feat: Epiplexity Demo 1 - Game of Life Emergent Objects

demos/epiplexity-01-emergent-objects/README.md

@@ -0,0 +1,80 @@
+# Demo 1: Emergent Discovery — LLM as Scientific Observer
+
+An LLM observes Conway's Game of Life and acts as a scientist — describing patterns, predicting future states, and building theories about hidden rules. Memory transforms a confused observer into a competent one.
+
+Validates **Paradox 1** from Finzi et al. (2026) *"From Entropy to Epiplexity"*: deterministic micro-rules expose extractable macro-structure to a computationally bounded observer with memory.
+
+## Research Link
+
+- Paper: *From Entropy to Epiplexity: Rethinking Information for Computationally Bounded Intelligence* (Finzi, Qiu, Jiang, Izmailov, Kolter, Wilson, 2026)
+- arXiv: https://arxiv.org/abs/2601.03220
+
+## Prerequisites
+
+- **Ollama** running locally: https://ollama.ai
+- **Qwen model** pulled:
+  ```bash
+  ollama pull qwen3.5:9b
+  ```
+- Verify Ollama is serving:
+  ```bash
+  curl http://localhost:11434/api/tags
+  ```
+
+## What the Model Does
+
+- Runs Conway's Game of Life on a 50x50 grid with three dramatic seeds:
+  - **R-pentomino** (center) — 5-cell methuselah that explodes into chaos, producing gliders, blinkers, blocks as emergent byproducts
+  - **Gosper glider gun** (top-left) — continuously emits gliders every 30 ticks
+  - **Pulsar** (bottom-right) — period-3 oscillator with dramatic rhythmic pulsing
+- A single LLM observer scans the grid through a 5x5 local window
+- Each tick, the observer:
+  1. **Describes** what it sees in free text (`llm:chat-with-template`)
+  2. **Predicts** the next 5x5 grid state (`llm:chat-with-template`)
+  3. **Reflects** periodically to update its theory of the world
+
+## Memory Modes
+
+- **Bounded**: Clears LLM history every tick. The observer has amnesia — perpetually confused.
+- **Persistent**: Accumulates history across ticks. The observer builds context, recognizes patterns, develops theories.
+
+Switch between modes live using the `memory-mode` chooser.
+
+## Run Instructions
+
+1. Start Ollama: `ollama serve`
+2. Open `game_of_life.nlogox` in NetLogo 7
+3. Click **Setup** to initialize the grid and observer
+4. Click **Test LLM** to verify Ollama connectivity
+5. Start in **bounded** mode, click **Go**
+6. After ~10 ticks, switch to **persistent** mode
+7. Watch the LLM Theory monitor evolve and accuracy diverge
+
+## What to Expect
+
+- **Bounded mode**: LLM descriptions are generic, predictions are random, theory never develops
+- **Persistent mode**: LLM starts recognizing patterns — "stable 2x2 blocks", "oscillating lines", "shapes that move diagonally"
+- **The Plot**: Bounded accuracy (red) stays flat. Persistent accuracy (blue) climbs. The gap = the epiplexity signal.
+- **The Killer Moment**: The LLM Theory monitor shows the observer independently discovering GoL patterns through pure observation with memory.
+
+## Files
+
+- `game_of_life.nlogox` — NetLogo model
+- `config.txt` — Ollama provider/model settings
+- `templates/describe.yaml` — prompt for pattern description
+- `templates/predict_grid.yaml` — prompt for next-state prediction
+- `templates/reflect.yaml` — prompt for theory generation
+- `results/` — CSV output directory
+
+## CSV Schema
+
+`tick, observer_x, observer_y, window_pattern, llm_description, llm_predicted_grid, ground_truth_label, grid_match_percent, memory_mode, llm_provider, llm_model`
+
+## Controls
+
+| Control | Type | Purpose |
+|---------|------|---------|
+| memory-mode | chooser | Switch bounded/persistent live |
+| show-observations? | switch | See raw LLM input in output area |
+| reflect-every | slider (3-10) | How often the LLM updates its theory |
+| episode-length | slider (20-200) | Total ticks per run |