Skip to content

Roadmap.md

Latest commit

 

History

History
71 lines (52 loc) · 2.07 KB

Roadmap.md

File metadata and controls

71 lines (52 loc) · 2.07 KB

LLM Social Simulation — Project Roadmap

Phase 0 — MVP: End-to-End Simulation Loop (Closed Loop + Tests)

Status: In progress (late stage)

Goal

Build a minimal but complete simulation framework that can run Iterated Prisoner’s Dilemma (IPD) on a fixed graph, produce analyzable outputs, and pass automated tests.

Phase 0 establishes a stable foundation or internal environment for all later phases.

Core Capabilities

  • Run multi-round IPD on a fixed graph
  • Output:
    • Cooperation rate per round (time series)
    • Final wealth distribution
  • Simulation-related tests pass (smoke tests)

Deliverables (Minimal Set)

SimulationEngine

  • Deterministic round-based orchestration
  • Pull observations from Gameworld
  • Collect agent actions and apply them each round
  • Append-only history of TickResult

Gameworld

Single source of truth for:

  • Agent graph (adjacency list)
  • Wealth accumulation
  • Last-round actions
  • Local observation generation per agent
  • Payoff computation using Prisoner’s Dilemma matrix

Agents

  • BaseAgent(agent_id)
  • decide(observation) -> Dict[neighbor_id, action]
  • At least two strategies implemented:
    • Always Cooperate
    • Always Defect
    • (Preferred) Tit-for-Tat

Simulation Entry Point (optional)

  • run_mvp.py
  • Single official entry point for running the Phase 0 MVP

Tests

  • test_gameworld_smoke.py
  • test_engine_dummy.py
  • All simulation tests pass under pytest

Acceptance Criteria

  • The following command runs successfully from the repo root: 
    python -m llm_social_simulation.simulation.run_mvp
  • Output includes: Cooperation rate over time and final wealth per agent
  • pytest passes

Phase 1 -- Small-World Networks & Config-Driven Experiments

Phase 2 -- Baselines, Norms, and Measurement Validity (Pre-LLM)

Purpose: Establish controls so LLM behavior can be isolated later.

Phase 3 -- LLM Agents, Communication, and Dynamic Networks

Purpose: Study how LLMs behave as social agents under constraints.