Wine quality assessment relies on expert tasters whose judgments vary with mood, health, palate fatigue, and the residual flavors of whatever they ate for lunch. The same wine evaluated by the same taster on different days can receive different scores. This isn't a flaw in the tasters — it's a feature of the sensory system. Human perception is contextual.
UV-Vis spectroscopy combined with support vector machines now predicts both sensory attributes (astringency, bitterness, herbaceous notes) and wine origins from grape juice alone — before fermentation even begins (arXiv:2507.20114). Origin prediction exceeds 91% accuracy. The system works at two stages of winemaking, meaning decisions about blending and processing can be made upstream, before expensive fermentation commits resources.
The surprising finding: the spectral signature of the juice predicts the sensory experience of the finished wine better than the spectral signature of the wine itself would predict the taster's score. The input that's furthest from the experience is more predictive than the input that's closest.
This happens because the juice measurement is stable — it captures the chemical fingerprint without the variability that fermentation, aging, and serving conditions introduce. The taster encounters the final product, which has been modified by dozens of uncontrolled variables since the juice stage. Predicting from the juice works better precisely because the measurement precedes the noise.
The broader lesson: in prediction tasks where the target is inherently noisy (human sensory judgment), measuring earlier in the causal chain — before noise accumulates — can outperform measuring later, even though later measurements seem more relevant. Proximity to the outcome is not the same as informativeness about the outcome.