Goals

Q4 2026 objectives

• Finish training encoder
• Evaluate
  • Benchmarks vs SOTA
  • Visualizations (PCA, RCDM)
  • Probes (MLP)
  • Weights analysis (SwiGLU activations, dead paths, attn softmax dist)
  • Kernels
  • Eval using labeling suite labels
  • Eval at various checkpoints
  • Check anisotropy
  • Multi-head ablation
• Train with findings applied on v3 data on cluster
• Scale labels to 10k
• Train with higher res cooldown
• Evaluate
• Generate 50k text-renderer samples
• LLM pipeline

Handbook