Record: SP8192 + 3-layer recurrence + byte-PPM mixer — val_bpb 0.99621 (3-seed mean)

[remg1997]

Summary

• Adds a 10min/16MB record submission combining PR Record: SP8192 + 3-Layer Recurrence + Parallel Residuals + QK-Gain 5.25 + Legal TTT — val_bpb 1.0810 (3-seed mean) #1493's SP8192 3-layer recurrence stack with PR Record: SP4096 + byte-level PPM adaptive-λ mixture (strict-legal gate) — val_bpb 1.01252 (3-seed) #1795's corrected causal byte-PPM mixer.
• Reports val_bpb 0.99621 as a 3-seed mean, beating the current merged SOTA 1.06108 by 0.06487 BPB.
• Includes three seed logs plus a TTT-disabled compliance run showing the eval-time configuration used for the submitted artifact.

Notes

• Training is unchanged from PR Record: SP8192 + 3-Layer Recurrence + Parallel Residuals + QK-Gain 5.25 + Legal TTT — val_bpb 1.0810 (3-seed mean) #1493.
• The PPM mixer uses an outcome-independent gate: cf = max_count / total at the deepest populated byte context before looking up the observed byte.
• Auxiliary local reproduction files (ppm_inputs_seed42.npz, tokenizer) are available on request but not committed due to size.

Test plan

• Verified packed artifact decompresses to valid Python.
• Verified per-seed ppm_mixer val_bpb values match submission.json.
• Verified local numerical reproduction for dump seed 42: 0.99535348, diff 3.47e-09.

[cocohearts]

Leaderboard audit note (pre-cutoff state): I don't think this is valid as a record row. The logged artifact accounting is over the 16,000,000 byte cap once code is counted: 15,976,086 model bytes + 66,546 code bytes = 16,042,632 total bytes. The compliance evidence also appears to rely on a 4xH100/extended-wallclock projection rather than direct record-track 8xH100 timing. Independently, the byte-PPM mixer would still need the full-normalized-distribution audit before it could be considered.