The Language of Letting Go: Text-Label Co-Production and the Limits of Computational Inference in CEO Departure Data

Can textual descriptions of chief executive departures recover whether an exit was forced, and what does such recoverability establish when text and label share a production process? We conduct a leakage-aware computational audit of an open database of 5,516 coded CEO departures in S&P 1500 firms from 2000 through 2018, whose narrative field consists of coder-constructed summaries rather than firm announcements. Among 5,365 discretionary departures with usable text, forced cases are substantially more negative than other exits (Cohen's d = − .89), and dominant-topic membership is associated with forced status (Cramer's V = .41 in the four-topic model, .56 in a seven-topic sensitivity model). A row-random benchmark produces ROC-AUC = .891. Keeping each firm's observations in a single validation fold and learning all text transformations from training data lowers the figure to .860 (95% cluster-bootstrap CI [.849, .872]); it remains .838 after direct label and coder-meta cues are masked and .824 under a strict topic-only specification. A model trained on 2000–2013 cases reaches .886 in 2014–2018. High-dimensional TF-IDF improves ranking (.937) but yields compressed, poorly calibrated probabilities. These results establish robust within-database recoverability, not evidence that firms' impression management leaks the truth. We develop text-label co-production as a measurable extension of familiar criterion-contamination and annotation-artifact problems, and propose provenance validity as the standard that connects a text's origin to the inferences it can support. The caution applies wherever coders both label events and write the narratives that describe them; testing firm communication directly requires a source-separated design.