diff --git a/records/track_10min_16mb/2026-04-17_SP8192_AdaptiveHessianClip/README.md b/records/track_10min_16mb/2026-04-17_SP8192_AdaptiveHessianClip/README.md
new file mode 100644
index 0000000000..3d0e05b0c3
--- /dev/null
+++ b/records/track_10min_16mb/2026-04-17_SP8192_AdaptiveHessianClip/README.md
@@ -0,0 +1,91 @@
+# Record: SP8192 + Adaptive Hessian-Sensitivity GPTQ Clipping
+
+**val_bpb = 1.0822** (3-seed mean, std 0.0009) | **~15.91 MB** | 8xH100 SXM
+
+## 3-Seed Results
+
+| Seed     | Sliding BPB | Artifact       |
+|----------|-------------|----------------|
+| 1337     | **1.0811**  | 15,906,928     |
+| 42       | **1.0826**  | 15,909,023     |
+| 999      | **1.0828**  | 15,911,535     |
+| **Mean** | **1.0822**  | **15,909,162** |
+| **Std**  | **0.0009**  |                |
+
+Merged SOTA (PR #1019): **1.1147 BPB**. Delta: **-0.0325 BPP**. Clears the 0.005-nat threshold.
+
+## Novel Contribution: Adaptive Hessian-Sensitivity GPTQ Clipping
+
+Standard GPTQ uses a global `clip_sigmas` parameter (e.g., 12.85) for all weight matrices. This submission replaces that with **per-tensor adaptive clipping** derived from Hessian sensitivity analysis.
+
+**Key insight:** Weight matrices with higher Hessian sensitivity (measured as `H_diag.mean() * row_variance`) suffer more from quantization error. These layers should use wider clipping windows (higher clip_sigmas) to preserve precision, while less-sensitive layers can tolerate tighter clipping for better compression.
+
+**Algorithm:**
+1. Collect full Hessian matrices from calibration data (same as baseline GPTQ)
+2. For each weight matrix, compute sensitivity: `sens = mean(diag(H)) * mean(var(W, dim=1))`
+3. Compute log-space raw clip_sigmas: `log_cs = -0.15 * log(sens)` (conservative exponent)
+4. Binary search for an additive offset in log-space such that the numel-weighted mean of `log(clamp(exp(log_cs + offset), 6.0, 24.0))` equals `log(12.85) + 0.012` (baseline budget + compression margin)
+5. Final per-tensor clip_sigmas: `clamp(exp(log_cs + offset), 6.0, 24.0)`
+
+**Result:** Clip_sigmas range from ~8.5 (early layers, high sensitivity) to ~19.0 (deep decoder layers, low sensitivity), while the numel-weighted geometric mean preserves the compression budget of the global baseline.
+
+Example per-tensor clip_sigmas from seed 1337:
+- `blocks.0.mlp.proj.weight`: 8.46 (high sensitivity, tight clipping)
+- `blocks.2.attn.proj.weight`: 14.05
+- `blocks.8.attn.proj.weight`: 19.02 (low sensitivity, wide clipping)
+
+## Other Techniques
+
+1. **SP8192 + GPTQ SDClip** -- int6 matrices, int8 embeddings, zero selective pruning (PR #1394 @clarkkev)
+2. **3-Layer Depth Recurrence** (layers 3,4,5, activate at frac=0.35) -- 17 virtual layers from 11 physical (PR #1331 @dexhunter, PR #1437 @dexhunter)
+3. **Parallel Residuals** (layers 7+) -- GPT-J style, attention and MLP read from same input (PR #1412 @Robby955, PR #1204 @msisovic)
+4. **QK-Gain 5.25** -- learnable per-head query scaling
+5. **Tuned Hyperparameters** -- WD=0.095, MLR=0.022, EMA=0.9965, warmdown=0.72 (PR #1445 @X-Abhishek-X)
+6. **zlib+base85 code wrapper** -- 61KB source compressed to ~19.8KB
+
+## Architecture
+
+11L x 512d x 8H / 4KV, MLP 4x, LeakyReLU(0.5)^2, Partial RoPE (16/64 dims), layerwise LN scale, tied embeddings, logit softcap=30.0. Depth recurrence: encoder [0,1,2,3,4,5,3,4] decoder [5,3,4,5,6,7,8,9,10] (loops layers 3-5, activated at step ~2038). Parallel residuals from layer 7: attention and MLP operate on same pre-residual input. Skip gates (sigmoid-gated U-Net connections).
+
+## Training
+
+MuonEq-R optimizer (row-normalized Muon, Newton-Schulz 5 steps), AdamW for embeddings/scalars. ~4612 steps in 588s on 8xH100 SXM. Linear warmdown to LR=0 over final 72% of training. EMA decay 0.9965.
+
+## Quantization
+
