Methodology

By Gil Guerra · Updated September 2026 · Evidence tiers

The Murillo broadcast corpus

Rosario Murillo, now co-president, speaks on state media almost every weekday at noon, and regime outlets publish transcripts of these routine broadcasts. The corpus used on this site contains 2,405 of those transcripts, from April 2012 through September 4, 2026: about 4.8 million words in 100,328 paragraphs.

Aug. 2015effective start of coverage
12 monthswith no transcripts, Sept. 2019–Aug. 2020
3 outletsCanal 4, Radio Nicaragua, El 19 Digital
Checked by the authordictionary search plus AI-assisted reading; minor errors may remain

Figure 1 above shows how many transcripts the corpus holds for each year, by outlet.

How the measures are made

Counts come from dictionary search (accent-insensitive keyword and phrase patterns, listed in section 2), followed by AI-assisted reading of the matched sentences to classify how a topic is mentioned, for example Russia as a security partner versus Russia as a cultural or tourism partner. These counts were checked by the author, but minor errors may remain. They describe her routine daily broadcasts only, not speeches at special events, other officials’ statements or everything the state does. That is why findings on this site are worded “in her daily broadcasts” or “by this count,” and never as claims about what the regime as a whole has or has not said.

Learn more: coverage, outlets, voice and quotations

Coverage. Only one transcript survives from before August 2015, so effective coverage starts in August 2015. Since then, the only calendar months with no transcripts are the 12 from September 2019 through August 2020. Weekends and missing days are unknown, not zero: a count of “no mentions” never includes them.

Outlets. The source outlet changes over time: Canal 4 (2015–2023), Radio Nicaragua (2022–2023) and El 19 Digital (2023–2026). Canal 4 transcripts before 2022 are roughly twice as long as later ones, so shares of broadcasts that mention a term are not strictly comparable across the switch. Where two outlets published the same broadcast, only one copy is kept. Standard boilerplate (share prompts, sign-offs, headings) is masked before counting.

Voice. Murillo often reads government communiqués aloud. The corpus does not separate her own words from statements she reads, so figures describe what was said on her broadcast, not what she personally wrote.

Quotations. Every quotation from Murillo shown on this site is an exact match to the corpus text, carries the broadcast date and links to the original published page. English glosses are the author’s.

Topic dictionaries

Each topic on the Murillo’s Broadcasts page is defined by a list of search patterns and exclusions. A broadcast counts toward a topic if at least one pattern matches after exclusions are removed.

Figure 2 above shows how many broadcasts each dictionary matched and the precision found in spot checks.

Rules that change on a known date

Some words changed meaning on a known date, so their exclusions apply only in a window. The bar shows when each rule is in force across the corpus.

Exclusions drop look-alikes, such as streets, schools or people that share a country’s or leader’s name. The table shows how many broadcasts each dictionary matched. A precision estimate (the share of matches a reader judged to be genuinely about the topic) is shown only where such a check was made; otherwise the counts are unchecked dictionary matches.

Full dictionary table: every search pattern, exclusion and date rule

Loading topic definitions…

Topic-mix validation

The “what does she talk about” measure assigns every paragraph to one of eleven broad subjects (public works, faith, foreign affairs and so on) using a keyword classifier, then reports each subject’s share of her words per year. A blind second reading agreed with it on 68.5% of paragraphs.

200paragraphs read blind (20 per year)
68.5%agreement; Cohen’s kappa 0.62
84%matched the reader’s first or second choice

Figure 3 above shows which measured shifts in her topic mix are robust.

Known weaknesses

The classifier under-detects weather and disaster content in 2017–18, so the decline of the “news desk” is, if anything, larger than shown.
“Ritual greetings” absorbs slogan-heavy paragraphs; its level is overstated by several points and shows no reliable trend.
Foreign affairs is under-detected (recall about 40%), so its levels are low; the fall from 2024 to 2026 is robust, but comparisons with 2017 are not.
2017–2021 transcripts come from Canal 4 and 2024–2026 from El 19 Digital. Copies of the same broadcast from two outlets differ by at most about one point per subject.

