Review Package

File	Contents
midas_recommendation.html	The dashboard itself — open in any browser, all data is baked in
personal_midas_quantile_raw.csv	Per-player rows: hero, role, buy time, personal NW advantage, expected win, actual win — primary input for the heatmap grids and model training
midas_quantile_raw.csv	Team-level NW quantile rows — joined with the personal data to build the 5×5 grids
midas_models.json	The 6 fitted logistic regression models: coefficients, intercepts, scalers, n, pseudo-R²
calibration_model.json	Baseline (no-Midas) logistic regression: 3 coefficients mapping NW difference + game time → expected win probability. Used to compute the expected_win column in every other file
personal_quint_ranges.json	NW advantage quintile boundaries (p20/p40/p60/p80) per role per time window — defines how raw gold values map to the 5×5 grid axes
midas_validation.csv	Non-Midas game sample used to calibrate and test the baseline model
validation_raw.csv	Out-of-sample test results: 312 unseen Midas games, per-row predicted vs actual win probability for both the Midas model and null baseline