HXO-EXP-0008

Identity-initialized learned post-diagonals recover at least half of the pure TensorAxis loss gap while preserving the edge-hardware resource envelope.

Six sequential 2,048-step WikiText-2 runs from source commit 0582756: seed 31418 dense→pure→diagonal and seed 31419 diagonal→pure→dense. Hash-pinned artifacts made all non-MLP tensors byte-identical within each triad; pure and diagonal factors and initial logits also matched exactly. Final scoring covered all 217,645 validation targets.

• ◆MLP projection: dense / pure TensorAxis / identity-post-diagonal TensorAxis
• ◆Matched initialization seed: 31418 / 31419
• ◆Order-reversed triads, quality-gap recovery, resource savings, and throughput floor

• →Mean dense / pure / diagonal validation loss: 5.73598 / 5.83298 / 5.83290.
• →Recovery by seed: 0.366% and −0.317%; mean recovery was 0.025% versus the preregistered 50% requirement.
• →Diagonal retained 22.35% fewer physical parameters, 20.65% fewer estimated FLOPs per token, and 29.78% lower peak MLX memory than dense.
• →Diagonal compute throughput averaged 32,275 tok/s versus 36,023 for pure TensorAxis — 10.40% lower, failing the execution floor.

• ·One seed improved by only 0.00041 loss; the other worsened by 0.00026. The sign change and near-zero pooled effect reject a meaningful rescue.
• ·The learned scales stayed finite and close to identity, so failure was not caused by gate collapse or explosion.
• ·The matched pure-to-dense loss gap remained 0.0970, ruling out the preregistered initialization-confound explanation.

The identity-diagonal rescue failed its quality and throughput gates. Stop the tested one-stage TensorAxis branch; do not scale it or optimize it with custom Metal kernels.

• HXO-EXP-0008 result JSON ↓Complete matched-run evidence, recovery calculations, gate decisions, SHA-256 receipts, and limitations.

A future attempt requires a materially different mathematical hypothesis, a new registration, and a fair dense control.