Does GIFT's claimed 7.6% Llama-600M pretraining speedup (and its anisotropy-distortion diagnosis) hold up under peer review or independent replication?

Raised while promoting the capture behind claim-gift-2026-gradient-anisotropy-isotropic-transform. GIFT (arXiv:2607.07494) reports a 7.6% end-to-end pretraining-time reduction on Llama-600M across 64 NVIDIA GH200 Superchips, with better downstream-task preservation than direct Euclidean FP8. The number was carried into the vault under [unverified-quant] and the note held at seedling, because it is a self-reported figure from a not-yet-peer-reviewed preprint — exactly the kind of specific quantitative + mechanism claim the sourcing floor requires Tier 1–2 (and ideally independent) confirmation to make load-bearing.

What to establish before the figure goes evergreen:

1. Peer-review / venue status. Track whether arXiv:2607.07494 is accepted at a reviewed venue (MLSys, NeurIPS, ICLR, or similar), and whether review changed the headline number or the anisotropy framing.
2. Independent replication of the speedup. Does any party other than the authors reproduce a comparable end-to-end pretraining-time reduction from a near-isotropic pre-transform before FP8/NVFP4 quantization? A 7.6% end-to-end number bundles model quality, throughput, and cluster specifics (600M params, 64 GH200) — confirm which of those the transform actually moves.
3. The mechanism as stated. Verify from the paper body (not just the abstract) that "highly anisotropic gradients incur direction-dependent distortion" is demonstrated (e.g. an ablation isolating the isotropizing transform), not merely asserted.

Medium priority — the mechanism is a plausible and interesting diagnosis, but the specific speedup should not be quoted as settled fact until a reviewed or replicated source carries it.