+Full-Hessian GPTQ with **adaptive per-tensor SDClip**: clip_sigmas derived from Hessian sensitivity (see above). int6 for attention/MLP matrices, int8 for token embeddings. Byte-shuffle + Brotli-11 compression.
+
+## Compliance
+
+- **Training under 600s:** ~588s on all seeds
+- **Artifact under 16MB:** All seeds under 15,912,000 bytes (15.91 MB)
+- **Eval under 600s:** Sliding window eval ~91s per seed
+- **No SLOT, no ETLB, no n-gram cache**
+- **Three seeds:** 1337, 42, 999
+
+## Reproduction
+
+```bash
+pip install brotli sentencepiece
+pip install flash_attn_3 --no-deps --find-links https://windreamer.github.io/flash-attention3-wheels/cu128_torch291/
+MATCHED_FINEWEB_REPO_ID=kevclark/parameter-golf python3 data/cached_challenge_fineweb.py --variant sp8192
+
+SEED=1337 torchrun --standalone --nproc_per_node=8 train_gpt.py
+```
+
+## Credits
+
+- **@clarkkev** -- SP8192 + GPTQ Embeddings + SDClip + MuonEq-R + depth recurrence (PR #1394)
+- **@dexhunter** -- 3-layer depth recurrence (PR #1331, #1437)
+- **@Robby955** -- Parallel residuals on SP8192 (PR #1412)
+- **@msisovic** -- Parallel residuals concept (PR #1204)
+- **@X-Abhishek-X** -- Hyperparameter tuning: WD=0.095, MLR=0.022, EMA=0.9965 (PR #1445)
+
+## Included Files
+
+- `README.md` (this file)
+- `submission.json`
+- `train_gpt.py` (zlib+base85 compressed wrapper, 19,846 bytes)
+- `train_seed1337.log`
+- `train_seed42.log`
+- `train_seed999.log`
diff --git a/records/track_10min_16mb/2026-04-17_SP8192_AdaptiveHessianClip/submission.json b/records/track_10min_16mb/2026-04-17_SP8192_AdaptiveHessianClip/submission.json
new file mode 100644
index 0000000000..cf829fd352
--- /dev/null
+++ b/records/track_10min_16mb/2026-04-17_SP8192_AdaptiveHessianClip/submission.json
@@ -0,0 +1,35 @@
+{
+  "author": "chris-colinsky",
+  "github_id": "chris-colinsky",
+  "name": "SP8192 + Adaptive Hessian-Sensitivity GPTQ Clipping + 3-Layer Recurrence + Parallel Residuals + QK-Gain 5.25",
+  "date": "2026-04-17",
+  "track": "10min_16mb",
+  "val_bpb": 1.08218,
+  "val_bpb_std": 0.00090,
+  "seeds": [1337, 42, 999],
+  "seed_results": {
+    "1337": {"val_bpb": 1.08114, "artifact_bytes": 15906928},
+    "42": {"val_bpb": 1.08265, "artifact_bytes": 15909023},
+    "999": {"val_bpb": 1.08275, "artifact_bytes": 15911535}
+  },
+  "hardware": "8xH100 80GB SXM",
+  "pytorch_version": "2.9.1+cu128",
+  "technique_summary": "SP8192 + Hessian-Sensitivity Adaptive GPTQ Clipping (per-tensor clip_sigmas from H_diag * row_var, exponent -0.15, binary-search offset) + 3-Layer Depth Recurrence (L3-5) + Parallel Residuals (L7+) + QK-Gain 5.25 + EMA 0.9965 + WD 0.095 + Sliding Window Eval + GPTQ SDClip + Brotli",
+  "compliance": {
+    "train_under_600s": true,
+    "artifact_under_16mb": true,
+    "eval_under_600s": true,
+    "no_slot": true,
+    "no_pre_quant_ttt": true,
+    "no_etlb": true,
+    "no_ngram_cache": true,
+    "three_seeds": true
+  },
+  "attribution": {
+    "sp8192_gptq_sdclip": "@clarkkev (PR #1394)",
+    "depth_recurrence": "@dexhunter (PR #1331, #1437)",
+    "parallel_residuals": "@Robby955 (PR #1412), @msisovic (PR #1204)",
+    "hyperparameter_tuning": "@X-Abhishek-X (PR #1445)",
+    "adaptive_hessian_clip": "Novel contribution (this submission)"
+  }
+}
diff --git a/records/track_10min_16mb/2026-04-17_SP8192_AdaptiveHessianClip/train_gpt.py b/records/track_10min_16mb/2026-04-17_SP8192_AdaptiveHessianClip/train_gpt.py
new file mode 100644
index 0000000000..4bfa51a5aa
--- /dev/null
+++ b/records/track_10min_16mb/2026-04-17_SP8192_AdaptiveHessianClip/train_gpt.py
@@ -0,0 +1 @@