Learn more: the blind check and what “robust” means

To test it, a second coder read 200 fresh paragraphs (20 per year) without seeing the classifier’s labels. The second coder was an independent AI-assisted reading, not a human panel. The two agreed on 68.5% of paragraphs (Cohen’s kappa 0.62, conventionally “substantial” agreement). In 84% of paragraphs the classifier’s label matched either the blind reader’s first or second choice, so most disagreements involve paragraphs that mix two subjects. An earlier, non-blind check by the classifier’s builder had reported 83%; the site uses the lower, blind figure.

What that means for the charts. Treat the topic shares as evidence of direction and rough size, not as exact percentages. A shift is called robust only if it (a) keeps its sign and a size of at least 2 points under five scoring variants, (b) survives correction for the classifier’s measured error, (c) has a 95% bootstrap interval that excludes zero, and (d) is more than twice the largest difference between two outlets’ copies of the same broadcast. “Direction only” means (b) and (c) hold but the size is uncertain.

Learn more: how well each subject was detected

Lending data

The NICA Act section compares lending to Nicaragua’s government before and after the Nicaragua Human Rights and Anticorruption Act of 2018 (NICA Act, P.L. 115-335). All amounts are nominal US dollars. Approvals (commitments at board approval) and disbursements (money actually paid out) are shown separately and never added together; disbursements lag approvals by years.

Dataset	What it measures	Coverage
World Bank Projects API	World Bank (IDA/IBRD) project approvals for Nicaragua, by board-approval year	2010–Sept. 2026
IDB Projects Dataset	Inter-American Development Bank (IDB) sovereign-guaranteed loan approvals; technical-cooperation grants kept separate	Downloaded Sept. 25, 2026
CABEI open data (“Préstamos”)	Central American Bank for Economic Integration (CABEI) loan approvals and disbursements, public and private sector	Files dated Sept. 14, 2026
World Bank International Debt Statistics (IDS)	Disbursements, commitments and net flows on public and publicly guaranteed debt, by creditor (IDA, IDB, CABEI, OPEC Fund, IMF, China)	Through 2024 (released Dec. 2025); 2025 projections dropped
US Treasury MDB voting records	How the US executive director voted on each Nicaragua operation at the multilateral development banks (MDBs)	2016–Aug. 2025
Treasury quarterly report on IMF lending (Q4 2020)	US position on the International Monetary Fund’s (IMF) 2020 emergency financing	Oct.–Dec. 2020
Nicaragua Ministry of Finance (MHCP) annual debt-policy evaluations	New external loan contracts by creditor, including Chinese supplier credits	2021–Q1 2026

Why NICA’s effect cannot be isolated

Timing. IDB and World Bank sovereign lending stopped in 2018, before the Act was signed on December 20, 2018. Loans that were never proposed leave no vote.
Other rules. Other US rules (trafficking-in-persons restrictions, human-rights voting law, the Pelosi amendment on environmental disclosure and Executive Order 13851) pushed in the same direction.
What was approved. The only approvals after 2018 were COVID-19, hurricane and disaster operations.
Substitution. CABEI and Chinese contractors replaced much of the lost volume.
An upper bound. The “about $1.6 billion” foregone-approvals figure is therefore an upper bound computed against the 2010–2017 pace, not an estimate of NICA’s effect.
Not public. Treasury’s reports and briefings required by the Act are not public.

Sanctions lists

Sanctions status was checked against the full published list files, not press summaries. “Not listed” means not found on any of these public lists as of the dates below.

