Methodology

Corpus

The reconciled census contains 1,527 University News articles dated 2023-10-01 through 2026-08-10, across administration, academics, and student-life subdesks. Full article text was used only in the private working corpus and is not redistributed.

Preregistration and codebook

Questions, measures, exclusions, directional hypotheses, and validation thresholds were specified in the preregistration and University News addendum. The codebook defines actor, framing, scrutiny, sourcing, verification, and follow-up fields.

Manual validation

A diversity-oriented audit of 230 articles included 50 independently double-coded rows. Because selection maximized strata and screening diversity, the audit is not a self-weighting probability sample.

Reliability gate

None of 19 automated semantic measures reached the preregistered κ threshold of 0.6. Their population-level findings are therefore suppressed, not promoted as evidence.

Statistical approach

Paired gaps compare administration and challenger treatment within eligible audit rows. Bootstrap intervals describe stability within the audited set. They are not population confidence intervals. Sample size and interval method appear with each estimate.

Why sentiment cannot prove bias

Words derive meaning from quotation, topic, evidence, and newsworthiness. Different treatment may be justified by different conduct. Sentiment cannot identify motive, editorial causation, or whether skepticism was warranted. Those claims require contextual coding, documentary evidence, comparison, and fair response.

Known limitations

Leadership groupings are descriptive time controls, not causal assignments. Private full text cannot be audited publicly word-for-word. Manual judgments remain fallible. Article updates may change after retrieval. Read the complete limitations, data dictionary, and reliability results.