+import zlib,base64;exec(zlib.decompress(base64.b85decode(b'c-qZ<+j8SZvf#VE0>fFK2GkIA>5JQ3(=qIp+}&Z_Mv}*#!%-k05)@&P02=^xv5q!ke_}syzhpDB?u7zLwLKfh5pID%RaRD3R#s+KR#rBf7iH}Yi|I5S)>)BP8<}`De-O{t)8a}zXN7p0CG|u+7uCkNEN0#u-c7SBkH0?Eij+w{DrWLmDxRtb*=C*1<m<&E8*PXm`C>MI@RG{Q=dxOw*J(aX=UF;T@l!RI!0VzMPK+PXD68r+yIR!g2n*v+vp}9l<3&z1NTyiiN7f_xb)J+7P-!aqwyz-WCv|cvK10?0{Zp;7sBn?yRZ)gso^Nc#Kc64OKlWdFJ@3b4TBRFx`Ow|);6JK2o+i~KPU<?3Gw5bHPN0t*jP?x?#`d&VujnN~(%QHv7wJZN57er?&jjY6EQ^x0I8Ul-W8<&h)|&r8(N7Qav_yi;QoycyPUVduAY0iG{D9%|SMT5+J^&x&IK*|G<<}K`83B#sQC324iYiL;+pH||=sK<a7yJE-e!Tzr>_zCkh_<od_KScjRVfU2&jT9zcBS)!gMFy5`|8!V0V`ZCfZC%b{Lb0W$MNU=PUyja`^c9O{`1-opd!cmqD-<JaP_8Xehoa2Z!_pvLUXDGh0hlUXZ?%MC&%Z2$leb8$1%A}%Gs#6g8?p+Ayghui=<iom;Tw&{>d-L@sDTy4*+2&di(ZU4Z>m`S9LmXYaOfqd>Wr$9GtTDdDR-b5?>|ta1z(WP1=h7#aaLJas0l2@!?Z^aq`mv@5bwQ-lj+YyfYwH>6aL&vlQ6*!QbP<gJTZIs6H*O<McKyAC>|*Joy+O{Nv#4p8(>Gq~|QTkMCf14yVQN1`rt*`KW58>rwyL_?Q0S;fKSM4?h7)A5M<<v1@O4cA_0M<o`&f*@(t>``{MlFujI{bx8M*{$UH@UhQdwOA*|1x*Z@WC)u#>w5Y13RKwbbC+FvqXs;C)R?t6S>G&=K!F?B}d2%(S>2y^TQ{OqK4nOaIKK>a0^7$ATGd?)(zdxi|_G0&iHd&5}5z0vrjW9)q^#axcLbw<vS8<j7p0*)-@}d7eKL7lm2LSTj?st0}QCJ<~Y4QMKq^08V&qwiL|DOkENa)>NUhTe0pjuVOZM6OMybsmRFXCgp+6=J=Ni=Jzb#$_Sa0ue)2w8Af;F!*?(h-rfz2?Eu`-6QVpj{h#cyk+1(qy!(k3an*{&dhMW__*ITvbif5u7ud&g0o)TDQ-T!_)Za=fewZCz9-Plg;Dnq)w}~SmmeBr}4-B#ld-#O=!|*X)I}0ErWFN`5=~*J2#<#(8%hzD#rCNnYVTagvRHK`26I@ix2%%0Pt!@cB(AqDNN`pZLNHEa(V!B`5bw27iG9v7e>2{^i&f@nwavfUTGbHNwt?q%_#sFCew5o=?<aphyLM#NwhCFFpL>$9$7mJ{QV~wJ!t#$@n^*be10F|Tm;$|a~L}{L*e(*33_tE79u%z1ukGoS+`We3b1_j&n`56r&?(~T2|-a7@6&jn3=?^F-I4)sMbC+MkPdXlJ+lX(S4&jE_J{%t+LTVEDAz%WW8acss+E({u!)1AU~*IOlV;%L4fdgdZ4$`)C#?Cdv!72+Vw{1IDxq;P&k=`5)C4>WHT8T&a!<T<BCb}CDYQJTULjrny4kX7A4HGN+@AuOFQq>93RCi8-)429~-UiylMhKW(E3=)&YY&3k2Th=)93y0_gMiOX=0WIQ#soL_=^Y&a<RkRuN}e|Ey7QwkYy=R-m%Gn6<b1^T~01baHfXeDU)U0r>7a9l%vG1XW|SRM{gK@B1G>+1QsV-y0Viu~Z$|%v%GeSgMk2E&_sk%R#R|v<a9YysN2n2)G2bBrnPtXkx!F6?6pk3{;Zile44#A=4l<k$08W$!=?B-ydA`cUix*@_S3m@3HcN#K|a`#jMxvI{p3rQOtXpZiyLkUBbK?PGMoiZnjqbcm~t!!y&Aw*gtCdm7zicV~^6M8YD3GJ`&~aYlF3x_LDJHOK0y!%lrDvzGE&dsYH#5O7Gqn;%Z4XHsc-jz?e*j$-|P0U{fA^=>L<?IJB&Ukp252$?NQQQvSXF@IDl?IY|5peo=uT|6RC@)ek2}r)LM}=O<*YUX?{X&A7FCJ+HsSAWO0<W;ZfRB5j*0A5Sm-9)n={{GM5n%%V6r*GG{+0xoHJo31o`2q55gaQ2S_&F<OVi*`&Ad6m_z^OuRp_n$8a4`Y~?RAa-$YrQjDBmS?l>seB*lFT#*&Od)V>T|>`Mp*|GL&b!&XIOrk=LdZhi#c;pXm5L7=d6GHQwRF0F@TqVUefvH<m_;ttQhMv)BlqE#!VvzG4wh+9bx1b<#e>fVA7j{db=H|BLS1eOIiZM7i2=Cy{w97s2I-6VhHRFl^Kklx1*$0G8}@W7b5<hx4o_5DJhuL3Hb_@ZwCljDgO?Bi+(Gz+!uZym@J|g49NJ!ILp(!^lEqKeO%2SmF@WSLiZYsvnkD=21wByfaQ?5*1Z{BW%&zbLO%xrme=c(4*-?17^^PnnJj4ZE!MjgU(M5qY|j^x5j9^Kp($)0{DJ)4f#+)bkvX=WqWZq>XeKPjsROM6-t;b-^AcZ}go7#5UKTj#EmZiu^GxUA$7ykuOufV6`WiiEKXR|@vz&<N4!Q=qPRNCb<8xf(z+M}m4&$7p@2Jl60{*0HyBq^=>j$dEJdf85D+@4lq&6Cx@Rs0%KdY`o5czpkOw%r@Al%_<H9(GC81odp+`^eFelvDK(Be&$7k7RTflL|WPrvicr|vgL-EYo2T<NTn*_=epxHEXHtEWrv5m0$z<Cl#aSZ9@n44EY*&V3FikgrRSXl334C9-6If(B&;5#+litvEmn#sx(4qPCDFd6xkbf_<F(Emc8SB_OfVd|05{px0T{<E?j{fPCa8tOjZ!O`+a6cqTWKC_5|?Z15}m;L3}LfEbR*SuwL{V!ukK<IroeZwSS1vmvk?-{7EMf$Akzj4JecRP$MM&bLcWal7P{4PH_DY<5%6Cy^<xjIkm@F@TX0N<NsSe}UPcy)vl`617UPkBgKx9Q?)@j62G!*L&0<J$a)dWvnnmJxEDY%~RNWY{iVv1o2;B5(Y)JbkTwyFhb$YM}mLwCy`tLU<4xF8*4#B4^e#efWEKkq7K>1Ndmu8VhmWKl3}WSgK2bCEb>uOKCssC8(?A*=xUMS!m00cVYy0lI5bEvID!;a<e}E;On}V0OE@Fch}^~Dck|l@2{}-Krs+WJZ<&8DyS%ly;or4KOzfphQt$ogdk>V9;bO{^Y?vyG8E`}A|Nfu<-9ZU~s$_AOGp6vW!mB4q?QwE%5xG4Y5DqjcQ;<)K9JRAyQI_ejo<2n0d765Yx}I0v?d@w2FN>=PwD)ZkGD)^AXWQ0wF&%GbfJ9ntzx(c6X696&R@cz8h>^A~V3qYd|61(5c?Y@{&J6gr$Zzm<5PJWTW81lqtUOy(wKqv_Q||?I?S<zP>LjtuE{?&P;D{4-uyd?O&|${vH94|4foI9RZxCG{QEtU5uFX+S<nstnDk|SskU~IwRD<a0!B-fx-M8H1(_rrPK;ZfqOLfwk0|Ep{{<R<=O<2urgt|1Sqt~g~>%dr{U_78zSf}c;nEIk>-~k-=2}TRn-Q4|YQKcjG#gudiGs_pLYR%ERfv9{b_kA14;drz)Vhor8Y<D>_Y?GY$F&iPH(hJp$js|^-$o*5{imN!<M8?Gd22Ql;Y^=8&3;H|Q?OtxI?ik2gpJ7C+I_N=sP3P`%QdTsv$%Io)qQ^JD3Kn7|h?Rv*>v6#uy1}@5YC=PvToag!FV+PXNVU_sUcssjvMTT5J~gg0v{I4qCMnJ2F;_;pRWrZ{4SyccRGQ-)f+8RzeN=}d{6|0yf=i(j;*v}N6hp9u{zrIuT+}~e|JZsz?i?4qu`#}7?Fc*{`O{NJm;LEIsmAkco=(9K^6ttEMi2yi3XtwWevh-Vs<+<k+-!}Cp{MA8ia!wBGtDt&iD8EsiHBhK4~U9ngjJKI9QlYQBLq`JbB6e35Yc?PUKESUXNWGy3y<qHn?qHU!