Agriculture & Food, Reinvented.

The forward model is a deliberately designed stacked ensemble — not one model, but three complementary base models whose predictions are combined by a fourth.

GLMM (Generalised Linear Mixed Model) provides rigorous statistical inference with proper uncertainty quantification and correct handling of nested farm data. XGBoost captures the complex non-linear interactions a linear model misses. Random Forest adds stability and resistance to overfitting in small datasets — hundreds of independent trees, each on a random subset, averaged so they do not all make the same mistake.

Each base model produces its own prediction for every farm. Those three feed a ridge meta-learner — a linear model with L2 regularisation that learns the optimal weighting of each contribution. The combined prediction is more accurate and more stable than any individual model.

The reverse model takes biological markers from a food sample and infers what farming system produced it — enabling independent verification, because a food’s biology cannot easily be faked.

The markers — odd-chain fatty acids, microbial diversity, polyphenol profiles — arise from fundamental biochemical pathways in living organisms. C15:0 and C17:0 in particular are synthesised exclusively by bacteria in ruminant digestive systems; they cannot be supplemented into milk or meat, so they are genuinely diagnostic of the farming system. These biological constraints are what make the reverse model fraud-resistant.

Raw probabilities from machine-learning models are often poorly calibrated — a model that says “80% confident” might be right only 60% of the time. Vitagri applies isotonic-regression calibration so the confidence scores are honest. For continuous outcomes it reports not a single point but a calibrated range — 4.8–5.6 mg/100g at 95%.

For the reverse classification model, the calibrated probability drives the certification decision directly:

The Brier score — the mean squared difference between predicted probabilities and actual outcomes — must be below 0.15 for the system to count as well-calibrated. Calibrated confidence intervals are what separate this from a marketing claim.

Verification reads the loop in two directions. The third use reads it as an optimiser. Once the forward ensemble can predict nutrient density from a set of decisions, it is also a fast surrogate simulator: score a proposed combination of farming choices in milliseconds instead of a growing season.

That turns “which farming change should we try next?” into a search over the model’s predicted surface. The system ranks candidates by expected nutrient-density gain while accounting for its own uncertainty — an explore/exploit trade-off handled with Bayesian-optimisation and active-learning methods, structured as a design-of-experiments programme so each trial is maximally informative.

The highest-expected-gain candidates are promoted to real field trials. The reverse model independently validates the outcome from the harvested food’s biology — so a predicted gain only counts once it is biologically confirmed. Validated results are folded back into the training set, the ensemble is recalibrated, and the search space narrows for the next cycle.

Every certified farm generates validated, labelled data that can retrain the model. More data means more patterns, better accuracy, and more credible certifications — which attracts more farms, a compounding virtuous cycle. The defensible asset is the accumulating verified dataset and the calibrated model trained on it, not any single algorithm.