Organizing the Arguments

A typology is only as good as the records it sorts. Before asking an AI to group arguments, make it write each argument down in the same form — and then check the form, not just the grouping.

Section 1A ledger, not notes

Reading notes are written in the order you read, in whatever shape each work suggests. They are good for remembering a work and poor for comparing works. An arguments ledger fixes the shape: one row per work, the same fields in every row.

The fields that matter most are the ones that force a commitment:

• Outcome. What is being explained? Two works that seem to disagree often explain different outcomes.
• Main argument, as one sentence: X causes or explains Y because Z. If the AI cannot fill this in, the work may not make an argument about your question.
• Scope: case, period, level. Many apparent disagreements turn out to be differences of scope.
• Data and method. What the argument stands on.

For this question the ledger has one more field: what the work says about on-site visits, tours, campaigns, or monitoring. A field for your observable measure keeps the dataset in view while you read.

A ledger is also a file, not a document. Saved as a spreadsheet, it can be sorted, filtered, corrected, and carried into Part 3, where it becomes the skeleton of the written review.

Section 2The prompt

The arguments ledger, for CORE articles only. For each record: item_key, short cite (Author Year), title, outcome / dependent concept, main argument (one sentence: X causes/explains Y because Z), mechanism, scope (case, period, level), data, method, key cited works it builds on or argues against (from its text), year. [Optional] what it says about [your observable measure].

Then place each argument on three levels: conceptual (the abstract claim), operational (how the concepts are measured), empirical (what evidence shows it).

Organize the arguments: propose a typology (2–4 families of explanation, e.g. by mechanism or by level of analysis), assign each core article to one family, and note borderline placements. List only DIRECT disagreements — where a text argues against a named position — with the passage; list contrasts and alternatives separately.

Write arguments.md and arguments.csv (same rows), and typology_and_entry_points.md.

Section 3One ledger entry

Here is the entry for B.-Y. Kim 2017, the one work in the corpus that counts on-site visits:

The one-sentence argument is the field to read first. It shows that Kim's claim is about who decides: the leader, not the plan and not the enterprise. It also shows what his evidence cannot do. Kim argues that inputs follow visits, but his data are the visits; the inputs are not observed. That gap matters in Chapter 4.

Section 4Three levels of an argument

An argument can agree with another at one level and differ at the next. The prompt asks for three:

1. 1ConceptualThe abstract claim. B.-Y. Kim: the leader allocates attention and resources at his own discretion, against the plan.
2. 2OperationalHow the concepts are measured. B.-Y. Kim: the number and sectoral share of visits recorded in official media.
3. 3EmpiricalWhat evidence shows it. B.-Y. Kim: visits cluster on the economy and the military, and their share does not track performance.

Shirk and Oi make a close conceptual claim — a leader's visit marks the moment a project or a village gets the centre's attention — but neither measures visits. Their evidence is fieldwork: in Shirk, projects that stalled until a leader visited; in Oi, model villages that won preferential treatment.

Laid out this way across all seventeen works, the levels show where the field is thin. Several core works describe visits or inspection tours; only B.-Y. Kim counts them. Elsewhere attention is measured indirectly — through ownership level, credit, transfers, or nighttime lights, which measure resources, not attention. Chapter 4 returns to this gap.

Section 5Four families of explanation

All four families answer the same question — why do some enterprises and localities receive more of the centre's attention? — in different ways.

Two things stand out. First, the families sort roughly by country: F2 is mostly China, and F4 is all North Korea. The comparative literature explains favour by structure and bargaining; the North Korean literature explains it by the leader. Second, the four families make different predictions about where visits should go — to large central firms (F1), to firms that ask (F2), to the regime's base (F3), or wherever the leader chooses (F4). A visits dataset can tell them apart.

Section 6Where the arguments disagree

A typology says which arguments belong together. Disagreements say where the debate is. The prompt asked only for direct disagreements — a text arguing against a named position — and the agent found five:

The third row bears most directly on the dataset. Kim notes that state media often reports substantial increases in production after the leader has given Spot Guidance instructions, and argues that the gains may come from inputs, not guidance. Any analysis of what follows a visit in KCNA reports has to take a side in this disagreement, or design around it.

Notice what is not in the table. The four families are rival explanations, but no core work argues against a core work in another family. Bottom-up accounts (F2) and top-down accounts (F3, F4) simply do not meet. That is not a debate; it is a gap, and Chapter 4 treats it as one.

Section 7Checking the typology

Four checks are worth doing on any AI typology:

1. Reassign without labels. Hide the family column, read five ledger rows, and assign them yourself. Where you disagree with the AI, the family definitions are probably unclear.
2. Look at the borderline works. Works that will not fit one family are often the most interesting in the review. Here, Park Jeong-jin sits where F1 and F4 meet: a party campaign that overrides the ownership hierarchy.
3. Check the “built on / argued against” field. The agent counted each cited name in the text, which confirms that a work is cited, not how. Do not treat it as a citation network.
4. Ask what is missing. The typology covers this corpus. A family that has no members here — say, explanations from the enterprise's own capacity — may exist in a literature the corpus did not reach.

Section 8AI: opportunities and challenges

Chapter 4 uses the ledger, the families, and the disagreements to find where a new argument could enter.