Finding Your Entry Point | Training | Organization of Power

An entry point is the place in the literature where your argument makes a difference. The AI can propose where those places are. Whether one of them is real, and whether it is yours, it cannot decide.

Section 1Two kinds of gap

Calling something a “gap” is easy and usually unconvincing. A gap is persuasive when it is built from the literature itself: when you can show which works say what, and where what they say stops. Two ways of building one are worth naming.

(a) An untested scope condition or factor

The literature agrees on an argument, but it has been tested only in some cases, periods, or levels — or a factor it implies has never been examined. You extend the argument, or show where it stops holding.

Locke and Golden-Biddle (1997) describe this as building on a literature that is coherent but incomplete: the works are shown to agree, and the contribution fills in what they leave out.

(b) An unresolved disagreement

Works make rival claims, or rest on assumptions that conflict, and the evidence has not decided between them. You design a test that separates the predictions, or question the assumption both sides share.

Alvesson and Sandberg (2011) call the stronger version problematization: rather than spotting a gap, the researcher challenges an assumption underlying the literature and builds the question from it.

Both kinds come straight from Chapter 3. Kind (a) comes from the scope column of the ledger: where each argument has and has not been applied. Kind (b) comes from Table 3.2 and from the families that never meet.

Section 2The prompt

Prompt 4 Candidate entry points

Using the arguments ledger and typology, list 3–5 candidate "entry points", each supported by specific articles. Tag each as (a) an untested scope condition or factor, or (b) an unresolved disagreement.

Phrase entry points as options a researcher could choose, not as a finished argument.

For each, say which families and articles it engages, [optional] what [your dataset] can and cannot do to address it, and what a researcher would need to check outside this corpus before treating it as a real gap.

Two lines do the work. “Options, not a finished argument” stops the model from writing your paper's introduction for you. “Check outside this corpus” reminds both of you that a gap in the corpus is not yet a gap in the literature. The optional dataset line ties each option to the dataset built in Generating Data I: Extracting Datasets from a Large Media Corpus.

Section 3Five options

The agent proposed five options: three of kind (a) and two of kind (b). Each names the works it rests on and says what the on-site visits dataset could and could not do.

Table 4.1 Candidate entry points

Option	Rests on	The dataset can	It cannot
E1 (a). Do visits go to priority sectors, poor performers, or showcases?	B.-Y. Kim (sector shares to 2009); Oi and Shirk (models; stalled projects)	Tabulate visits by sector and province, 2012–2025, against Kim's shares; check whether production reports come before or after a visit	Observe inputs, or verify output claims
E2 (b). Pull or push: do enterprises attract attention, or does the leader pick them?	Pull: Heilmann, Shirk, Oi. Push: B.-Y. Kim, Park Jeong-jin	Code whether visited enterprises were earlier or later named national models or emulation targets (따라배우기)	See lobbying, which is not reported
E3 (a). Sanctions as a scope condition: does attention shift to elite areas?	Y. Park (lights in elite areas); periphery works on 2017–2019	A province-by-year panel against a sanctions index, with Pyongyang, provincial-capital, and border indicators	Rule out under-reporting of sensitive sites
E4 (a). Central or local industry after the 2024 20×10 policy?	Walder, Naughton, Painter; S.C. Kim; Park Jeong-jin	Code central versus local (지방공업) enterprises; compare shares before and after 2024	Rely on ownership being stated; the post-2024 window is short
E5 (b). Is a visit monitoring or reward?	Monitoring: Edin, Ericson. Reward or propaganda: B.-Y. Kim, Oi	Code tone (criticism versus praise or gifts); test whether criticised enterprises are revisited	Escape under-reported criticism

E2 and E5 are kind (b) because the families make rival predictions about the same visits — though, as Chapter 3 noted, no core work argues the point directly. Building the disagreement is part of the contribution.

A caveat for all five

KCNA and Rodong Sinmun publicise selectively. Some visits are never reported; secret, military, or failed sites are left out; reports are timed and framed for propaganda, as B.-Y. Kim says explicitly. The dataset measures publicised attention. Without outside data it cannot show where resources went, and it cannot show that output gains reported after a visit are real.

This is not a reason to drop the options. It changes what each can claim: an argument about how the regime presents its attention is open to the dataset; an argument about where attention actually goes needs more.

Section 4Testing an option

An option from a corpus is a hypothesis about the literature. Before treating it as a gap, test it:

Search outside the corpus. E4 depends on a 2024 policy; work published since the corpus was built may already address it. Run the option as a new search — the same way the corpus was built in Part 1.
Read the works it rests on. E2 depends on Heilmann's patron-seeking and Shirk's account of projects stalled until a leader's visit. Read those passages, not the ledger's summary of them.
Check the kind. A kind (b) option needs two sides. If one side turns out to be a contrast the AI promoted to a debate (Chapter 3, Section 6), it is a kind (a) option, or no option.
Check the data. Each row of Table 4.1 says what the dataset can do. Confirm that the codes it needs — sector, province, ownership level, tone — exist or can be added.

Section 5Choosing is yours

The entry points are among the least reliable parts of the AI's work. That is not because they are bad. It is because the choice depends on things the corpus does not contain: what you know, what data you hold, which debates you want to join, and what has been published since the corpus was built.

Treat the options as a map of where the literature could be entered, and make the choice yourself. Then write down why, in a sentence that names the families and works you engage. That sentence is the bridge to Part 3.

Example An entry-point sentence

“Accounts of central attention in reform-era China explain who is favoured by bargaining from below; the North Korean literature explains it by the leader's discretion, but has counted visits only to 2009. Coding which enterprises were named models before and after a visit, 2012–2025, separates the two predictions for publicised attention.”

Illustrative only. It combines E1 and E2, and it would still need the tests in Section 4.

Section 6What carries into Part 3

Part 2 ends with files, not prose. Part 3, From Arguments to Prose, starts from them:

corpus survey

The map of the field: what it asks, when, and where. Becomes the opening paragraphs of a review.

rq analysis

Core and periphery, with the four analyses and verified quotes. Decides what is cited and how closely.

arguments ledger

One row per core argument. Becomes the body of the review, family by family.

typology

The families and disagreements. Becomes the review's structure.

entry point

Your chosen option and the sentence that justifies it. Becomes the review's last paragraph and the paper's first.

process notes

What was checked, what was not, and what was corrected. Stays with the files.

Section 7AI: opportunities and challenges

Step 4 Finding an entry point

Opportunities

Seeing the whole structure. The AI can hold the ledger, the families, and the disagreements together and see where they do not meet.

Several options at once. Five traceable options give you something to choose among, rather than the first idea that occurs to you.

Data in view. Asked what the dataset can and cannot do, the AI ties each option to a design — and to its limits.

Challenges

Corpus-bound. The AI sees gaps in the corpus, not in the literature. Every option needs an outside search.

Steerable. Name a dataset in the prompt and the model will find reasons it fits.

Not your judgment. Feasibility, data, and what you want to argue are not in the files.

Terminology introduced in this chapter

entry point: The place in the literature where a new argument makes a difference.
scope condition: The cases, periods, or levels to which an argument is claimed to apply. An untested scope condition is one where the argument has not yet been applied.
unresolved disagreement: Rival claims, or conflicting assumptions, that the evidence has not decided between.
problematization: Building a research question by challenging an assumption the literature shares, rather than only filling a gap in it (Alvesson and Sandberg 2011).
entry-point sentence: One sentence naming the families and works your argument engages, and why your case or data decides between them.