What lies behind the peaks and dips?
An article-based audit of aid, foreign-policy and defense coverage. Real policy events appear in the series, but their interpretation depends on who is being evaluated, which outlets are reporting, and how many articles support the change.
What the first audit establishes
A broad policy disruption, with several stages
Under the n10 screen, all 21 qualifying countries decline in January; 41 of 44 decline in February. Liberia’s February change is −92.3 points, from 53 articles across eight outlets. Malawi articles describe university disruptions and health-service risks. The chronology separates the 20 January order, 24 January stop-work instructions, reported waivers and local consequences. February’s comparison period already includes January’s disruption. Read Liberia →
Diplomatic conflict followed by reengagement
July’s −53.3-point change accompanies the Askarov award dispute and the decision to terminate a cooperation agreement; October’s +59.9 accompanies renewed diplomatic engagement. These are identifiable bilateral episodes. The negative stories nevertheless mix U.S. criticism of Kyrgyzstan with Kyrgyz criticism of the U.S. Read July → · Read October →
An agreement brings competing political voices
The +40.9-point shift, based on 82 articles across four outlets, coincides with EDCA and Obama’s visit. Officials offer security assurances while organized opposition questions sovereignty and U.S. credibility. This is a change in the balance of reported frames and speakers; it does not establish a uniform Philippine response. Read the episode →
Some apparent signals should not be interpreted
Malawi’s June 2025 aid rebound rests on one article about community-led development with no clear U.S. actor. A Ukraine comparison month includes Ukrainian military acronyms mistranslated as U.S. references. Senegal’s January 2020 series includes the Paris-based Nouvel Obs. Low count + relevance → · Translation → · Outlet geography →
A real event and a classification problem can coexist in the same month. The audit records both. It does not relabel the database or calculate error rates from this purposive sample.
Broad changes before country-level interpretation
| Month | Topic | Matched countries n10 screen | Equal-country change | Pooled-article change | Equal change n5 screen | Equal change n20 screen |
|---|---|---|---|---|---|---|
| 2025-01 | Aid | 21 | -62.8 | -59.2 | -47.6 | -55.2 |
| 2025-02 | Aid | 44 | -47.4 | -50.6 | -56.9 | -48.9 |
| 2016-11 | Foreign policy | 55 | +18.4 | +16.7 | +18.2 | +17.1 |
| 2021-08 | Foreign policy | 56 | -13.2 | -13.5 | -12.8 | -14.6 |
| 2014-12 | Defense | 34 | -20.7 | -17.7 | -22.2 | -18.7 |
| 2020-01 | Defense | 51 | -15.5 | -16.2 | -16.9 | -15.1 |
Changes are in net-tone points, relative to the preceding three months pooled. n5/n10/n20 require at least that many articles in the focal month and in the whole three-month baseline. Country sets change across screens; pooled article weights can change between periods. These are illustrative large movements, not a complete ranking or causal estimates. November 2016 policy coverage includes election expectations, not an assessment of Trump’s governing performance.
Why large numerical extremes are not automatically the best cases
Only 2 of the 24 selected n10 extremes also meet n20. The audit therefore adds 12 larger-volume cases and three country illustrations of broad changes, as well as six very-low-count cases and three ordinary comparison months. Aid has only 30 country-month observations eligible at n50, versus 4,723 in foreign policy and 2,407 in defense. A common threshold does not yield equally informative coverage across topics.
The evidence timeline
Browse all 48 episodes, or filter to the 15 larger-volume cases and global illustrations. Country means the outlet’s assigned country. Each card separates publication month, reported event dates, tone, attention, media composition and reading limits.
Scope and methods
What was measured. The frozen paper panel contains 12,354 country-months and 37,062 country-topic-months, January 2012–June 2026. Counts reconcile exactly with the 71 country exports and with reconstructed source totals. Aid contains 58,516 topic articles, foreign policy 944,256 and defense 381,813. The main change compares the current month with pooled counts in the previous three calendar months. Adjacent-month changes and sentiment levels are supplied in the ranking files.
What was read. For each episode, a deterministic sample balances sources and predicted sentiment: up to 16 focal articles and eight preceding-period articles, with all focal articles included when the live count is at most 16. There are 972 reviewed episode-sample records (964 distinct article IDs). English stored bodies, translated titles/leads and selected original-language texts were inspected to differing depths; each episode reports its scope. These are AI-assisted diagnostic judgments. They are not independent human adjudication, full-text reading of every original article, or a representative estimate of classifier accuracy.
Media and country checks. We count active topic outlets, the largest outlet’s share, shared-outlet sensitivity, changes in country all-news volume, repeated exact URLs and repeated titles. Shared-outlet checks refer to topic-active outlets, not verified continuous all-news collection. Syndicated publication is genuine media exposure but does not provide independent corroboration. No per-outlet all-news denominator is available from these exports. Source metadata contain 415 country–domain assignments across 414 domains; Balkan Insight is assigned to both Serbia and Kosovo. Nouvel Obs is assigned to Senegal although ACPM records its Paris address. The corrected analysis removes Nouvel Obs from Senegal; these archived figures retain the original assignment.
Snapshot differences. The 28 October live retrieval matches frozen counts in 45 of 48 episodes. Bangladesh aid September 2015 has 9 versus 10 focal articles. Belarus policy September 2021 has 50 versus 52 focal and 193 versus 203 baseline articles. Turkey defense December 2014 has 94 versus 95 focal and 441 versus 458 baseline articles. The rankings retain the frozen counts and the article review uses the live snapshot; individual cards flag this distinction.
Selection and inference. Current and pooled-baseline count screens of 5, 10, 20 and 50 are exploratory. Candidate months also require current country feed volume to be between half and twice its prior-three-month mean. This does not prove collection stability: gradual feed declines can pass a short-window screen. Selecting extremes creates regression-to-the-mean and multiple-search problems; tests on these same selected months would not independently establish event significance. Baselines can contain earlier stages of the same event. There are no causal effects or p-values in this report.
Next research stage
- Adjudicate the measurement failures first. Use the saved IDs to review U.S. relevance, topic, evaluated actor, speaker, original-language entity names, publication date and outlet country. Double-code a new sample spanning ordinary months, neutral labels and excluded articles before estimating error rates.
- Build an event catalogue independently of peaks. Record announcement, implementation, suspension/termination, waiver, visit/agreement, military operation and information disclosure separately. Link event country and publication country; retain competing events and attribution uncertainty.
- Study attention and framing as separate outcomes. Compare topic share, positive and negative attention per collected-news volume, and tone conditional on coverage. Begin with official dates and a stable outlet panel. Country and media hypotheses—aid dependence, alliances, regime and media freedom, ownership and government links—need dated external measures; this audit does not estimate their effects.
The immediate paper contribution can be how policy stages and local political voices shape media attention and evaluation. Public opinion, policy efficacy and an instrumental-variable design require separate validation and identification; the observed sentiment series alone establishes none of them.