API inputs used title/maintext (66 English-original,109 non-English-original articles). A separate 33-article diagnostic using title_translated/maintext_translated was submitted on 16 October 2026. Its results are outside the historical counts shown here.
Submission and local outcomes
“Outside this submission” identifies the part not included in the submission captured here. This snapshot does not assert what may have happened in later, unseen runs.
Parser request acceptance
Request-level acceptance is mechanical. A single failing record can quarantine its entire request. Pending requests have not been scored as failures.
Source text review
A hold is a downstream review flag, not proof that an article is unusable. Stored-text fidelity does not establish publisher completeness, language quality or scientific validity.
Assistant review has found errors in the returned labels, including some records that passed mechanical checks. The charts below display unchanged model decisions for diagnosis; they include possible errors. They are not corrected labels or evidence of scientific accuracy.
Across all 175 returned records, assistant substantive review identified 74 correction cases, 54 requiring adjudication and 47 with no obvious issue. Within the 88-record chart subset, these counts are 40, 27 and 21 respectively. These are review judgments, not human-gold labels or accuracy scores.
Provisional model output diagnostics
Only records with exactly one returned output, mechanical request acceptance and no source-quality hold enter these counts. They remain unvalidated model outputs.
Coding meaning and limitations
Actor and target categories describe institutional role in the selected action. They do not establish direct affiliation with the incumbent or cabinet, and neither role nor membership alone establishes action alignment.
Frame +1: article voice clearly favors gov. Frame −1: article voice clearly disfavors gov. Frame 0: descriptive or meaningfully balanced text; reporting a favorable or unfavorable action or condition; attributing a source’s position without adopting it; or no clear presentation-level cue.
Frame concerns observable article-voice presentation toward the coded government, not the underlying event’s positivity/negativity or the author’s private intent. Missing and null are displayed separately from zero.
These are unweighted calibration counts. They are not population estimates, prevalence, trends, model accuracy or a research release. Action tags can have up to two values per article; missing, null, empty and invalid values remain separate.
What this preview leaves out
This distributable snapshot contains no article text, titles, evidence quotations, article or request identifiers, source domains, account identifiers, credentials or local filesystem paths. It has no upload, retry or submission controls and makes no external network requests.
Human adjudication and research validation are still required. The current human-validated and scientifically accepted counts are both zero.