Add non-record AR GPTQ XSA ROTQ Hadamard submission

[vermissa0ss]

Summary

Adds a non-record submission folder for a rotation-aware GPTQ variant on top of the public AR self-generated GPTQ +
XSA-all + BigramHash stack.

Folder added:

• records/track_non_record_16mb/2026-04-01_ar_gptq_xsa_rotq_hadamard/

What this submission is

This is a non-record submission.

It uses the same evaluation metric and artifact accounting as the main leaderboard, but it is not a main-track
record claim because the key run was trained on 1xH100 for 4800 seconds, not reproduced under the official
8xH100 SXM / 600 second budget.

Main idea

The new ingredient is modular rotation-aware GPTQ:

• choose a compact right-rotation for selected large MLP matrices before GPTQ
• quantize in the rotated basis
• invert the rotation after dequantization during roundtrip evaluation

The best result here used per-layer Hadamard right-rotations on mlp_up and mlp_down with block-size search over
{128, 256, 512}.

Results

Best result in this folder:

• exact sliding val_bpb: 1.11290586
• exact sliding val_loss: 1.87908996
• total artifact bytes: 15,826,148
• hardware / duration: 1xH100, 4800s, seed 314

Export-only ablations on the same checkpoint:

• base export: 1.11296252
• mlp_down Hadamard 256: 1.11291713
• mlp_down per-layer Hadamard: 1.11290938
• mlp_up + mlp_down per-layer Hadamard: 1.11290586

So the best rotation-aware export improves the same-checkpoint export path by -0.00005666 BPB while staying under
the 16MB cap.

Notes

• README is explicit that this is a non-record submission.
• Included logs are recovered copies of the exact auto-generated pod log content because the pod stopped before raw
  log sync completed.
• This PR only adds the new records folder.