The Narrow Corridor — Trajectory Atlas

State vs. society power paths, scored by LLMs. 72 panels · 12 countries.

Method & code: GitHub repository. Click any plot to play its animation.

Iran (Persia)

Iran (Persia) — anthropic/claude-opus-4-8 — anthropic/claude-opus-4-8

Iran (Persia) — openrouter/qwen/qwen-2.5-72b-instruct — openrouter/qwen/qwen-2.5-72b-instruct

Iran (Persia) — ensemble-mean (4 models) — ensemble-mean (4 models)

Iran (Persia) — gemini/gemini-2.5-pro — gemini/gemini-2.5-pro

Iran (Persia) — openai/gpt-5.5 — openai/gpt-5.5

Iran (Persia) — V-Dem — V-Dem

France

France — anthropic/claude-opus-4-8 — anthropic/claude-opus-4-8

France — openrouter/qwen/qwen-2.5-72b-instruct — openrouter/qwen/qwen-2.5-72b-instruct

France — ensemble-mean (4 models) — ensemble-mean (4 models)

France — gemini/gemini-2.5-pro — gemini/gemini-2.5-pro

France — openai/gpt-5.5 — openai/gpt-5.5

France — V-Dem — V-Dem

United Kingdom

United Kingdom — anthropic/claude-opus-4-8 — anthropic/claude-opus-4-8

United Kingdom — openrouter/qwen/qwen-2.5-72b-instruct — openrouter/qwen/qwen-2.5-72b-instruct

United Kingdom — ensemble-mean (4 models) — ensemble-mean (4 models)

United Kingdom — gemini/gemini-2.5-pro — gemini/gemini-2.5-pro

United Kingdom — openai/gpt-5.5 — openai/gpt-5.5

United Kingdom — V-Dem — V-Dem

United States

United States — anthropic/claude-opus-4-8 — anthropic/claude-opus-4-8

United States — openrouter/qwen/qwen-2.5-72b-instruct — openrouter/qwen/qwen-2.5-72b-instruct

United States — ensemble-mean (4 models) — ensemble-mean (4 models)

United States — gemini/gemini-2.5-pro — gemini/gemini-2.5-pro

United States — openai/gpt-5.5 — openai/gpt-5.5

United States — V-Dem — V-Dem

China

China — anthropic/claude-opus-4-8 — anthropic/claude-opus-4-8

China — openrouter/qwen/qwen-2.5-72b-instruct — openrouter/qwen/qwen-2.5-72b-instruct

China — ensemble-mean (4 models) — ensemble-mean (4 models)

China — gemini/gemini-2.5-pro — gemini/gemini-2.5-pro

China — openai/gpt-5.5 — openai/gpt-5.5

China — V-Dem — V-Dem

Chile

Chile — anthropic/claude-opus-4-8 — anthropic/claude-opus-4-8

Chile — openrouter/qwen/qwen-2.5-72b-instruct — openrouter/qwen/qwen-2.5-72b-instruct

Chile — ensemble-mean (4 models) — ensemble-mean (4 models)

Chile — gemini/gemini-2.5-pro — gemini/gemini-2.5-pro

Chile — openai/gpt-5.5 — openai/gpt-5.5

Chile — V-Dem — V-Dem

Colombia

Colombia — anthropic/claude-opus-4-8 — anthropic/claude-opus-4-8

Colombia — openrouter/qwen/qwen-2.5-72b-instruct — openrouter/qwen/qwen-2.5-72b-instruct

Colombia — ensemble-mean (4 models) — ensemble-mean (4 models)

Colombia — gemini/gemini-2.5-pro — gemini/gemini-2.5-pro

Colombia — openai/gpt-5.5 — openai/gpt-5.5

Colombia — V-Dem — V-Dem

Democratic Republic of the Congo

Democratic Republic of the Congo — anthropic/claude-opus-4-8 — anthropic/claude-opus-4-8

Democratic Republic of the Congo — openrouter/qwen/qwen-2.5-72b-instruct — openrouter/qwen/qwen-2.5-72b-instruct

Democratic Republic of the Congo — ensemble-mean (4 models) — ensemble-mean (4 models)

Democratic Republic of the Congo — gemini/gemini-2.5-pro — gemini/gemini-2.5-pro

Democratic Republic of the Congo — openai/gpt-5.5 — openai/gpt-5.5

Democratic Republic of the Congo — V-Dem — V-Dem

India

India — anthropic/claude-opus-4-8 — anthropic/claude-opus-4-8

India — openrouter/qwen/qwen-2.5-72b-instruct — openrouter/qwen/qwen-2.5-72b-instruct

India — ensemble-mean (4 models) — ensemble-mean (4 models)

India — gemini/gemini-2.5-pro — gemini/gemini-2.5-pro

India — openai/gpt-5.5 — openai/gpt-5.5

India — V-Dem — V-Dem

Lebanon

Lebanon — anthropic/claude-opus-4-8 — anthropic/claude-opus-4-8

Lebanon — openrouter/qwen/qwen-2.5-72b-instruct — openrouter/qwen/qwen-2.5-72b-instruct

Lebanon — ensemble-mean (4 models) — ensemble-mean (4 models)

Lebanon — gemini/gemini-2.5-pro — gemini/gemini-2.5-pro

Lebanon — openai/gpt-5.5 — openai/gpt-5.5

Lebanon — V-Dem — V-Dem

Somalia

Somalia — anthropic/claude-opus-4-8 — anthropic/claude-opus-4-8

Somalia — openrouter/qwen/qwen-2.5-72b-instruct — openrouter/qwen/qwen-2.5-72b-instruct

Somalia — ensemble-mean (4 models) — ensemble-mean (4 models)

Somalia — gemini/gemini-2.5-pro — gemini/gemini-2.5-pro

Somalia — openai/gpt-5.5 — openai/gpt-5.5

Somalia — V-Dem — V-Dem

Zambia (Northern Rhodesia)

Zambia (Northern Rhodesia) — anthropic/claude-opus-4-8 — anthropic/claude-opus-4-8

Zambia (Northern Rhodesia) — openrouter/qwen/qwen-2.5-72b-instruct — openrouter/qwen/qwen-2.5-72b-instruct

Zambia (Northern Rhodesia) — ensemble-mean (4 models) — ensemble-mean (4 models)

Zambia (Northern Rhodesia) — gemini/gemini-2.5-pro — gemini/gemini-2.5-pro

Zambia (Northern Rhodesia) — openai/gpt-5.5 — openai/gpt-5.5

Zambia (Northern Rhodesia) — V-Dem — V-Dem

Prompt-language ablation

Does the prompt's language bias the scores? Each country's ensemble trajectory scored with English prompts (blue) vs. prompts translated into its own language (red); gray connectors show each period's shift. The paths stay close, so the trajectories are driven more by the history the models encode than by the prompt language.

Iran (Persia) — English vs Persian prompts — Iran (Persia) — English vs Persian

France — English vs French prompts — France — English vs French

China — English vs Chinese prompts — China — English vs Chinese

Chile — English vs Spanish prompts — Chile — English vs Spanish

Lebanon — English vs Arabic prompts — Lebanon — English vs Arabic