List	Publisher	As of
Specially Designated Nationals (SDN) and Consolidated lists	US Treasury, Office of Foreign Assets Control (OFAC)	Published Sept. 23, 2026
EU consolidated financial sanctions list (Nicaragua regime, Reg. 2019/1716)	European Commission	Generated Sept. 22, 2026
UK Sanctions List	UK Foreign, Commonwealth & Development Office	Report date Sept. 21, 2026
Consolidated Canadian autonomous sanctions list	Global Affairs Canada	Latest listing Sept. 22, 2026
Section 7031(c) public visa designations (a provision of the annual State Department appropriations act)	US Department of State press releases	Searched through Sept. 25, 2026

What the lists cannot show

The State Department can bar officials from entering the United States privately under Section 7031(c) and Presidential Proclamation 10309; State said in June 2026 that visa restrictions cover over 2,350 Nicaraguan officials and family members. Those names are not public, so an official “not on a public list” may still be barred from entry.

Learn more: how names were matched

Matching. Names were normalized (accents removed) and matched on a surname plus a second surname or given name, across primary names and all aliases, on every sanctions program, not only Nicaragua programs. A looser surname-only pass was run over entries mentioning Nicaragua, and distinctive tokens were searched in the raw files. False positives (unrelated people with similar names) were removed by hand. OFAC designation dates come from the list’s own event records.

Gold trade data

Four sources: importer-reported trade from UN Comtrade and the US Census Bureau, Nicaragua’s own export totals from its central bank, and the Federal Register for US import-prohibition authority.

UN Comtrade A
US-reported imports from Nicaragua of HS 7108 (unwrought gold, almost entirely bullion and doré), annual and monthly, in dollars. Canadian and Swiss imports from the same source show where the rest of Nicaragua’s gold goes.
US Census Bureau A
Its International Trade API, for the state each import is declared for, the port where it clears customs, and 10-digit tariff lines, whose quantities give the gold content in kilograms. The 2025 Utah figure matches the Kem C. Gardner Policy Institute’s published Utah International Trade, 2025.
Central Bank of Nicaragua (BCN)
Export reports, for Nicaragua’s own export totals. The 2025 gold export total ($1.97 billion) is as reported by La Prensa citing the BCN C, because the BCN’s own tables could not be retrieved. Nicaragua’s coding of export destinations differs from what importing countries record, so destination shares use importer-reported data.
Federal Register and OFAC notices
For the import-prohibition authority added by Executive Order 14088 (Oct. 24, 2022). No determination under it has been published as of Sept. 26, 2026.

Trade values are nominal. US imports are reported at CIF value (including freight and insurance) unless labeled customs value; the difference for Nicaraguan gold is under 0.1%.

Records of Chinese, Russian and Iranian activity

The Russia, China & Iran section draws on a catalogue of 130 public-record items (66 Chinese, 32 Russian, 32 Iranian) with events from January 1, 2018 to September 24, 2026 (a few earlier assets appear only as context), compiled from provider and recipient government documents, company filings, legal texts, Nicaraguan budget reports, and independent and regime media. Taiwan is excluded from Chinese totals.

66 China 32 Russia 32 Iran

Records in the catalogue, by provider (130 in all). A record is not a delivered asset; see the first rule below.

Counts are records, not delivered assets. A record can be a project, program, agreement, delivery series, company portfolio, proposal or a lead that was checked and excluded. Trend charts count only eligible records with a dated event in a given year, and recent years are better documented than earlier ones.
Money measures are never summed across types. Loan commitments, contract values, disbursements, budget accruals and debt service are different things. Chinese and Russian loan figures come from the Finance Ministry’s registered contracts.
An official record shows what it records. A signed agreement or announced handover is not proof of delivery or use. Equipment origin or dual-use potential does not establish foreign access, intelligence collection or basing; those need their own evidence.
The catalogue is not exhaustive. Searches expanded the original leads substantially, but undisclosed transactions, complete arms-transfer registers and customs records were not available.

Security officials profiles

The Military & Police section draws on the author’s compiled profiles of Army, National Police, Interior Ministry and related posts compiled in September 2026 from open sources.