}2usfzIA;fHtqt6TD6FD5%Hleh*f&O+QUh0=-D<*L<4Qxa!qK;Z+lg0s%CQ-JjkgZ$}W5PYs6Opex#a>Fr+XY-ACjOEEhQenAI6=!G<wpxb+I-a^Mf7!%+6o$s^Pogji?nvs<+DK0B$MHhQ^Q3jlqLMsXJ@R>f(<^{nI*=pB=wgtDGf9jv@d$_r9aOQnJzBu>(gT0H7`X4`k5EN(dBU%{$!jK2*p$5Xf40bMQwcgv=d%g2!Z^!z&%f9X|9j%{pU|U$5(Ite$6rI9xdc?knVZp61BQ}7yvUkZHRUC|1#|_HA)T?jmG!}iurcij-WIM*9&j!?`9XTiLsopHBnC7iI(Up+AM+4>))3!RA<3ni`ErSZ2AO=-+EA)!-7$$O0?|2}(CjraRR5e$0i$$F!qvnvvnS!DLag8b7%vX&i6)`V7E%?HL;c-1$!b#nZ8QU>djNkwB;^2JsG~tzwSqw=0q5t92LDw5)!+Jmx9Gb!<@bPIQK04~3HVXjcSvs2~b6$*1>lk|(E4tY88o)AA-K6OL#-Jt=eav^A3@7AOa3M$DL8v8%gO_$%?3}8s$RC;K|Ltb5d*ZqXHUMJb*yK8&<zk$GnoRP>Y!vN*lEd3RTr=Dh(zv5BLQr7Y=*(uaOHQ`YT9MfBAw$<euTyrc`OIgpV?u;TS19P`A=lnC#08}|LJB?MC^?@j#^Y%^I#0hW&|F2AOIj{-n7pt&CNI`=nHXO6iT2a-tjOpLQ0?@14YL?)1O0l;UKSsdsd6o^p@fMxsESa<G<8gZbNOGMt(F<B3AUGA5@^dd2a^al0=iwDT9;;}GQVa##KM51k7<t4k${VOclzPu+t)q=7FdlM%jTfTt6Xs-;z=Ye99Z~8mXE5gH_fVgK>whWMa~-xE(u251_M7<nX)s{G*RUO%RHhlwDra(xnE`6anFUc_Jvrs%9O5dPF9RxoKIL3(k$bx3I+s?GbC$&C$vd5sIp7=ZWq^RE1*g*8jKU#M(H&M<*IgHVl9#<^3@OCj+esv$2Wh3Zd393iR0wE@!5s{2?p1c6?WQqF(`3SvL!<z2UB{ss7WvOjrM}J!KPLL8PO99_NIYB5Y+{;g|CejEDSZ<i{&H5_6mgB!MtPk=9d~vmv`(^4K5Q5in$HoNFd)9OPh_x0vj5OLFGcsY@v<hWjf<A%rsAvvP>SpNJ1uOldF;wL(%q9bIj0o>f6@}mwRb%rL%ec;QI~aLX-9l6J3Y)@TWq<4qa)JxdF_Y&UH~IxUnOt6jh6v9~fmwKgD9EZTEHZW4Ro$Oa1wS_neN;%i^li%Xr(Yo+%MCThI(A1rXgPMCi?X1TipMk%$X%U{3L9Ud%1LBtA^=*lseWJT<g1g$h;+tSu;vg_fL{R55q!H0UNrVKy5`M7ydFFHNzGX1VEUl`$N0It{K!kTDzv-L2hA<NJg8UH7kXb?P33o3E)QU_oG%2}Z@^b{TD2hTg*(rJlJ-#v;D^dc^nhx600r&X0?7=I40?@D@`sMIG87$zNh=@4sj*<r~DhuD5`d1dL=^(^@}R`~v)|)&ZDDcWSx5?{Zn*aF%Ek|1m0O6$Nkk_o3(ClL_N*?FQ)52H-u>Rv;DK4`6Jjqr)sulM?&Np0vI`2aP*1SMJj6dQ#)a-ZN`db@wWxXc)GU1<!gR@=J>kx+TGue%f?zba~2fhYsI`1YLFR45UK&u+9*UvKe1l(TgiavnOJ#L-SM#6{W9xuR|4iEne?DW8hhQKK$F<;=>Tbq_PYFJ6L;(F`!3;t7jmw`XRL{9f2&rjmKsB1v?n+fN&+xwzs+I`CJaM6-8)(5?tu*nX4dC@zowEURFt4-!9W@(1Ota0%jM`t0Qpegc2PP)Sx9|F)-JeFT?~Qs1Qr9!j4s7SXBI@U}JN_+VHY`dGVG>qH~Gc3gl#xFB$i**4|+7e23nz`X%Bnj6%2|TOC?a&LkvmEm{m_6j#j)1N1Kq3)Ip#O1Z2`p0Jf0!Q($T7{95!h|>q6*0!TobJa}m#BmEX>wwJ}Fl2bZbDRF%!jl3`eWf{9B+SNqaVlIeiONyi#Ig5f0Dd{Lu)yitU4|En0N}q-+Hzo~!l{%DzOBhfW?>+(l;AAk&LWS}FKX|jSirRQYc-97<^0=Y^&@!Y2Loi$(CdaCf4Fp1)|gsUhF9xWwnawMPN8L{!RSawA_b#B*g(cE!QF7ld_H}MOIn!lFo7>uRP;nv>IXnlKO9Spz5X`jI);C$H{W|o8&K?WA7cqv)?+-D3vllTQ4|5KWJORD`8$P!BIG2Qj*Zm?%)LDYeHSWYKsRfKgT^`tReOi8G^p&8v3=&>@500c0)gs!kB1h4JOBe<`djcEi1$z=pyj*=9koYMhiCNMs=~XaD5m<Q`eb#d(Gy3TU)2YtsV%BxdJb&XuW=0|m&9{48+kZS`z?<7u~$tAB~PSrl_)gMYsr^Q?Pxs}2$&WXpl`H;gMVpQvkmWgM;(f!fR-!NUv9H1!yU(~2L-HSw4~bj#ysM<PLrx~Ym)+A$KgyAha%;3vxs)$QvYcdq@zz#XGFGaLKI50Ec=bOw^koz1<}RymWKkJ6{iPUw*WVWLlgN;vuzXmVT=osVRAqj;!&o$JN3F@eB-Reuvl90_UkIn%i_1MKoqM*F-9bb(~L$?3YGanvj@jGk50vA0oN18i|N$&<$$8mrq4FFZXL8M{X6z2Ty)?)H;ascVunvo&Ub50_l#dwn~N%q@2f;c87WbOhn~~%I2&du_r!p~ua>pjrPkzoP=G`4Q|Rpj<cEm1bL7WQH#oMpTte?DjIPUKfg8b}aB@H?<<X0;2~L?3B%g!7xeZwzXh%W0#Uc7JlE=OLTPcvEMP7Ybq^Z>|=NTWH*@XH5d9GL;V7WA%<1>rY#_#spL>0FPOkm?ksC|`1epUTW_6SquXwKeSl$1Im%ufGao8;>KjsE^ty}#ApBb+Z1bE9EX42GtNFk8G?0YM0wRv?Tm{vePqH2Tt+0=B3BMRTT@C~F08px}+OAgt-6(<!ru&q0RTy)+CPIyK_b&>Fvl5~6K|o5(e4-mFuT#!|>~CbE=1D=_2=wxMQU8V>YErWnK<5P-=vByEPf_w1F63lwv`dJv?uq{;)pGyV1t?%_8^Sl8=u&iM~Y(>gjleV%3|QW!i#^O<C$kvwfLQy^@bNHX8>(76;b0MA^+M5V7<$Nq?CQqlRxXqqNBz>?{7!JXej5F?<Drq@ZG;laTvFk}ypx<?>rlahk9#n`+@$MB!6_nHIvXR5&tSacvi&PMmrmJ4P21bL!|4a@p&M4Ds#)UNPcTSc&s%S#{-;+UqPVRVDIM-%6!NYzvbQ+rvaLI85mGEO`KF2=*vOX2BpC{Ps1%@noOni970XThQj4Dl?xcf+dc1en$W3@p!Hun=~969lXQ--a*vBOWKinhyFhZxbd(gGo}>zT_&}JV!|?Q*w}l7s{`k!0@=UbTYFJnmtzQLc`U6AJcjVg$6s9=4x>Xq7G5TZ@aE<N_V9}FX+6AWCGu0dLLT~#Cl~BOUH+SA-rDNETwe>zB|`A#7v;$fQo-#TFj+Gh3~CkMJ<PWS4>==Agu1D*nT4{rQrN{da;h!bevpV*OH$^W&#}CmyT8$opfqct0YX%c~ND!8h~<M)EFaV6l+8YH4^wttXGZVj#M5K*L!OsS}A@kysTD=8wAE=!ht}=H&0PvfPt2*$)c#4$cawBN#}<n0*ImRT=`8<p$^Ec3Yb0zH%tvgqX8Gprn9<<c4J$jfkbnw23H_<Vxv^Kxh9d8s>^Up$~vQTS)Hvm;MQIpsw5Q~Wj4JIF{Ec;bb{$E7R9tRbYdAMlSr(A_^j4c$0ougGS^uvLNHVsM=w*VMsFe|h~>Y>XjT<BqEkc_ytMe#a6MZmca+OY>pl*MMt}weL{u3?!bsfQ=Aic^!wtvBq-9sBQN?tKSiTUQ?3OpAL1NinZ{o72c#sF4W?!y?O&L8nlQY^{vW=GWRdZXHFs_LxEG^Mq>nUix)DqtoA*!XgYjA<ZRXl>wnr=FX!2IfnSJF6RkX^DNRjw#yvb3cM(}X%t#x=wh$JD{i9;(b@jSjK|Vk!EVD56iy9+g-u^!Tro0&;e(jXVr+&GJ#kf*D(21kJe`H|8oKhz&GprZp0!