world-model-synthesis

[!IMPORTANT] AI Context & Knowledge Heritage

• Subsystem: Agent Skills Registry / world-model-synthesis
• Architecture: @docs ARCHITECTURE:Documentation
• Failure Path: Information drift, legacy terminology, or documentation mismatch.
• Observability: Traceability via execution/parity_guard.py ([SKILL])

World Model Synthesis & Graph Search Planning Skill

Knowledge Heritage: Inspired by Schema Harness (ARC-AGI-3 ~99% Public benchmark).
Source of Truth: execution/graph_planner.py, execution/backtest_engine.py

Overview

When agents encounter complex multi-step environment state changes (e.g. database schema migrations, deployment rollouts, infrastructure topology transitions), agents MUST synthesize a programmatic world model rather than guessing step-by-step actions in natural language.

---

Operating Protocol

Step 1: Synthesize World Model Script (.tmp/world_model.py)

Write a lightweight Python module representing the system state and transition rules:

def get_initial_state():
    return {"step": 0, "status": "PENDING"}

def is_target_state(state):
    return state.get("status") == "COMPLETED"

def get_successors(state):
    # Returns list of (action_name, next_state)
    actions = []
    if state["step"] == 0:
        actions.append(("VALIDATE_SCHEMA", {"step": 1, "status": "VALIDATED"}))
    elif state["step"] == 1:
        actions.append(("APPLY_MIGRATION", {"step": 2, "status": "COMPLETED"}))
    return actions

Step 2: Backtest against Recorded Telemetry

Run python execution/backtest_engine.py to backtest the world model against historical transition logs:

python execution/backtest_engine.py --trace-file .tmp/history.json

Step 3: Run Graph Search Planner

Execute execution/graph_planner.py to derive the optimal zero-token action sequence:

python execution/graph_planner.py --world-model .tmp/world_model.py

Step 4: Execute & Monitor

Execute the returned action path deterministically. If an unexpected state occurs, trigger the Discriminative Probing workflow to falsify competing hypotheses before modifying the world model.