HXO-EXP-0005

A same-width TensorAxis MLP can reduce parameters, arithmetic estimates, and memory on Apple Silicon while remaining numerically healthy enough for a real-data pilot.

Two forward/reverse-order dense and TensorAxis pairs at d_model 384, two layers, six attention heads, sequence length 512, FP32, eight optimizer steps, and 4,096 deterministic smoke tokens per run. Attention remained dense. Both final checkpoints were reloaded and reproduced their validation values.

• ◆MLP projection: dense versus one-stage TensorAxis
• ◆Run order: forward versus reverse
• ◆Measured resource accounting, peak memory, and ordinary-step throughput

• →TensorAxis used 22.37% fewer physical parameters and 20.68% fewer estimated FLOPs per token.
• →Measured peak MLX memory was 32.0–32.9% lower across the two orderings.
• →Ordinary-step throughput was 1.6–4.9% higher, but end-to-end timing was materially order-sensitive and was not accepted as a stable speedup.
• →The deterministic corpus and eight-step budget do not support a language-quality claim.

• ·The operator, backward pass, optimizer, checkpoint, and telemetry paths were healthy at the target 10M-class shape.
• ·Lower resource estimates were sufficient to authorize a real-data pilot, not to claim equivalent capability.

Engineering gate passed. The structured operator was cheaper and operational, but quality remained untested; proceed to a controlled WikiText-2 pilot.

• HXO-EXP-0005 result JSON ↓Hash-backed run IDs, configs, timing, memory, accounting, checkpoint hashes, and verification evidence.

HXO-EXP-0007 — three-pair WikiText-2 pilot. Open HXO-EXP-0007 →