_n_U6e1Ui?Hlm<`4ex_BNG1#}IQh9c4Ps!5W3>)&Da(uTxLJNML90LLa?p+~Jt5Lp!r&mZeW$k7$;8SWoR$?L2v%aY7Mxc4%JP&5k;U6oJYZKwxLKIPl#-t;WNgS5)HeIbiG`(GtYJQ>WRC8(n1%G!jj_4w$QY@jiZnsdl2%W7N8I`KCe9DK_}KBjUjpHKW29X|ILXs^mX_Bo1}hH@+=<?-qInCS8x5FLVSYDKi8<Z%h-LEwmUTpxixtOQ%}g|)R|tx`-I+-uv}hhv7J9QZ$$J!&1=@M9O`SJmZ*!m&ogrDUqn3=7;3Z|8ftNBRw4Q~c%5oT?e3<&I5GK<$1LUk4*J%x_P$`Rd)PmUw8$KOZ&xzAx4uE&L8Fi9QF%ZZo6y>C1^}IdW_&2|iGZbua?}wiE_B9r8BH=l!?ec0XFfYn_Qe0y?a|?a2;6PRC?M6GzJ&c&Fr;6L2Sw3IXpr=+#oe<Im!5PDs)y!D%CBcwQJxH@aT2&$0uqv$Yf86h&6F^NF5I17=;c;Ae>_b~^@?nfgyCqa>nUN}*p`9?9wS4W-!t>20>=o#%n+RLYu9t1bl{q0v6%-U`A+uENaQyCftQ8f>8zu8OY9`t!NZC4{Jjggxr@)n+S7NUbar)hN-{~n9WKw=xIYbyZ{P=g*wooY`9v2_5z(UdU#S9N5;Yk2SgQoZqRm7^%SnP40GcI#S!;Pr83}4r(4g~DWruWX_nj%s+<K6Z7qpjer^GEK`%E%tfjaOtN-vY4=JsUq~3+gDyG%e@8Mh##jkVh_wTD4${u+USf0ck80C?HM7!?8gsXiU=tFSzEOQpN~hJH2+6-4$hz4o}^sOdf5d#GplO2|~>1wx?gC7;^(NL)zg7rf#_+1En{2sau#)Wi8jR!$qvjdtW`WRe(2*4pjE3)q`pekU;vS2Htk>@-plDjU_dwb4CybTY7tG#9NVdg3(EUzGMbEXk{k$_oddd)hQ~_+9lrP8sVs`W%@GKl#xbCRIkBp*yc>*Cm#<2&I0~sx2aDk|BSw%7|~Q~DhRTy3(ztJn5jd;&#9knaHzt|6nYbo?!?hi(l=*K&-SKBBa`G4UkA4BDBT*H(opC^4;ElN0<}uXV@=ADF||d9!Y5CI3`@F%0#h!{AAAQ*WNZM2<?vTRGnCtJI-XhucweTY1tpN|b<hy$h`3(bWXZ-4#KJQlZ41hbL%oB)GDx?>2a(dTVlj-P@9@EqFB==^9by@oZn1;2HCj>}K5K=^H)gcxci1++dP(VGIape{*!!Rf8=-ad7}riwPtx+anK^m#_~hB?IRD0DYKz%Co2Ep5w43I|9cY`kSrX^zU0vkWaI%>G{>J~P1y#$lvUGaKIJ(utRM_7(xG@R6tI$Iymj5bx{rWXn*ITcnZ@+y5wfCZ(SG#W*7yb$*K1NprYxk`$UT*iiU&)^JF}xQKc=2?xc{ML83OWDC_-`lheh|;QDg}uwsm1}D`Zc=PaCiXO`CL`Bk43$|$>OvXfolZJ@ID|kq`!aj`Uda>-u?<--(yXEpaG1D6jPIbRiQJBA{P!K9NtLY5YWEm$g*m}WGrPMg~;J9Ul>{E+afFS6weS9w?<yn&|}h)CE!+rjUEJ7pof;@ef?+*^L#4nj}<a8to>=(o0h6`J@H_5uV?&P($AiFF#Fikf2`_b&-kUK5#vIg&$YaW@f0{`Yf5_y$ngOjPK#<$rakM-BVGq&+d`Pdg_T~Bc$J^MX18UqrX312Yk@=5%o8&^HvYn7cgby%O%Wi>MU-~pJq}_<{U2#CjLaNV6c+^b)(=?tx@w69T4-5cDCM=RioaVb?HI-)3t<&*Dq^RY-dWm$Ob|>T2Oajib7|}RNM`n}Ny0|*KB8->fKE#bpmu`<FCyMp1?`(6cxmIcRYfqgDlOhv1?jwZED3<2wirk^7n#wt>?W*`sT(flBXR_fr-?fG#JtQpWO@3D{RDHxJijMW=y9QMwlQ3kr4fWJeIFr1OgRBKicX(m=6z$ZB2R;k{5y|uVytIyv$F0p|25#CopP_!!Q5?Eqb?z@*I-9ArLI5~yDff2-pHU)`_iO-Jk4%W-(jD#{Ip}rNaP08RwbG($Ro*$2T?K_#ceI%ToG$XYfq4x*9gW3=%JTP=aWQ=@>aOS&Me=(A#qpug>`vNb4X~YOs8;$mq)M^&*m}i1c64g^KG!aD|o0-w18&Sb(`rfYAn;WuDOn7u3>GkHd*3W=)kqqfwsjnrhwVp#4C@Swe=}Hpt;s%Q|_fFrNAXoOw-?+;*V<dR25rxYNEP1LL}O8Pm!#G)Yowey<*;@0%!E`WFDRWd}QT!7zshgI745!JM5mM^+bv$=$t|X8v3Cbj3w8R2-7mMiGp#3TetvhLZv3XaQxSzd~}<ocX3kr^FZ0^^ETF+^Jyuj)7G?n&RB8n6x9LHL_eGyU!0vB#uo?2=O<_JasTKbKJ8yz9Go3vnmjsU?enX$C<m2Pr1@|2q6``ul}vdV^o~-WI?6{Bs>dpnP8Ri4h0=+pp5QHuyAL(TZD{O#Du-#P93vcsAfhauekbfmQ$qAs1yf7qTyB4@nGBv%-O{#e7Xu45;YALfm~E}e<(c-mRUPgHinLPvTGl{ILPfW+NDyW;)x0qS*bX1fgC+|0q@G*5Wt7t_fAF>OCzIIZiB>Uc{kM!mv<RvV-C3WM|K6^C4?Tsde?}*t+dV)+8+x6CWg6Q?FpWVHvCGo}C0N!{V@ra@+ga)lVnZ<6aB=T3J+Bo7BiPWAJySOS(vDu8mhG`)JX{RQ=wLMddlZ&L*BfXEnBK?B&_mmxD^Y;oRM>d=WF^Hbgv7!8U@A!fxeEHDWcG`1srdpg(~iZP!m6GRsyAMxhA|*?cUWMSLOo;>iObN;Z-&7vC~`M43ym=cJ1JGVUEI<kyiyF$Ff8oj5{4)XEkg_1mNx;0PCE-r==a)%?M9X;y2OEC^_$;=W|lD>1iX;((nXG@<!aJqi6JX3L&iRrEL2pGrH$Zh!mMT@ou7?KCOCK!$vYJdCf43RUH+RHkl$kW_54ROw>if`*HDCRQntzL^@2ndG#<ff4@)xvN}Hq+@urrfnMC~NPivI;TYHSwAE#Ag)ixI9csRx+%5uE511%f}jaEivHAre^9f~P(fGcWEKRtf*ErhhK8VXHfsU4@&wx}(f)KI%QrL7qZ?;{D&$LxW10pcQ0MFO{m+c5aLMyS3u2h5&4H(Q*3z2dXBA+S53uLGN>@p%4fk4qJw>&5D@jde2@^j1ZpF@|>pp}7Ol^tf-!Wx#{3xPM^7(iRal)m~HcHginTmuG5@%N0ZBab<AH2*krIMu4Kjw@G%X*VBH;?(to>VmeKSHB67H%95PlHDnJ!s9n)FWr>_y5<Ih#4vS|)WjnPpd9jSvu8TCqKGWORA-lLprBXK%dkN&b%>pdZIu$AkI%6`pRC8&T+@vvn(6$@(Os`|A6_0btJ{jKrfrrSN8`toO`8VriIPn96k2Kb}m!{Jz_cs5%vBAod4Ggw+FB_nUhq(7p(x3rTd>LSI%U$SBsdX;{7enKDb%L!1xL~6l6?9q5X0d-CT|hexX9Oqx#ztEVsy-9HZvHzWJ(Ugf_RX7DZ>?K{$l2}KsrFua9S)(>7$&R*==!v!9P-^?hr5zERM;&XE1B0%?BEl(ntW?CF0MVr+|SJDXo;2}@^WJKcv&q(W88dY$M&%TD(J(QID{fh6bv+Iong<MWnzZVG0<hxY@Mn0%<1;5Dfe~L?oXb2&FQ!P6m(2OCzonWw2W1CKIoWMu;N(8PIzk{A5Co=Q|_LcI*+z*?