Presidential decrees (acuerdos presidenciales) and La Gaceta
Official Army and Police notes
OFAC and State Department releases
UN reports, including the Group of Human Rights Experts on Nicaragua (GHREN)
Named investigative outlets: Confidencial, Divergentes, La Prensa, Artículo 66, EFE

Each post-holder, date and relationship carries its own evidence tier. Dates of lower precision (month or year only) are flagged, and tenure bars use them as approximations.
Sanctions entries in those profiles were corrected against the lists described in section 5; for example, one designation date was moved from Dec. 12 to Nov. 7, 2019, and two officials previously shown as EU-listed are not.
Projected term ends under the September 2026 constitutional reform are the author’s calculations from the reform’s transitional clause, which counts seven years from each chief’s latest appointment. They assume ratification in January 2027 and are shown as ranges where the appointment date is ambiguous.
Public officials and sanctioned persons are named. Relatives who are private individuals are described, not named, and no personal identifiers beyond those in official sanctions listings are published.

Political prisoners and purges

The Known Instances of Political Imprisonment page and the purge section of the Military & Police page combine the author’s compiled records on political prisoners (research cutoff September 26, 2026) with a second, AI-assisted research stage conducted on September 26, 2026.

Monitor registers: the Mechanism for the Recognition of Political Prisoners and Asociación Memoria y Justicia
UN Group of Human Rights Experts (GHREN), UN human rights office (OHCHR), IACHR
US Treasury, State Department, USTR and Federal Register; EU, UK, Canadian and Swiss lists; OAS and Holy See records
The Murillo broadcast corpus (section 1)
Named outlets: Confidencial, La Prensa, Divergentes, Artículo 66, 100% Noticias, EFE, AP

Stocks are not flows. Counts from different organizations are never added or subtracted; overlapping totals are never summed; no death rate, regression or risk score is computed.
Independent review. Ten studies (eight on prisoners, two on purges) were each checked against primary sources by a separate reviewer; the published findings apply every correction and omit claims that failed. This stage has not been verified by a person.
Privacy. Only public figures are named: officials, officers, party and Indigenous leaders, bishops and prominent opposition figures. Everyone else is described by role, and the data build fails if a withheld name appears in a published file.

Evidence tiers

Claims across the site carry a letter grade for the strength of the underlying evidence.

Official or primary

An official dataset, legal text, government record or primary document read directly: for example a sanctions list file, a lender’s dataset, La Gaceta, or a transcript’s exact wording.

Reputable secondary

Reporting by named, established outlets or research institutions, ideally more than one. Attributed in the text (“according to Confidencial”).

Single report or regime claim

One press report, or the Nicaraguan government’s own claim about itself (for example trainee numbers or handovers). Always attributed; never stated as fact.

Calculation or inference

The author’s computation or analytic judgment built on A–C inputs, such as projected term ends or a foregone-lending gap. The inputs and method are stated.

Counts from Murillo’s broadcasts are a separate category: the transcript text itself is primary, but the counts and classifications are produced by dictionary search plus AI-assisted reading and checked by the author; minor errors may remain.

Corrections and how to cite

Errors are corrected promptly. A correction note with its date is added beneath the affected figure or paragraph, and the change is not made silently.

How to cite the site

Guerra, Gil. Data for US-Nicaragua Policy. 2026.

How to cite a single figure

Guerra, Gil. “[Figure title].” Data for US-Nicaragua Policy. 2026. [Figure link]

Add the figure number and title, and the figure’s own link (use the Link button under the figure).

Corrections. Numbers that were revised during the research are shown in their corrected form only; for example, the US share of Nicaragua’s gold exports is roughly a third by importer records, and topic-classifier agreement is 68.5%, not the earlier 83%. To report an error, contact the author through Hudson Institute.

Data downloads

Every chart on the site has four actions beneath it. Figures and data may be reused with attribution.

TableShow data table: the exact values plotted, readable by screen readers.
CSVDownload data (CSV): the same table as a spreadsheet file.
ImageSave image (PNG): the chart with its title, legend and source line, ready for a slide or memo.
LinkCopy link to figure: a direct link that opens the page at that figure.

The processed data behind each page are also published as JSON files under /data/ on this site. Source files that are not public (for example working spreadsheets) are not republished.