How fudandisc/maser works
MASER generates legal-interaction dialogues by configuring a case scenario, creating plaintiff and lawyer agents, and running their exchanges under supervisory checks; the resulting histories support synthetic legal-scene data. Model engines provide interchangeable LLM backends. The README also describes MILE’s interaction and goal evaluation phases. The graph reflects documented and excerpted relationships; some runtime wiring is dynamic or unsampled, so those links are omitted or marked unverified. Training is represented only as a documented capability, not as a delivery pipeline.
Main entry points: law_fact.py Case Scenario · supervisor.py Dialogue Supervisor · lawyer_fact.py Lawyer Agent · eval.py Interaction Evaluation
Components 24
Legal Simulation
- Run Entry
src/run_law_fact.py - Scenario Configuration
src/utils/options_law_fact.py - Case Scenario
src/legal/law_fact.py - Class Registry
src/utils/register.py
Interaction Agents
- Agent Memory
src/agents/base_agent.py - Plaintiff Agent
src/agents/plaintiff_fact.py - Lawyer Agent
src/agents/lawyer_fact.py - Dialogue Supervisor
src/agents/supervisor.py - Dialogue Guidance
src/utils/guider.py - Speech Styling
src/utils/styler.py - Fact Perturbation
src/utils/tamperer.py
Model Engines
- Engine Interface
src/engine/base_engine.py - LLM Backends
src/engine
MILE Evaluation
- Interaction Evaluation
src/evaluation/src_eval/eval.py - Goal Evaluation
src/evaluation/src_eval/eval.py - Evaluation Prompts
src/evaluation/src_eval/prompts.py - Evaluation Utilities
src/evaluation/src_eval/utils.py - Evaluation Runner
src/evaluation/src_eval/main.py - MILE Evaluation Data
src/evaluation/data
Data and Training
- Case Profiles
src/data/legal.json - Dialog Histories
src/legal/law_fact.py - SynthLaw Dataset
src/data/legal.json - Fine-tuning Configs
src/train
Other components
- Researcher
Connections 32
- Researcher → Run Entry: runs
README.md - Run Entry → Scenario Configuration: loads options
src/run_law_fact.py - Run Entry → Case Scenario: starts scenario
src/run_law_fact.py - Scenario Configuration → Class Registry: resolves classes
src/utils/options_law_fact.py - Case Scenario → Case Profiles: loads cases
src/legal/law_fact.py - Case Scenario → Plaintiff Agent: queries
src/legal/law_fact.py - Case Scenario → Lawyer Agent: conducts interview
- Case Scenario → Dialogue Supervisor: requests review
src/legal/law_fact.py - Case Scenario → Dialog Histories: records dialogue
src/legal/law_fact.py - Plaintiff Agent → Agent Memory: inherits
src/agents/plaintiff_fact.py - Lawyer Agent → Agent Memory: inherits
src/agents/lawyer_fact.py - Dialogue Supervisor → Agent Memory: inherits
src/agents/supervisor.py - Dialogue Supervisor → Dialogue Guidance: uses guidance
src/agents/supervisor.py - Dialogue Supervisor → Speech Styling: applies style
src/agents/supervisor.py - Dialogue Supervisor → Fact Perturbation: perturbs facts
src/agents/supervisor.py - Lawyer Agent → Class Registry: resolves engine
src/agents/lawyer_fact.py - Dialogue Supervisor → Class Registry: resolves engine
src/agents/supervisor.py - Class Registry → LLM Backends: selects backend
src/engine/base_engine.py - LLM Backends → Engine Interface: implements
src/engine/LawLLM.py - Lawyer Agent → LLM Backends: requests responses
src/agents/lawyer_fact.py - Dialogue Supervisor → LLM Backends: requests responses
- Case Scenario → SynthLaw Dataset: generates dialogues
README.md - SynthLaw Dataset → Fine-tuning Configs: supports fine-tuning
README.md - Evaluation Runner → Interaction Evaluation: runs evaluation
src/evaluation/src_eval/main.py - Evaluation Runner → Goal Evaluation: runs evaluation
src/evaluation/src_eval/main.py - Interaction Evaluation → Evaluation Prompts: formats prompts
src/evaluation/src_eval/eval.py - Interaction Evaluation → Evaluation Utilities: requests scoring
src/evaluation/src_eval/eval.py - Goal Evaluation → Evaluation Prompts: formats prompts
src/evaluation/src_eval/eval.py - Goal Evaluation → Evaluation Utilities: requests scoring
src/evaluation/src_eval/eval.py - Evaluation Runner → MILE Evaluation Data: evaluates against
README.md - Dialog Histories → Interaction Evaluation: scores dialogues
README.md - Dialog Histories → Goal Evaluation: scores goals
README.md