_5A^DIG$Mp0SQaTC#9)CXFSP7T23hqW0|F3T-RE<~~qj3wurMf`l{YAJA-8+@^kD3OP4(H5keo%He$BqxauQi48pnK~&^A09`sQZjj|pA;zJ1GVp_belmY*w==7#fL9_afWFe$$YuN$`@(!~@}@$J8`*moZmde*`Ab}s8M@E!<QC3w3c`qU3z*jxW@ee<BG1Zdd3ru!LZe1Kt_Wzx)+kG^{ZE0qB>B?-zlT&x-P{I7&YXonVDh<7tXdFZKoduYy~AjJpKqCA9o{4+#n<vhUgP}{zJP*lRYRCvDtj|eNa?JGF9S{>UV!y}(4~a0vrB_atPp;~9u26#rQ8xBGO?GkIz#yrw678hP6}9TZysWJ3Cdl{16HQE6kN-%K*yk2w^0Z_hTa0?NvWGU=q+J~%EegLBigOwdr)*=fqp;urmV89@t(Dpx4T1|DViJ=o6Z$&jVkRJ-~HW^+KR;*x$55u>fROQ4Iz!RGbm3x*Skt$?Fdk6KIq!Z9zTng$@o%!8Wk%k5%fzeVvbqcIBeN&H`_CZT?~UPjXwrdP?VSD4$9KmaX@Jc%9;+V-&BXhcPkhr9#RMpM-=h<&D?7qaddcr-x>wJa1fmNa6tJGmUtVA&zjhVF@(vpr3OFL8Nq<~cSa_l>~F~Y)PQ0PHJ;MFKqTP}ZSnG)S1`bTjDpt;4ge~$zG7PP(hN7&Y0;%j_fSGhi$+n)7D>97S^+EWcMptAWQ(rVie!jA)(Vd7-^`(hFUH!yD8pn8=pDuyfxBuWxl;HWGAtoRyUSsHg{#c7c5_i%4VtwyK*iN>V834SH3C#QN<y^64ul4hZW92%kxO$oSE-3I4Yv-66(fVeeGQmtxUl1eGSY)+#x57r)f$SXp^{3)rIo#TtDzR0L{V6TAf7^+7Z}PzccChS=KCJAXee6N4xwFfIqIhvmnL(Ow6qVAc1M%98k3@EzGbdvxd`8~L|h=y`MOEO@EbdrI1W<h!Dt7Djai9{e}EZc6zZ#;1{k7)QpV*LxKdW|=B_2<O}FVN&e$*`sd64frMHInQ7=+sR~Ga`kVly2|1s}+TRYM28$1VqS%~m*dVc!{`j^SVRqY8*c~b3CfHl#C*Oo4SZC%8mSyWT@G!)`ujN8%~{)PHmdvq8S|83K*8;2KX8A@%R!Gb0jFqDGksr0*<gD!*oQtvd5>087AnP%@bJ>j?Xe(rO6VGTUcG74Nd#MQ<wV$7=q;!=}1w+-?;bjmUExk4g2S)jEw4rWLy>c!cmYAXq(>-$dD(pl4(0xQWyWnXecq=`kapbS4%32d$%sr+vzsjS3DB-*tJ+|c7Fu-XZp+!f|A8(an?unwlMt|y>PTu(e-!0Rj*idXD>h1mL}pU&e6e9xGVFqYde2A8H7UO&0H=@S&Xly*xCgmr<nxI{-qoBm5SNQNx(mvI-x=&;duN)Z!dU?JfF7<cgPBUXLtdXK}Z8|{tt>{4Gs($(9IOvz87WzmMZh1nZcx;oc$X?7ZmkWcIjZG?Vl{Z7F{R7TOs;EQnw(e^Mh_7O+s33y_Ii*dj6@$};FUNTDNbmmvLMf^{!BU?C~F$-5#KLp*5vIFpV%>$04oRRtp;nQa^XJH0jg9SvX9k{z54EqWu-8g~nP+&Yxa%&QEuOs%SId|Em#CQW#R4uOj^2NWRe~n)Lj~Ai$;zg+5vM?k1hM(9^c%ox18c#?R2k=NvC7)Ud6tVawQ99!Lwv|o}jw`=LQGp9gvLLTL!dGN8#3%Z#z}xlgT7*OpAQS4EEgSuiD8jYcQ7So-bX&hwe+=97#?ZYQ0m0`}l%3Vc0d3i4MzL6~zp4=?cL<P8YpXfZSw1ds(Lgp3bKEo@umHczWQ7`t4aRKmsYgH9MwB6@G)YRL3c+_9H)ZI;ym2QR-`?sO+grPD$yu+VfG%~*A!<dFH&^RgtQ-oCsZId()e^Iu8eFhBgihEdX^f+c%sJhNmC`ZrzvxD$(wwvO4$~>~q!gjI>0^?KO)O79>C(s`XzOTmEog0&Jyl%Gc7x>eizQ2p-=BZ_k=^Hwr=&Xgvm<WM#mk}NiftVFxEtAJ*HB~3bD52#KX8vG2z`>r{#XfA4t7ypz&tz%A;8o9tg^j<grAv>S`O7M#{7aL{HR^vwoo>I&eJVHtw793R-IjBn7gZu(yHdF;EsxiEh_zJ8KS9i<#4i~11`|g07`bdtb}3{92Nn$D;s!8=uu?~(55GC^F<WvY31~R(ue3Jqq0BXO{2Mh+bea!-Ur6eEpo@;ImSvKoHsaH_+!w0eTjS4TJuYWjE#1W)qH(vbAh+#tUGwa6B^8u2XyeSHSF>d<ToY*a2U^F0vO4!;Q@XXSx%RIvL4cLYYaav@amb|`$l|V$DIO$M|9>HbS213+WDtW3^-<8f#FNlvKC<|KK9gCYl=5H15c5HysOL!E#4NTI>oCo36D!?N(6ORxdebKL1+5=Ea^1H9N$I;0--mj%Qt)6uS=PJELc}%QBN}sS~i=bKE&SJZLqg?E1)<`L96TSa(}4vhVq-dxXTSvJgOgmDWBxwXG-<*6=YmZC-OBAnXD%MIC%&(r%nrSeVzGQ3aPeiA%s@Gq2VQWoSs!$vh#@ks^ELRzbGSJJWRQxXw5v3l;RxydaO5Cgm9_2q`((E{gB^T)Z?vpoj@}31Wi)B*c7+XhtM(Ibwo*AlH1gXhL%RRAE@kXbG&AL!`O9{I22WcpHac1lRxe{Oocudq#21YpGOVY6XULqUF!oB@z|E#qw2mY&7>cgp7RB&syAKl*76FFEX1n|DD2+dThqxU?_$@(h5XKNF-kft6(nd?bSN}1jk2yVlk#hiTVj=mi-4!?jFJe7qKJ33#xx~$HcRow>M;nlMUK&b0k^@R8s9O7V(!o<e<F6|1H<jD%r+~^ibWp_LWD*u&_4l9Ss2cp-5v}(+SZOw{3OeU8b+NuzyOcPyaK7<Nf8T%=+Qw8P%P`6ZnQgo5*$MIQ!Gp?_6uzdVewmA28=_Lp$|`-cUV<+LNP0%_sCbDPdNRxdGgq~>vXMVJ2UwT+kOfaGy_W#BJvZ?7IM(~dLgiJG2gUW)nr{{IVGF)I(Yf8ArBNNR*b&I;%eqbrI(sSSwm`2TtqR)p^F4%N3ojmK#;quPT(cL^i~ep=&pn**1)e_I(V{ymyde)(FC{q(Iz`CsNdD%YL>yM!_1>|xJv+Q<T7E36<K``8`PAh{qi-y6^cj_AEpQ+8b+cps$z<X(+htRXqP|Qi7H%itO`Q#O!=s4w#dhgH5s97GJ+a>fLfvbftL};)CBCaIMU4b>_H|enWarnU2n1ATGK@f7@sBR@){;A2@G_;qYJ80`bD<h@Xu(-4L}fFhMw_ehHn^^18WSa53xo}*$y*4hqV!Xz!*(ce`ZMZ0ScC{3y6;K5N3Qezw+s!!t568Fc691khq^cYZ?mA^1=43*jOX(JimhkMSCI=ma*f8#i3Qa{7F>YR(mvi(GI!;e66<FdE3@75`F@mhyq+Fn&2~(h>Ou)Ul}rt)tRmgI%PH@YVV_N?hIkxSWYZ;AM}Jm(2N^znWCYGg;BX%`z2U!FcY<01DmbUV`wBdp?>c*uIFuN&4xhl{G)#uzwcjs_!OUi{?7w;`2#E%w`eIQH7rI6yugUfit8#E(?bam{-<Z|Z;U&5PlO4t_t@+utggSDoE`4V0;<38$%gsytiK=kKYaN4C_cY9I6cP-HbtJ~h0nc5{_xbXj_c^es@V0o4*H?