First Year

A marriage simulator that demonstrates attractor-based narrative generation — a multi-agent architecture that breaks LLM text adventures out of autoregressive local minima.

The Problem

Single-prompt LLM games suffer from context poisoning: the model conditions on its own outputs, converging to whatever tone was established early. A passive player gets a pleasant, flat experience with no surprise. The narrative falls into a local minimum.

The Solution

This game uses three independent LLM agents with different objectives:

• Narrative Model — writes the scene the player sees
• Director — a separate LLM that reads the full context and generates pressure directives (confrontations, disruptions, revelations)
• Partner Agent — a background scorer that tracks the partner's emotional state (drift, tension) and maintains attractors: unresolved relationship patterns that accumulate charge when ignored

When an attractor's charge reaches threshold, it mandates — the narrative model is forced to cash it out as a concrete in-scene event. This breaks the autoregressive feedback loop because the pressure comes from outside the conversation context.

Setup

Requires Claude Code installed and authenticated.

git clone https://github.com/NicholasARossi/first-year.git
cd first-year
python play.py

Options

python play.py                    # full game with attractor system
python play.py --no-attractors    # baseline mode (no charge dynamics)
python play.py --no-director      # disable the director
python play.py --hide-thinking    # hide model reasoning
python play.py --load session.json  # resume a saved session

A/B Backtest

Compare baseline vs attractor system with passive-player rollouts:

# First: play a few turns, then copy a save as the fixture
cp saves/session_XXXXXXXX_XXXXXX.json backtest_fixtures/session_frozen.json

# Run the backtest
python backtest.py --rollouts 10 --turns 5

Results are saved to backtest_results/.

Architecture

Player Input
    │
    ▼
┌──────────┐     ┌──────────┐
│ Director  │────▶│Narrative │────▶ Player sees this
│ (Opus)   │     │ Model    │
└──────────┘     │ (Opus)   │
                 └──────────┘
                      │
                      ▼
              ┌──────────────┐
              │Partner Agent │
              │  (Opus)      │
              │              │
              │ drift: 4/10  │
              │ tension: 6/10│
              │ attractors:  │
              │  - resentment│
              │    charge: 8 │
              │    MANDATED  │
              └──────────────┘

The Director and Partner Agent run as separate LLM calls. The Partner Agent runs in a background thread while the player reads and types — zero added latency.

Blog Post

This repo accompanies a blog post on attractor-based narrative generation for LLM games. See rossidata.com.