Methodology

The feedforward multilayer perceptron uses standardized experience and skill, plus one-hot encoded education, role and department. Hidden layers use ReLU or tanh; the salary regression output is linear. MSE gradients are computed by the chain rule and optimized with Adam or minibatch SGD.

Reproducibility & leakage prevention

A fixed seeded shuffle reserves 20% for testing. Validation is taken from the remaining records using the training seed. Numerical scaling, categorical vocabulary and target standardization are fitted only on training data. Unknown categories use an explicit unknown slot. Test data is evaluated only after the final epoch. Artifacts store weights, biases, preprocessing, configuration and evaluation together.

Data provenance

The initial 480 records are reproducible synthetic examples generated with seed 2026. They are not real employee salaries and carry no market-validity claim. The synthetic generator defines a teaching dataset; predictions always come from the trained ANN, never from that generator. User CSV uploads replace the demo and are marked unverified.

Limitations & intended use

Educational exploration only. Not for hiring, compensation decisions or employment screening. No fairness audit, causal feature attribution, confidence interval or external validation is provided. Salary outputs can be negative or implausible; they are not silently clipped. Protected attributes are not predictive features. Categories may still carry societal biases.

Artifacts & storage

The active model and up to six version summaries are saved per user. Predictions are saved only on request, up to 500 records. Dataset changes invalidate the active model. Stored artifacts are validated before loading. Storage failures are reported and the current session remains usable. No server or database is provisioned by this browser edition.

Verification

The engine checks numerical gradients against finite differences before training. Training rejects nonfinite loss. Metrics are computed in original annual salary units on the held-out test data. This is not a substitute for a full independently executed automated test suite or production security review.