_tc@Gn{^P}c!7tP!RzYq#(T<B%yJ7-@{N%m{3I~;xPL_dRbzj!P)ZTU-pH*UM%bhjFJfV4N#^11@Fvm5*n*a0dYc-FxS^5Vdsj2;yfL4ZfX8AZRkuyoq+7X^|#Q9JtM~ZkA3X+EnCb1mA3Kkn6+&{J9zia@J@U7_B<W6#j4P?`K+c(x=4V#F_(#RWDzO^AKj1JBn$o9J&E~WZxCRk`6BCu!Qz7D+{?FHP}F~be@5NX0=(B0a-G-9P27g6DV53>Ajca7p2BP5F&>Qyz@uwp`0SbMH)F`>uPkTbL$**J;Ld<`TX#6*A(fz-Sl>|stF=Kg+=NpXw{#)N&@)0(sPrePjn7D^j_1DM{ngx6L96sudzu4jd?Tep&5*SyR_dl_m#m@kso$TDS`_H<DXI3z8htmGm)Z;uY}r7n0U3BUqP6SwB8c>-gsL-zh->CyTBCBcAUbcJz$cp)P{CXGH&PLIoA4HPZW_OhoV-5fo-iBuh|cvd>B?nYa(6P<rnB|CA|HMGL#S6BUStF6gyij4QoZ3L;PnjLANW;PaB>R()e@`krM{ae&P-PLCuXHz0xY&*T$TcIUtZYymu1CEG)K0k<m?7y<}Z^q+X#jUbzi@K-JJT(AP?jcWR*^u2e15GccQ&5)X6HP9gK2^PyR=69hHC)`>#Dfq#S(_-kCCk#@f~HhSHF$+qB@ESkRPgnR&fwiuw7TC_MmV$%Iy9SZ3HridF6KYc@c-Tb#KT+JD~5=0(->&67G@7|rF1t@e7VX8by8QI)hZv<R2Ui93=oknLcSCcCx2Ko{ByLzKSv{6UjfVv7QtatR0~3Hg|<hORTTzTSr;I_)T&u2_p0AHbBOv{t`O3Pm}+SCU0jpP?y_=dSq5P&jp<K^csHfcW&PeBaq87dqsJzo%A+Sc-BhOTQm!)Ju{#PlX8KfXPSQIjaiH!NwB`IP&yw9WltX=w+@%BGRoy$>DcyBxB`av8jpYCc(Y@pQrPLnjx{cxwt;Kh{7J2`OTI)>H$&)!tdGB-FE$+;xlx=lc66fb8a&9Aok=40TNXzn!Nentx%7@1N(#$gWqb$m~4Y|sdb{{FY!`BQv;KlaR^|ixs4@3c7i5XqtUmX03jeb*2bf#Ao)`5b=waG?olBxYO`rol-*fF(0ZAP_T@gzs9nHmlhW4c{+Wy$(#xdE_FmjT9lI{y-Ve;rl-GFSr{OU1wX_vrCotHb{R{e4Z{UC%wWOo3nRY352h?g@@!(tUwos8q9~60i_#;(<i?vw3rePs<f)t`UO}jH-uxI4O&q5(2N@ifwq4(Mr#KyY<zrq|;<x;dV(ynap=~q8)N@x9ocAyQq)U_a})<_fbAY@5WC8Gxi!E=q+ip2)_HM+U`a><ENRmzvc3nZ>-#bfbSPL$ap*Bs-kQ#>9jNRi>$jv7Pi`3e1iPdJ5OtGF=uw8#f4wHk)|N<{di_djU?1I3u`TtigsI_X-Wn7{rywlY*SDNJr2Er9*b}pVoXdj!Zn}g)-^5<H4Qk?f(c#YHL;sVnMof}e#z{eDsibc&y8e|G6Y(Vip&f|f{sGuhz(Taoskbq%WpZ+<7R?JLYTF}A~-Ie(kbpjj2AfzWv7nRf_K>#;eC7u@@`1yUd<_r=M(g=RSGNksA9wD{`7-b>t!l+I?QFY;mpp}D7A$KG{rVxoLPCB%2qZxj@I@!mhG~#7?ASlu1$gfCS3{)Lzjs49_^-;tdI$81kvJgnGEUf0+!lelWzQIKFMQxGBTvNh|r7#7uTEseA*Pe2h$dlW6|UXq<0(1TQup)D%cItF}?4!u1qnr+SJF6EJqov=%k#1HJ4Mb>6UaSy1Uv?Ks_q_L(gLgdMzkU0Ejow=Zj@<tJA7R;aNBbnzzSe>qgQp)Y|bc-qXOZjaIE{8PDuwTvD^*vyHsjOAT-Hgf>V&07zq}-E-U6mTX*~)*=(lycZ?R^{h-OVIsMl2rjO+WG%{;dp*ZisBy-LC+K+i(-U<3TeldRn+v8ZT#XrqW8>C<bML|D`|XatM1r(#R@inCRvA*n^vrJOv)4+786!V0<~-9SaxK5@wY`)Qr}_*t%A-4MS5BMFc5rOZjj(K#awo>rX3BHZW2$QwkR(P^u2gz$S@djlmNcw2&eL)>GIefBaj;=ZIT}<tA4WHxHx7!($kj!iO{<7XQXMQF`z<;_*nq5=-5PDJG=W^GDaI3P`4K5FY&<s%39Yk>ejbTzDjuaI0(~A?cBq!)q*fk_mj7JmSJ#R&)zQ*2de;0p<-xF+KLAeY93Om^Ls7vB+$?ixs(6O|C9u_~q>E;qc$UZ!>xfo_&Dq8gXALwgkaP$`uSHI2?;Frqm?y--2EH<pt>fbKOM!x;mFaqq7MQ2)M-z&EHgwgM@Onm5^M+N&-HZ$wx?7#K*Oj+avl~{skNW<Nd0Fg}$1hto3R%~^K6RRx$3K4z*B{6KuVd*X;nxi6s(~{vel<#(5qb_$Va4iMLObL)vA{U|p3QwV8N=53Z31j&OSN&L$KW|?+j3264K2!G^shC!glE9R;b3;?ENO_}Mpiu`dv)Quwp_&)C_6Jhn$CZ1DBBVF0zD$k!V{(w*#Gl58P_yO+@kURizOb^Zo~RIAJT@He)5dg-2@D8qO!KHipD!HrSw;2^3W8t@N6W>r<XtU?EYx$Z8a$t(-H21qa!7y3q?i4*aC~bBkx>#o0boxc}dXUc#ZEk|IG_C%@$l5L<<ECj?)L~<@)Y2J5ok>+R5-wb<jcjmO@YSGMK{Y5xY+EM~#N&iSsm9+-uDh=P6&@>#;4Z)}zBxjuJ7QV)&YuQLn=+-y2NkUT>X(d)z_zBxlp4oIZ3tWozSYq)cp(^U-^>N0j>&tk_K*G_3(H!gp3Ae7!2u<Yq%Vc&zzcn~e1=zT%cm(*vSAKLX>95Sn1h&9F=lk++((uVh$4_!Q(td$LUZ$2A$eWSsy?MTSO#4FomaeWZV*SLj~;Hz+xV%fa+vFbBrQhQrEsB<r;F^EH-`CiBY$i)Wmo+m__2!jmjEmoRZ767%Dz@|qMTt){2>B9i&MEbe2+8ugV-YfF60)5H9(U?`g@k|RO%YbEoI63BPHp$y8^z^*FP!6m{caNLWa)(wegw|T1w0!xQtf!y?5-WhM(^V^8I%_(O}jO6y3aAf3euZv|iR?5b<^+GYWkcTX8lO11ROGn26&5)z*$i0P*v2$`WCzx5a&4sGgT@u4c+HFEgk(fSotUD7l%N*Dpll~-gmK>_psZE^IhV&23x~(cKOPjrx``p4X=BBxDV6vZd^Q2ZPSQ@!u?}awOJN876`o9KkNMY<;QA{0xjN>fr3%Mk&2En%5%}Q+2s9Zv7RqM=Q#q?(Btbq9oy~J};U^E~@NF@bH&q~CygijG>yE}WY#os6ID0{E5-?+oK9NM9Y0xiHg8yfHZ)%%4K`M`{Yd37yzZs-s$NZBqh+VlY}j)SAV$oa#U7QBZp)ef1pECm}%9$03IjP@}g@pXY^Cb5?rxq6s3$`bQaw!Gwkm~nj9+``lMSb;a#runl@BC45G0)S!`RVfV3g#PeM74xX=6N5HO@&$!@I=}Mamqo~O6~&C;S|Gl;+iM?oW-1HoZ8ZTC7G)TQ(b3Iyr%8?8dxTV_bp`Y-DFmpPV)E<gdRkohFUDD(-lbPeG{u|IRhGX9*dC^8K}COYm3~R=soec+A0@fjz<0r$rif0jm^PD(h!RW)LsLaWbmnEc^+mf31(SEwTfljpP?|gX!EVjoG1213Mb!b;(C7n^GiGOA<A4rt#u&onlfH-NI2=g`oGu6wl2=JtW-`FSIQD9sqBTc+neWaxcInb~+vBEJZ0#$^UlA?cfUTYHw8v7Or~SIIBkJ(Tb{>2hb1hy=|KMsbHfw3ki!Uu`=%<aD@R^aznYCwJ#*OQ(bw6iE9qPk7Ds~0$sKADndWR<;`iJpZ|M(}Yw<GGA7l?Q0yLl}rt(bbV=4Z(xlr>VxD>GipJ}7wxdN)v&9hK~S_<6sN3pDqyjnq}18ad4UV@H+}g(5p&d6nYC#<YGwe>^n#s2zb+O}sPjo&L-E&GdADtj$qIM>^kmRpB+ZQDK#Oz$nQ&g-=rzAFwQahMhb^_kpB%s~VNICVaaLp$kuWrUJGT=JPxBdYybYoOZN;nTA$pNg)j%4x<uFRSiZT>V7br>hY^R-FR+R$n&K|v70<|T+ARt2hw4NDGL0~1QeEzfs2s-!y+GN*K+eQ9*ySGtIf`^U_NtmmG^~2uhFDJWhtKU*rv%kiDF<{;6)m+ZdQH=AXF$8?la8vbz}L$ah(Li3JkJNuY<w~6xnI2M<7Wx%|f!P6>s|2MQQ?`Zt*q&R8Z4`4oIWD8AO1S_a<Uz8qh`Hcj8Xawvps{k2jeA;r6K&6T-XKVX-z#dpGWYEFhQn=|epMar0O`R1scR0OR}AS-b894!%&B<Cp`8;}sy<k<>z`lwN2#?3d0p%<xtg57u-;p^qynU1l4{l*uV(>=g4{PHa8c_<t?va+m')))
