Record: GDN-Hybrid + Sliding Window Attention + compressed-code warmdown1000 (cold-cache, 1.01671 BPB)

[joshkmartinez]

Summary

Joshua-owned SAFE_SUBMISSION update for the GDN-Hybrid family. This supersedes upstream PR #1564 with a better clean-lane mean from the repaired compressed-code warmdown1000 bundle.

Authoritative result

• evaluation: quantized_bpb
• 3-seed mean: 1.01671233 BPB
• 3-seed std: 0.00134386 BPB
• best seed: 1.015700 BPB (seed 1337)
• artifact bytes max: 15,903,365
• legality lane: SAFE_SUBMISSION (all pulled artifacts under 16,000,000 bytes)

Per-seed authoritative results

Seed	Steps	EMA BPB	Quantized BPB	XSA BPB	Artifact bytes
42	2227	1.007164	1.016200	1.021202	15,733,879
1337	2242	1.007164	1.015700	1.020105	15,903,365
2024	2227	1.009032	1.018237	1.024111	15,713,422

Delta vs PR #1564

• PR Record: GDN-Hybrid + Sliding Window Attention (cold-cache, 1.01710 BPB) #1564 mean: 1.01710033 BPB
• this PR mean: 1.01671233 BPB
• improvement: -0.00038800 BPB
• PR Record: GDN-Hybrid + Sliding Window Attention (cold-cache, 1.01710 BPB) #1564 artifact bytes max: 15,981,262
• this PR artifact bytes max: 15,903,365

Notes

• same SAFE_SUBMISSION / fixed-predictor Track-A legality lane
• same GDN-Hybrid family, but with the compressed-code warmdown1000 repair that already proved itself on seed 1337 in run050-safe030
• final authority comes from pulled TensorPool artifacts, not live logs

[joshkmartinez]

Closing this stale submission after run055-reeval-gdn-bpbfix showed the advertised 1.01671233 BPB used non-canonical SentencePiece byte-accounting. A corrected clean restage of the strongest still-authoritative Joshua-owned artifact is now open as PR #1622 (run039-safe019, 1.01710033 BPB).