Issue #62: Loss-curve validation on held-out data #59

Task

Validate loss curve on held-out data (90% train / 10% val) to ensure no overfitting before Parameter Golf submission.

Acceptance Criteria

• Split TinyShakespeare: 90% train / 10% val
• Run 1000 steps on train set
• Compute BPB on held-out val set
• Check: val_bpb < 2 * train_bpb (no severe overfitting)
• Plot loss curves for train vs val

Deliverables

• experiments/loss_curve/train_loss.csv
• experiments/loss_curve/val_loss.csv
• Plot: experiments/loss_curve/loss_curve.png
• Overfitting diagnosis

Configuration

• Schedule: flat_3e4 (Issue feat(Φ8): Reduce IGLA-GF16 size from 18.8MB to ≤16MB #54 winner)
• Model: IGLA-GF16 with real forward pass (Issue feat(phd): Lane A scaffold — LaTeX skeleton with 34 chapters + 8 appendices #55 fix)

Refs

• feat: integrate Chrome Extension (Dioxus/Wasm) into Cargo Workspace crates/ #61: 3×1000 steps scaling
• feat(phd): Lane A scaffold — LaTeX skeleton with 34 chapters + 8 appendices #55: BPB fixed (realistic training)

Go/Pivot Gate

Only proceed to Parameter Golf if val_bpb is plausible (not severely overfit).