USG SENTIMENT EVENT AUDIT — 28 OCTOBER 2026
Scope: 71 countries, January 2012–June 2026; 37,062 country-topic-month observations. Frozen counts reconcile exactly with source exports.
Selected: 48 episodes in 32 outlet countries, 16 per topic. 24 n10 stress tests; 12 n50 larger-volume cases; 3 global illustrations; 6 very-low-count cases; 3 ordinary comparisons. Only 2/24 n10 stress tests also meet n20. Screens are exploratory, not reliability guarantees.
Retrieved: 6,801 unique stored articles. Read: 972 episode-sample record occurrences, 964 unique IDs; text depth varies, and not all original-language bodies were read in full. Evidence table: 510 selected annotated records. AI-assisted review, not independent human labels or accuracy estimates.
Tone = 100*(positive-negative)/all topic articles including neutral. Change = focal month minus pooled prior-three-calendar-month tone. Attention = topic articles/all collected news in the outlet-country feed. Tone and attention are different outcomes.
Country denotes outlet assignment, not event location, author nationality or audience. Shared-source sensitivity uses outlets with topic coverage in both periods, not verified continuously observed outlet feeds.
Frozen/live mismatches: E02 Bangladesh aid (10 vs9 focal); E42 Belarus policy (52 vs50 focal,203 vs193 prior3); E46 Turkey defense (95 vs94 focal,458 vs441 prior3). Frozen counts retained for rankings.
The timeline distinguishes genuine events, information releases, policy stages, mixed targets, classification/translation/geography issues, and sparse samples; these can coexist. Dates primarily come from articles; external official verification is attached only where performed. No causal attribution or statistical significance is established.
Read event_audit.html for findings, article links, diagnostics and proposed next steps. Selected aggregate outputs and evidence summaries are supplied with the public report. Do not treat an eventual live re-query as the frozen snapshot.