🔬 physics

Mapping the Narrow Corridor with Large Language Models

This paper proposes a reproducible, provider-agnostic pipeline using large language models to operationalize Acemoglu and Robinson's "Narrow Corridor" framework by quantitatively scoring state and society power for twelve countries over time, thereby transforming the theory from a qualitative narrative into a data-driven trajectory atlas.

Original authors: Ebrahim M Songhori

Published 2026-07-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ebrahim M Songhori

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to map the history of a country not as a list of dates and kings, but as a walk through a giant, invisible hallway. This hallway is the "Narrow Corridor," a concept from a famous book by economists who won a Nobel Prize. They argue that true freedom doesn't happen when a government is super strong or when the people are super strong; it only happens when the two are balanced, walking side-by-side down this narrow path. If the government gets too strong and the people get too weak, the country wanders into a "Despotic" zone (like a prison). If the government is too weak and the people are too strong, it falls into chaos. If both are weak, it's a "Paper" or "Absent" state where nothing really works. For decades, historians have tried to draw these paths, but it's incredibly hard work. It requires reading centuries of history and making thousands of judgment calls to decide exactly where a country is standing at any given moment. It's like trying to measure the temperature of a storm with a ruler; it's messy, subjective, and takes forever.

Now, enter the new kids on the block: Large Language Models (LLMs). These are the super-smart AI computers that can read almost everything ever written and answer questions about it. The big question this paper asks is simple: Can we teach these AI computers to be the historians? Can we ask an AI to look at a country's history, period by period, and draw its walking path through the Narrow Corridor? The authors, led by Ebrahim Songhori, decided to find out. They didn't just ask the AI to guess; they built a careful, step-by-step recipe. First, they made the AI list the major events of a decade (like a revolution or a new law). Then, they asked the AI to use that list to assign a score from 0 to 10 for how strong the government was and how strong the people were. They did this for twelve different countries, including giants like the US and China, and smaller ones like Somalia and Lebanon. They even tested four different AI models to see if they agreed with each other, and they compared the AI's drawings to a famous expert-made map called V-Dem to see if the AI was on the same page as human experts.

Here is what they found. The AI models were surprisingly good at drawing these paths. When they looked at countries like the United Kingdom and the United States, the AI correctly placed them inside the "Narrow Corridor," showing a long, steady walk where the government and the people grew strong together. When they looked at places like Iran and China, the AI showed them walking up into the "Despotic" zone, where the government became much stronger than the people, especially after major revolutions. For countries like the Democratic Republic of the Congo and Somalia, the AI correctly placed them in the "weak" zone, where the government barely exists. The most interesting part was that the AI didn't just copy the book's ideas; it actually calculated the path based on the events. For example, it showed Chile taking a sharp dive out of the corridor during a military coup in 1973 and then slowly walking back in after democracy returned.

However, the authors are very careful not to say the AI is perfect or that it has "solved" history. They found that while the AI agreed well with human experts on how strong the people were, it sometimes saw the government differently. The experts often thought governments were stronger than the AI did, especially in older times. The authors suggest this is because the AI is trained on a lot of modern, English-language data, which might make it see the past a bit differently than a human historian who has spent a lifetime studying specific archives. Also, because the AI was trained on data that likely includes the same expert maps it was compared against, the agreement might just be the AI remembering what it read, not necessarily discovering new truths.

In the end, this paper suggests that AI can be a powerful, cheap, and fast tool for drawing these historical maps. It can do in a few hours what might take a team of experts years to do, and it can do it in a way that is totally transparent—you can see exactly what events the AI used to make its decision. The authors propose using the AI as a "first draft" or a "baseline" to generate these paths for many countries quickly, and then letting human experts step in to check the tricky, controversial parts. It's not a replacement for human judgment, but it is a new, exciting way to turn the messy story of history into a clear, visual map that anyone can explore. The authors even released all their code and an interactive gallery so you can click and play through these animated history walks yourself, watching countries grow, stumble, and find their balance in the Narrow Corridor.

Technical Summary: Mapping the Narrow Corridor with Large Language Models

Problem Statement
Acemoglu and Robinson's The Narrow Corridor framework posits that liberty arises from a contested balance between state power and civil society power, visualized as a path through a two-dimensional space. While the framework is theoretically rich, it remains purely qualitative because operationalizing it requires labor-intensive, judgment-laden coding of centuries of history into specific $(society, state)$ coordinates. The authors identify this as a measurement problem: turning narrative history into a sequence of quantitative coordinates is vulnerable to human bias, selective reading, and lack of reproducibility. The paper asks whether Large Language Models (LLMs) can operationalize this framework to produce reproducible, checkable, and useful trajectories for multiple countries.

