Algorithm 1. Corrected Fold-Local Bandit-Style Sample Weighting

Input:

Training partition $T_f$ for fold f; held-out partition $V_f$; labels y; action set

A = {0.25, 0.50, 1.00, 1.25, 1.50}.

learning rate α, discount factor γ, and exploration rate ε.

Output:

Fold-local sample weights $w_f$, trained residual MLP for fold f.

1. Initialize a new fold-local Q-table $Q_f\in {\mathbb{R}}^{N\times \left|A\right|}$ with all values set to zero.

2. For each training sample ${i\in T}_f$, select an action index $a_i$ using an ε-greedy policy.

3. Assign the preliminary sample weight $w_i$ = A[$a_i$].

4. Train a probe residual MLP on $T_f$ using the preliminary weights.

5. Predict the labels of the same fold-local training samples $T_f$ using the probe model.

6. Compute the reward:

$r_i=\left\{\begin{array}{c}1 if {\widehat{y}}_i=y_i,\\ 0 \mathrm{otherwise}.\end{array}\right.$.

7. Update $Q_f(i,a_i)$ using the observed reward.

8. Reassign final fold-local training weights using the updated $Q_f$.

9. Train a fresh residual MLP on $T_f$ using the final weights.

10. Evaluate the final model once on the held-out fold $V_f$.

11. Discard $Q_f$ before proceeding to the next fold.

Note: The held-out fold $V_f$ is not used in Steps 1–9. It is used only in Step 10 for the final performance estimation. No Q-values, rewards, selected actions, sample weights or prediction outcomes were transferred across the folds.