HostelPunk classifier evaluation, May 25, 2026 This is an aggregate excerpt from the retained internal release review, 2026-05-25_classifier-prod-run-postmortem.md. No guest review text is included. It documents a historical evaluation, not current production performance. Constraint: reduce repeated LLM work without degrading review analysis. Alternatives: retain existing outputs and methods, or use a new classifier trained on a larger, cleaned production-data corpus. Decision recorded by the review: do not deploy the replacement checkpoint. Comparison on the same database-backed labeled reference set: Signal New classifier F1 Existing stored-output F1 Overall positive 0.9048 0.9677 Overall negative 0.9388 0.9697 Described as social 0.8571 0.8627 Described as antisocial 0.5455 0.0000 Paid-for-review mention 0.0000 0.8913 Asked-for-review mention 0.0000 0.0000 Suspicious-review mention 0.7632 0.9730 Combined review-suspicion signal 0.1497 0.9922 Macro F1 0.5199 0.7071 F1 combines precision and recall. Macro F1 is the unweighted average across these eight signals. The reference is stored database outputs, not a single fresh baseline model run. These are signal-level results, not ranking accuracy. Limitations: rare-signal support was sparse, label definitions and source mix shifted, and the new train/test split was not comparable to the older split. Threshold calibration had only three paid-review positive validation examples. The comparison does not establish accuracy over the entire review corpus. No inference-cost saving or throughput measurement is claimed for this rejected candidate. The original review records a no-deploy decision, not a new audit of current production state.