Record: PR #1908 base + GPTQ module-damp + Asym Logit Rescale — val_bpb 1.06048 (3-seed mean)

[dexhunter]

Record candidate: val_bpb 1.06048 (3-seed mean, std 0.00074)

Extends the PR #1908 native base (SparseAttnGate + AWQ-lite int8 + LQER asym rank-4 group-64 + BOS-masked SmearGate) with GPTQ per-module damping: GPTQ_DAMP_EMBED=0.005, GPTQ_DAMP_MLP=0.02, GPTQ_DAMP_ATTN=0.01 replacing uniform damp_frac=0.01. Composes with PR #1945 Asymmetric Logit Rescale.

Results (8×H100 80GB SXM, 600s train / 600s eval, Phased TTT)

Seed	val_bpb	eval_time_s	artifact_bytes
314	1.06006259	404.2	15,868,104
42	1.06003198	413.3	15,869,714
7	1.06133559	442.4	15,865,773
Mean	1.06047672	419.97	15,867,864
Std	0.00074396	—	—

All gates pass: train ≤ 599.083s, eval ≤ 442.4s, artifact ≤ 15,869,714 B.

Compliance

• C1 strict-causal: flash-attention + cu_seqlens packed-doc, no future-token leak
• C2 full normalized softmax over SP8192 alphabet at every scored position
• C3 score-before-update: phased TTT scores each chunk before optimizer step; per-document LoRA reset
• C4 single L→R pass: each val token contributes exactly one BPB term in quantized_ttt_phased
• Section V: canonical BPB via sentencepiece piece table; full val shards; 16M-byte cap; no SLOT/PPM/n-gram cache/casefold/oracle/pre-quant TTT

Lineage

• PR Record: PR #1855 base + activation-aware GPTQ mixed precision - val_bpb 1.06081 (3-seed mean) #1908 base: SparseAttnGate, AWQ-lite int8, LQER asym rank-4 group-64, BOS-masked SmearGate
• PR Record: V22 = V21 + PR #1953 levers + EVAL_SEQ_LEN=2816 -- val_bpb 1.05877 (3-seed mean, all strict <600s) #1945: Asymmetric Logit Rescale (softcap_pos / softcap_neg, learned during pretrain)
• HEV20_07: GPTQ per-module damping (this submission's contribution)
• PR Record: SP8192 + CaseOps + GatedAttn + QuantGate + Loop4-5 + PhasedTTT + MLPClip12 — val_bpb 1.06453 (5-seed mean) #1769 (our prior merged): val_bpb=1.06453 (5-seed mean, banked 2026-04-22)

See records/track_10min_16mb/2026-04-30_HEV20_07_module_damp_3seed/ for full artifacts.

[dexhunter]

Withdrawing — submission was filed at 1.06048 (3-seed mean), which is below the current open clean leader PR #1953 (1.05855) and far below new open clean leader PR #2014 (1.05759). Doesn't clear the empirical merge floor. Apologies for the noise — this was filed in error during the deadline crunch when our preferred candidate (single-seed best 1.05767) couldn't complete 3-seed verification in time.