The signal is smaller than the noise: what 21,684 glaze recipes taught me about the limits of machine-learning colour prediction

Industry specifications for ceramic glaze colour prediction commonly target R² ≥ 0.985 and ΔE₀₀ < 0.5. A gradient-boosted ensemble trained on 21,684 real, already-fired glaze recipes reaches R² = 0.35 (MAE 38.0 sRGB units), ahead of all four single-model baselines evaluated here, so model capacity is not the constraint. Two improvement hypotheses were tested and rejected. (1) Features: 41 physics-derived features from classical glass science — thermal expansion via six independent additive models, boron coordination ratio, polymerisation index, weighted chromophore loading — changed R² by −0.0000 (95% CI [−0.0048, +0.0055]), because those quantities are already implicit in the raw oxides. Thermal expansion is a linear function of the oxides; the model could already compute it. Renaming is not adding. (2) Labels: four independent annotators over 200 samples agreed exactly (0.00 disagreement), so labels are not the noise source. Measuring the labels instead revealed the real limit: the two candidate colours auto-extracted from each photograph sit 150.4 sRGB units apart, and a human cannot adjudicate in 17.7% of cases. 82.7% of colour variance is within-chemistry variance (mean dispersion 41.7 sRGB units) — chemically identical recipes fire to visibly different colours, because firing curve, layer thickness, particle size and quantified atmosphere are unrecorded, and colour is read from uncontrolled photographs rather than a spectrophotometer. The ceiling is a property of the data-generating process, not of the learner. Both natural modelling remedies are measurably empty. This is a negative result published because it is actionable: it shows where the information actually is, and saves the next researcher the two experiments that do not help. AI assistance is declared in AUTHORSHIP.md. An AI research assistant was used as a tool under human direction and verification, credited in the acknowledgements rather than as an author. Its own hypothesis — that physics-derived features would improve accuracy — was the one the data refuted.