Personal Content Filtering with a Local Student Model: Architecture, Abstention and a Failed Precision Gate

Personal content filters can give individuals control over unwanted engagement, but hiding legitimate criticism can also reduce access to useful information and impair accountability. This report examines Nullifier, an implementation that obtains offline teacher labels, trains a small local classifier and applies decisions to browser responses or Telegram messages. The inspected student uses frozen MiniLM sentence embeddings, optional author features and a calibrated logistic head. An archived training configuration records 673 examples, including 66 positive teacher labels, an out-of-fold ROC-AUC of 0.7635 and average precision of 0.2762. No threshold met the configured precision criterion, so the saved configuration disabled text-based filtering and retained list-based decisions. This negative result is central. Source inspection also identifies global preprocessing and model selection that limit interpretation of the reported cross-validation. We describe decision provenance, sink-specific recovery and an evaluation protocol focused on user control, false-positive burden and preservation of contestable evidence. No improved well-being, harassment reduction or financial-access outcome was measured. Preparation of this report was assisted by Codex. The research programme is directed toward human agency, financial inclusion, and evidence-based accountability. Intended social benefits are distinguished from measured technical results in the manuscript. The manuscript and newly authored analysis are offered under CC BY 4.0. Any supplied source files retain the license terms stated in the supplement; the record license does not replace third-party source licenses.