Asking a Research Question

The survey told you what the field asks. A supervised query tells you which parts of the field have asked your question, which have asked something next to it, and — just as useful — what none of them has measured.

Section 1From a field to a question

A literature review is organized around a question, not a topic. Two researchers with the same corpus and different questions should write different reviews, citing different works at the centre.

The query sorts the corpus into three groups:

• Core works ask a question like yours. These are the arguments your review must engage one by one.
• Periphery works address your question only in part or indirectly: they supply a mechanism, a context, a method, or a case. They are cited in support but not argued with.
• Everything else. Most of the corpus. For this question, it is background.

The line between core and periphery is a judgment, and the AI must be made to show where it drew it. The point of drawing it is the analysis that follows: once the two groups exist, you can compare how each asks the question, what each argues, and what each stands on.

Section 2The research question

The worked example uses one question from the lab's research on the organization of power:

The question is chosen to meet a dataset. The lab's companion tutorial, Generating Data I: Extracting Datasets from a Large Media Corpus, builds a dataset from 946 KCNA and Rodong Sinmun reports on North Korean enterprises (기업소), 2012–2025. Each report is coded for sector, event type, province, and leader_visit: whether the leader visited. In that dataset a visit (현지지도, on-the-spot guidance) is the observable measure of central attention.

A literature review for that dataset has two jobs. It must say what the literature already argues about why the centre attends to some places and not others, and it must say what a dataset of visits could add. This chapter does the first; Chapter 4 does the second.

Section 3The prompt

You are running a literature analysis for one research question. Accuracy matters more than polish: every article you name and every quote must come from the files. Never invent works or quotes.

Files (read-only): 3 PDFs/citation_index.csv (434 works) and 3 PDFs/text/<item_key>.txt (full text for 309). Use Python to search the texts.

Research question: [your question here]

Identify CORE articles (primary research question similar to this one) and PERIPHERY articles (address it only in part or indirectly). Screen all 434 by title and abstract, with keyword search over the texts for each channel the question names, then read the candidates' full text (introduction, conclusion, key sections). Aim for roughly 10–20 core and 15–30 periphery.

Summarise each in 2–4 sentences. Then, for CORE and PERIPHERY separately, analyse: (1) variations of the research question; (2) main arguments; (3) data; (4) methods.

[Optional] List every work that discusses [your observable measure] substantively, separately from works that only mention it.

Quote one short passage (40 words or fewer) for each CORE article, with the text file as the source. Mark abstract-only summaries.

Write rq_analysis.md, and process_notes.md: your steps, how screening narrowed 434 to core and periphery (counts at each step), borderline core/periphery calls with keys, errors you caught yourself, and what a human must verify.

Section 4Screening: 434 to 47

The first finding came before any summary was written. The question has four channels, and the corpus is very uneven across them: more than a hundred works on ownership and credit, six on campaigns. A question that leans on its thinnest channel will have a thin core.

Section 5The core and the periphery

Each core summary came with a quote. Two show how a quote anchors a summary — and how two works from different countries can describe the same moment:

The 30 periphery works fall into a few groups:

• North Korean visits and sanctions (13). Works that report visit counts or use visits in passing — Kong 2014 on visits to military camps [7ZMILDVP], Noland 2019 on the 2018 shift to economic sites [QSL3BVBS], Kim, Kim, Park & Sun 2023 on the costs of sanctions [MIS3S8V6] — and works on reform politics and rents.
• Chinese fiscal and administrative relations (13). Revenue sharing and fiscal dependence (Wong 1991, Oi 1992, Wedeman 1999), federalism (Montinola, Qian & Weingast 1995), collusion against inspection teams (Zhou 2010), and the promotion of officials (Chen, Li & Zhou 2005; Shih et al. 2012).
• Other cases and theory (4). Russian governors (Reuter 2012), the breakdown of Soviet and Chinese hierarchies (Solnick 1996), Vietnamese state enterprises (Beresford 2008), and soft budgets in banking (Berglof & Roland 1998).

Sections 6 to 9 are the analysis. Each takes one of the prompt's four comparisons and sets the core beside the periphery.

Section 6Analysis 1: variations of the question

The core does not ask one question. It asks six, and they differ in where they look for the answer:

Only two core works ask about the leader's own instruments, and both are on North Korea. The comparative core asks the question through ownership, credit, and bargaining; the North Korean core asks it through the leader. That split shapes the rest of the analysis.