Methodology
The authors propose a provider-agnostic, reproducible pipeline that scores countries period-by-period (typically in 10-year intervals) using a specific prompting strategy:

Anchored Rubric: Instead of free-form requests, the system uses a fixed 0–10 rubric defining bands for state power (from "collapsed/nonexistent" to "overwhelming penetration") and society power (from "atomized" to "pervasively checks the state"). The rubric includes explicit neutrality guidance to avoid presentism and treat axes as independent.
Chain-of-Thought (CoT) Elicitation: For each period, the model is first prompted to list major historical events and trends affecting power dynamics, separating slow trends from discrete events. This narrative is then fed into a second prompt to generate the numeric score, ensuring the number is conditioned on explicit reasoning.
In-Context Temporal Anchoring: To ensure temporal coherence, the pipeline accumulates the $(society, state)$ pairs from all prior periods and feeds them into the prompt for the current period. This discourages the trajectory from drifting and allows the model to reason about changes relative to the previous state.
Structured Output: Responses are validated against a fixed JSON schema containing the scores, changes, key events, and justifications. Malformed responses trigger a retry or a numeric fallback.
Provider Agnosticism: A unified interface allows swapping model providers (Google, Anthropic, OpenAI, Alibaba) by changing a single string, isolating model effects from prompt effects.

Experimental Setup

Scope: A proof-of-concept atlas covering 12 countries (Iran, France, UK, US, China, Chile, Colombia, DRC, Lebanon, Zambia, Somalia, India) selected to represent the book's four "Leviathan" types (Despotic, Shackled, Paper, Absent).
Models: Four models were compared: three closed-weight (gemini-2.5-pro, claude-opus-4-8, gpt-5.5) and one open-weight (qwen-2.5-72b-instruct).
Validation:
- Consistency: Scores were compared against the Varieties of Democracy (V-Dem) expert index (specifically the civil society index and state authority over territory). The authors explicitly note that because V-Dem is likely in the models' training data, this measures "concurrent consistency" with expert consensus rather than external accuracy.
- Reliability: Inter-model agreement was measured using Krippendorff's $\alpha$ on period-to-period changes.
- Language Sensitivity: Prompts for five countries were translated into their native languages to test if the English-centric training data skewed results.

Key Results

Trajectory Atlas: The models successfully generated trajectories that largely aligned with the book's theoretical expectations.
- Shackled Leviathans: The UK, US, France, and India were placed inside the corridor, showing co-growth of state and society.
- Despotic Leviathans: Iran and China were placed in the state-dominant region, with trajectories showing a shift toward state dominance following major revolutions (1979 Iran, 1949 China).
- Weak States: The DRC and Somalia remained in the low-state region (Absent Leviathan), while Lebanon and Colombia occupied the weak-state region (Paper Leviathan), though the models placed Zambia inside the corridor rather than below it, interpreting its post-1991 stability as genuine balanced growth.
Ensemble Robustness: An ensemble mean of the four models produced a smoother, more robust reading than any single model. While models disagreed on the degree of growth (e.g., US state capacity), they largely agreed on the category of placement (e.g., all placed the US inside the corridor).
Consistency with V-Dem:
- Society Power: Strong positive correlations were found between LLM society scores and V-Dem's civil society index (e.g., $\rho = 0.89$ for China).
- State Power: Correlations were weak or non-existent. The authors attribute this to a mismatch in operationalization: V-Dem's "authority over territory" proxy is near-saturated for consolidated states, whereas the LLM captures broader state-capacity dynamics and historical building phases.
Reliability: Inter-model reliability (Krippendorff's $\alpha$ ) was generally moderate to strong (overall $\alpha \approx 0.66$ for society, $0.68$ for state), with the highest agreement on countries with clear historical arcs (Iran, Chile) and lower agreement on complex cases like the US.
Language Sensitivity: Translating prompts into native languages resulted in minimal displacement of trajectories (mean shift of 0.33 points), suggesting that while training data bias exists, the prompt language itself is not the primary driver of the scores.

Significance and Claims
The paper positions itself as a proof of concept rather than a definitive historical analysis. Its primary contributions are:

Operationalization: It demonstrates that LLMs can translate a qualitative, narrative-heavy framework into a reproducible, quantitative trajectory atlas.
Scalability and Cost: The authors highlight the dramatic cost and time reduction compared to expert panels (approx. $15 and a few hours for 1,600 calls vs. orders of magnitude more for human coding).
Transparency: Unlike human expert panels where bias is often opaque, the LLM pipeline offers explicit artifacts (prompts, rubrics, chain-of-thought) that can be inspected, versioned, and re-run.
Role Definition: The authors argue LLMs are best suited as a "scalable first pass" to generate candidate trajectories and surface areas of disagreement, freeing human experts to focus on contested cases where accountability and nuanced judgment are paramount.

Limitations Acknowledged
The authors explicitly caution that:

Circularity: Agreement with V-Dem does not prove accuracy, as V-Dem is likely in the training data.
Bias: The models inherit biases from English-centric, Western-leaning training corpora, which may flatten contested or under-documented episodes.
Path Dependence: The in-context anchoring method means an early mis-score could propagate through the entire trajectory.
Generalizability: With only 12 countries, the study is not a basis for cross-national generalization.

The paper concludes that while LLMs cannot replace expert coding for high-stakes accountability, they provide a powerful, reproducible tool for visualizing and exploring theoretical frameworks like the Narrow Corridor at a scale previously unattainable.

Technical Summary: Mapping the Narrow Corridor with Large Language Models

More like this