# Interpretation limits

The source scores are ordered sensory assessments. MAE/MSE regression treats score differences as equally spaced numerical errors; that is a modeling assumption, not evidence of an interval-scale physical quantity. Source quality does not certify a wine, establish health/safety, explain a causal chemical effect or predict general consumer preference.

On the341-row red holdout, the white-trained neural model's descriptive three-seed mean MAE is .746975, worse than the white fit-mean constant's .636872. All four source/target mean baselines remain visible beside neural metrics. On3969same-seed/source-case comparisons, the in-domain model is worse1585times (321red-target/1264white-target); its average advantage is not a per-case guarantee. The fixed default red:0002 also favors the cross-domain model at seed17. It was selected before current-model tests and is retained.

All6497original records,1179duplicate-excessrows, unused training/validation rows, all1323test cases, low-support/unsupported quality groups and every error remain. Exact input identity is only one leakage definition; it is not a producer ID or proof of independence. Quality-specific groups can be tiny or empty, and repeated seeds reuse the same cases. No confidence intervals, significance, exchangeability or population-level uncertainty claim.

The red/white models have equal training/validation row counts and update budgets with paired initial arrays. Their source labels, feature support, normalizers, duplicate weights and own-source checkpoint criteria differ. This does not isolate a causal distribution-shift mechanism, and does not measure unlabeled adaptation or a future unknown domain. All original source labels and distributions were already audited; the first-test gate controls this run, not historical blindness.

The demo accepts vectors within combined-source marginal bounds, including unrealistic combinations of maxima/minima. Marginal membership is not joint chemical support or physical validity. Exact numeric source matches establish vector consistency, not provenance of user measurements. Continuous outputs are not clipped; current test outputs within0..10 do not bound predictions for arbitrary accepted inputs. No calibrated per-input uncertainty is supplied.

Native exports contain all11values, all6outputs, source matches, normalizers, artifact hashes and attribution. Inputs stay in page memory until cleared/refreshed; deliberate downloaded files remain where the user saves them. Hashes establish byte consistency, not authenticity or permission to reuse arbitrary supplied measurements. Original public data and derived weights/results/figuresCC BY4.0; original codeApache2.0. No employer/private or competition material.