Section 7Analysis 2: main arguments

Grouped by what they say decides who gets attention, the core arguments fall into four clusters. Chapter 3 turns them into a typology; here they are in brief.

1. 1The level of ownership structures attention.Lower-level owners monitor more closely (Walder). The centre grasps the large and lets the small go (Naughton). Restructuring falls on local SOEs (Painter). North Korea's county industry is left to fend for itself (S.C. Kim).
2. 2Particularism from below.Enterprises and provinces bargain for deals (Shirk, Solinger, Kornai 1986). Rescue motives keep budgets soft (Kornai, Maskin & Roland). Model status brings favours, and experimenters seek patrons (Oi, Heilmann).
3. 3The regime targets its coalition.Resources shift to elite areas when sanctions bite (Y. Park). Provinces are created to win votes (Malesky). Strategic townships are watched more closely (Edin).
4. 4The leader chooses.Visits and the discretionary fund move inputs to favoured sites without improving performance (B.-Y. Kim). A party-led campaign mobilises labour for selected counties (Park Jeong-jin).

The periphery holds the corpus's only direct argument about what officials are rewarded for: Chen, Li & Zhou 2005 find that provincial leaders' turnover depends on relative growth, and Shih et al. 2012 find no evidence that growth was rewarded. Chapter 3 lists it among the disagreements.

Section 8Analysis 3: data

Section 9Analysis 4: methods

Put Sections 8 and 9 together and the pattern is clear. The core makes arguments about why particular units are favoured, and tests them with fieldwork in China; the periphery has more systematic data but asks other questions. For North Korea, no work in either group has data at the level of the enterprise or the province. That is the opening a visits dataset could fill, and Chapter 4 tests it.

Section 10What the corpus says about on-site visits

Because the question names one observable channel, the prompt asked for every work that discusses it, separately from the core/periphery split.

Only one work in the corpus counts visits systematically and relates them to an outcome. B.-Y. Kim 2017 tabulates 1,447 Kim Jong-il visits by sector, 1994–2009, and finds no correlation between the share of visits to the economy and economic performance. Its series ends in 2009.

For the dataset this is the most useful finding in the chapter. No work codes the province or the enterprise visited. None codes what follows a visit. None covers 2012–2025. The variable leader_visit in Generating Data I: Extracting Datasets from a Large Media Corpus is not a replication of anything in this corpus; it measures something the literature discusses and has not measured.

Section 11Checking the analysis

The process notes for this run are as useful as the analysis. Three parts of them should be read before anything else.

Errors the agent caught

1. Vocabulary. The first visit search missed “Spot Guidance” (B.-Y. Kim's term) and “onsite” (Noland's). A second pass with spelling variants fixed it. Every field has its own words for the same thing.
2. A second language. The English searches missed the one Korean-language core work. A separate pass on Korean terms — 현지지도, 현지료해, 기업소, 지방공업, 속도전 — found it.
3. False positives. “Inspection” matched nuclear inspections; “visit” matched diplomatic visits to Beijing. Reading the passages around each match separated them.
4. Invented detail. Four small details in draft summaries — a citation, a phrase, an appendix label, a source attribution — were not in the texts. They were removed because every claim was checked.

The first two are why a human who knows the field must supply the vocabulary. The last is the error to watch for in any AI analysis of literature: a model knows a great deal about famous works and will write what it knows rather than what the text says, unless the prompt forces it back to the text.

What a human must verify

• The visit counts in B.-Y. Kim 2017, against the book.
• A robustness table in Y. Park 2025 that was jumbled in extraction, before relying on its result.
• The translation of the Park Jeong-jin 2024 quote, by a Korean reader.
• Periphery summaries that rest on the opening and closing pages only.
• The 125 works with no text and 17 with empty or near-empty text: a visit study among them would have been missed.

Your spot-check

Pick five core works. For each, open the PDF, find the quote, and read a page either side. Then ask: is the one-line summary what the work argues, or what it is known for? Start with the works whose summaries rest on numbers read from tables.

Section 12AI: opportunities and challenges

Chapter 3 takes the seventeen core works and turns their arguments into a ledger and a typology.