The Ontology of the Alien: Escaping the Median Trap in LLM Ideation

Large Language Models asked to "be creative" produce solutions that converge on a small number of archetypes — the Median Trap. This repository contains the experiment code and full dataset (196 solutions across 8 conditions) for a systematic comparison of methods that escape it.

We test three novel architectures against baselines:

• Semantic Tabu — accumulating constraints that block previously used mechanisms
• Solution Taxonomy (Studio Model) — a dual-agent system where an Explorer proposes and a Taxonomist curates an evolving ontology graph
• Orthogonal Insight Protocol — constructing coherent alternative physics, solving the problem within them, and extracting mechanisms back to reality

Prior work: This paper extends Algorithmic Creativity via Strange Worlds (Westerberg, 2025), which introduced the Orthogonal Insight Protocol and tested it against a single baseline.

The Problem

All eight conditions tackle the same hard problem:

"How do we build a retirement system for people who don't know how much they will earn next month, where 'consistency' is impossible?"

Conditions

Condition	Inspiration	Novelty Mechanism
A: Semantic Tabu	None	Tabu list
B: Solution Taxonomy	None	Graph
C: Random Seed	Seed word	None
D: Seed + Tabu	Seed word	Tabu list
E: Seed + Taxonomy	Seed word	Graph
F: Orthogonal	Alien physics	None
G: Orthogonal + Tabu	Alien physics	Tabu list
H: Orthogonal + Taxonomy	Alien physics	Graph

25 runs per condition. Conditions B and H had 23 solutions accepted into their taxonomy graphs (2 rejected each as structurally redundant), yielding 196 distinct solutions.

Key Findings

• The Studio Model (Conditions B, E, H) exhibited emergent metacognition: active commissioning of research, structural coaching, and ontological accommodation (restructuring categories when data defied classification).
• The system independently derived advanced economic concepts including antifragility, metric dissolution, and flow rights as alternatives to accumulation.
• Different architectures produce different solution space topologies: Tabu forces vertical depth, Seeds create lateral branching, and Orthogonal Insight extracts epistemological stances.

Repository Structure

paper/                  LaTeX source and compiled PDF
src/taxonomy_graph/     Graph data structure and embedding service
agents/                 Agent prompts (explorer.md, taxonomist.md) and orchestration
run_experiment.py       Main entry point
analysis/               Result analysis scripts
seeds.json              25 seed words used for Conditions C-H
schema.json             Solution output schema

Data directories (25 JSON files each):

Directory	Condition
semantic_tabu/	A
taxonomy/	B
random_seed/	C
seed_tabu/	D
taxonomy_seed/	E
strange_worlds/	F
strange_worlds_tabu/	G
taxonomy_worlds/	H

Each JSON file contains the full agent output (world-building text, solver reasoning, extracted solution) for reproducibility.

Usage

Requires Claude Opus 4.5 and a valid Anthropic API key.

pip install -r requirements.txt
python run_experiment.py --condition [A-H]

Try It

The Orthogonal Insight Engine is available as a standalone open-source tool: emergent-wisdom/orthogonal-insight-engine

License

MIT License