installation

Use a local Python environment and install the repository requirements:

pip install -r requirements.txt

The dependency file covers the release smoke path and non-vLLM repository components. Install and configure vLLM separately for macro runs that import vllm.

Environment configuration

For real micro runs, copy .env.example to an untracked .env file and fill local values there. The micro LLM configuration keys are:

• LAW_SIM_LLM_PROVIDER
• LAW_SIM_LLM_API_KEY
• LAW_SIM_LLM_BASE_URL
• LAW_SIM_LLM_MODEL
• LAW_SIM_LLM_MAX_TOKENS
• LAW_SIM_LLM_TIMEOUT
• LAW_SIM_GM_LLM_MODEL
• LAW_SIM_GM_LLM_TEMPERATURE

Pollution and log output locations can be controlled with:

• LAW_SIM_OUTPUT_DIR
• LAW_SIM_LOG_DIR

Macro model configuration can use:

• LAW_SIM_MACRO_MODEL_PATH

Macro output locations are set by command-line arguments such as --output_path for the generic macro runner and --output_dir for the theft baseline.

Some older code paths may still read compatibility fallback variables named in .env.example. Keep all secrets, private endpoints, and real model paths in local untracked environment files only.

Set LAW_SIM_LLM_MODE=mock only for no-network smoke tests and constructors. Do not set it for real experiments.

No-network smoke tests

Release smoke tests should run without real LLM, API, GPU, or vLLM access:

LAW_SIM_LLM_MODE=mock python -m pytest tests -q

This command validates the no-network release smoke path. It is not a real micro or macro experiment.

Real LLM / vLLM run caveats

Real micro runs use the repository LLM interface and require local API-compatible LLM configuration.

Real macro runs import vLLM directly inside the macro runner and require a local model, GPU access, and vLLM installed outside requirements.txt.

The macro path accepts macro-format scenes directly. Bridge outputs and bridge-related scene files are optional inputs for selected mapping experiments, not required infrastructure for every macro run.

Data and artifact policy

Keep generated outputs, logs, and local model or service configuration out of git. Use relative output locations such as outputs/, Results/, or legal_simulation/logs/.

Do not commit API keys, bearer tokens, passwords, private endpoints, real model paths, generated run artifacts, or machine-specific absolute paths.

Limitations

• This README documents code-backed entrypoints and release checks only.
• Full paper reproduction has not been claimed here as validated.
• Micro labor/company and micro pollution runs require real LLM setup.
• Macro runs require separate vLLM/GPU/local model setup.
• The bridge is an optional API utility and has no committed CLI.
• The repository does not currently include a verified license file.

Citation

If you find this repository useful, please cite our paper:

@inproceedings{wang-etal-2026-law,
    title = "Law in Silico: Simulating Legal Society with {LLM}-Based Agents",
    author = "Wang, Yiding  and
      Chen, Yuxuan  and
      Meng, Fanxu  and
      Chen, Xifan  and
      Yang, Xiaolei  and
      Zhang, Muhan",
    editor = "Liakata, Maria  and
      Moreira, Viviane P.  and
      Zhang, Jiajun  and
      Jurgens, David",
    booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {ACL} 2026",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.findings-acl.396/",
    doi = "10.18653/v1/2026.findings-acl.396",
    pages = "8061--8104",
    ISBN = "979-8-89176-395-1"
}