HXO-EXP-0007

Replacing only the dense MLP projections with same-width one-stage TensorAxis operators improves the validation-loss-versus-resource frontier on edge hardware.

Six sequential 4,096-step runs from source commit 5c758b2: three dense / TensorAxis seed pairs in dense→Ω, Ω→dense, dense→Ω order. Each arm trained on 2,097,152 tokens with FP32, batch 1, sequence length 512, dense attention, fixed-window validation every 256 steps, and exact full-stream validation at the final step.

• ◆MLP projection: dense versus one-stage pure TensorAxis
• ◆Seed: 31415 / 31416 / 31417
• ◆Run order, wall time, throughput, memory, and held-out validation quality

• →Dense mean full-validation loss / PPL: 5.56445 / 260.98. TensorAxis: 5.72141 / 305.34.
• →TensorAxis was worse in all three seeds; its mean validation-loss penalty was 0.15696.
• →TensorAxis saved 22.37% physical parameters, 22.83% estimated active parameters, 20.68% estimated FLOPs per token, and 32.19% mean peak MLX memory.
• →Dense reached fixed-window loss ≤5.8 in 24.13 seconds mean; TensorAxis required 39.65 seconds, 64.3% longer.

• ·The structured arm was nominally faster on ordinary steps, but one dense run thermally slowed; no stable pooled speedup was accepted.
• ·The resource savings were real, but they did not buy an equal quality frontier.
• ·Same numeric seeds did not force shared layers to identical initialization, motivating the stricter matched rescue.

The pure one-stage TensorAxis MLP failed the quality/resource-frontier gate. Do not scale it or authorize custom kernels; run one bounded matched-initialization rescue.

• HXO-EXP-0007 result JSON ↓Complete six-run metrics, hashes, resource accounting, validation results, verification, and limitations.

HXO-EXP-0008 — identity-diagonal matched rescue. Open HXO-EXP-0008 →