src/speech_decoding/models/v14_converged_v3/towers.py

Selected training excerpt · lines 50–50 · source SHA-256 5610fc7e8fc4

50PRED_LAYOUT: tuple[str, ...] = ("L1",) * 6  # Design B (Ben 2026-07-15): L1-only, 6 layers; class W dropped ⇒ no cross-shaft predictor