HELIX-BAKEOFF-0611

The original 'pulse' absolute-time-gated HSL recurrence is the right design for HELIX's oscillator-bank state layer.

Three 13.85M-parameter arms (pulse / nopulse / rotation HSL variants) trained 10,000 steps each on WikiText-103 under seed control, overnight while a sibling project occupied the GPU. A same-morning confirmation suite replicated on fresh seeds and separately tested per-pair vs. per-bank ω parameterization.

• ◆HSL variant (pulse / nopulse / rotation)
• ◆ω parameterization (per-pair vs. per-bank)

• →Val PPL @ step 10,000: rotation 194.76, nopulse 278.13, pulse 290.04.
• →Confirmation suite replicated the ordering on fresh seeds: rotation 196.69, nopulse 279.29.
• →Per-pair ω (196.47) bought nothing over per-bank ω (196.69) — the cheaper parameterization was adopted.

• ·The founding 'pulse' hypothesis lost to even the plain 'nopulse' gate — a clean falsification, not a close call.
• ·A speed A/B showed rotation ~2.8× faster than nopulse, but only under GPU contention; the lab explicitly forbade citing that number publicly.

Rotation-HSL adopted; pulse retired the same day. Every subsequent HELIX generation (56M, 127M, 209M) trained on rotation-HSL — the one architecture decision that survived the entire program.

Launch the 56M flagship run the same day.