# Research basis and scope

Actual source read: [UCI186 Wine Quality](https://archive.ics.uci.edu/dataset/186/wine+quality), DOI10.24432/C56S3T, original native ZIP download, every source names line and both actual CSVs. Authors Paulo Cortez, Antonio Cerdeira, Fernando Almeida, Telmo Matos and Jose Reis (2009), CC BY4.0. The source describes red/white Portuguese Vinho Verde chemical tests and sensory scores, ordered and imbalanced. UCI summary4898 is white only; original names and CSVs show1599red+4898white, all retained. Score source is median of at least3expert evaluations,0..10scale. Individual chemical units not present in the actual files and are not invented here.

Original source cites [Modeling wine preferences by data mining from physicochemical properties](https://doi.org/10.1016/j.dss.2009.05.016), Decision Support Systems47(4):547–553. Original author HTTP portal navigation timedout, fullpaper not read/downloaded/reproduced. This experiment uses no sourcepaper training method, published score or figure as its own evidence. No paper text/figures/code redistributed; original names are the unmodified licensed dataset documentation.

The experiment compares two independently trained source-domain MLP regressors with the same ROW/update budgets and paired initial arrays. It tests both directions on group-separated, fixed known-data test subsets, without target-aware split or sample selection. It does not isolate covariate shift: source quality distributions, feature support, train normalization, distinct own-domain validation, effective duplicate weighting and source access differ. Source-only checkpoint selection is deliberate. No population/cross-vintage/causal chemical interpretation.

There are5318exact11-feature groups in6497rows,1179duplicate-excessrows,2cross-color exact groups, no conflicting-quality groups in current bytes. Group hash excludes color/quality and prevents a vector crossinginitialfit/val/test. This prevents exact duplicate leakage under the stated definition, not unobserved producer dependence or arbitrarily correlated wines. Sampling fit/validation pools can retain only some duplicate rows within the same initialrole, leaving unused rows explicit. Quality classes are not rebalanced from targets.

Tanh-MLP MSE/Adam code reuses the author's own basic math pattern from accepted AirfoilSplitLab under Apache2; new11-feature arrays, normalizers, source partitions and all six actual fits are fresh. Architecture/optimizer are simple fixed controls, not a novel method or reproduction of Cortezetal's best source model. Continuous outputs not rounded or clipped. Training mean baselines are analytic constants, not extra neural fits.

Protocol/splits/sourceaudit/config frozen beforefit and first current-model test. All source labels/distributions/duplicates were already audited, so PRETEST is prospective for this code/model evaluation, not a historically blind benchmark. Full failures and target-domain subsets remain, with explicit uncertainty/causality/measurement limits.