\ No newline at end of file
diff --git a/records/track_10min_16mb/2026-04-17_SP8192_AdaptiveHessianClip/train_seed1337.log b/records/track_10min_16mb/2026-04-17_SP8192_AdaptiveHessianClip/train_seed1337.log
new file mode 100644
index 0000000000..b2fb5faa80
--- /dev/null
+++ b/records/track_10min_16mb/2026-04-17_SP8192_AdaptiveHessianClip/train_seed1337.log
@@ -0,0 +1,205 @@
+W0417 04:54:56.899000 41906 torch/distributed/run.py:803] 
+W0417 04:54:56.899000 41906 torch/distributed/run.py:803] *****************************************
+W0417 04:54:56.899000 41906 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. 
+W0417 04:54:56.899000 41906 torch/distributed/run.py:803] *****************************************
+[04:55:09] Hyperparameters:
+[04:55:09]   adam_eps: 1e-08
+[04:55:09]   adam_wd: 0.02
+[04:55:09]   beta1: 0.9
+[04:55:09]   beta2: 0.95
+[04:55:09]   compressor: brotli
+[04:55:09]   data_dir: ./data/
+[04:55:09]   datasets_dir: ./data/datasets/fineweb10B_sp8192
+[04:55:09]   distributed: True
+[04:55:09]   ema_decay: 0.9965
+[04:55:09]   embed_bits: 8
+[04:55:09]   embed_clip_sigmas: 20.0
+[04:55:09]   embed_lr: 0.6
+[04:55:09]   embed_wd: 0.085
+[04:55:09]   embedding_dim: 512
+[04:55:09]   enable_looping_at: 0.35
+[04:55:09]   eval_seq_len: 2048
+[04:55:09]   eval_stride: 64
+[04:55:09]   gptq_calibration_batches: 64
+[04:55:09]   gptq_reserve_seconds: 12.0
+[04:55:09]   grad_accum_steps: 1
+[04:55:09]   grad_clip_norm: 0.3
+[04:55:09]   head_lr: 0.008
+[04:55:09]   is_main_process: True
+[04:55:09]   iterations: 20000
+[04:55:09]   ln_scale: True
+[04:55:09]   local_rank: 0
+[04:55:09]   logfile: logs/8124e954-a9bc-4a08-bd7b-bf5c97127188.txt
+[04:55:09]   logit_softcap: 30.0
+[04:55:09]   loop_end: 5
+[04:55:09]   loop_start: 3
+[04:55:09]   matrix_bits: 6
+[04:55:09]   matrix_clip_sigmas: 12.85
+[04:55:09]   matrix_lr: 0.022
+[04:55:09]   max_wallclock_seconds: 600.0
+[04:55:09]   min_lr: 0.0
+[04:55:09]   mlp_mult: 4.0
+[04:55:09]   model_dim: 512
+[04:55:09]   model_path: final_model.pt
+[04:55:09]   muon_backend_steps: 5
+[04:55:09]   muon_beta2: 0.95
+[04:55:09]   muon_momentum: 0.99
+[04:55:09]   muon_momentum_warmup_start: 0.92
+[04:55:09]   muon_momentum_warmup_steps: 1500
+[04:55:09]   muon_row_normalize: True
+[04:55:09]   muon_wd: 0.095
+[04:55:09]   num_heads: 8
+[04:55:09]   num_kv_heads: 4
+[04:55:09]   num_layers: 11
+[04:55:09]   num_loops: 2
+[04:55:09]   parallel_start_layer: 7
+[04:55:09]   qk_gain_init: 5.25
+[04:55:09]   quantized_model_path: final_model.int6.ptz
+[04:55:09]   rank: 0
+[04:55:09]   rope_base: 10000.0
+[04:55:09]   rope_dims: 16
+[04:55:09]   rope_train_seq_len: 2048
+[04:55:09]   run_id: 8124e954-a9bc-4a08-bd7b-bf5c97127188
+[04:55:09]   scalar_lr: 0.02
+[04:55:09]   seed: 1337
+[04:55:09]   skip_gates_enabled: True
+[04:55:09]   sliding_window_enabled: True
+[04:55:09]   tie_embeddings: True
+[04:55:09]   tied_embed_init_std: 0.005
+[04:55:09]   tied_embed_lr: 0.03
+[04:55:09]   tokenizer_path: ./data/tokenizers/fineweb_8192_bpe.model
+[04:55:09]   train_batch_tokens: 786432
+[04:55:09]   train_files: ./data/datasets/fineweb10B_sp8192/fineweb_train_*.bin
+[04:55:09]   train_log_every: 500
+[04:55:09]   train_seq_len: 2048
+[04:55:09]   val_batch_tokens: 524288
+[04:55:09]   val_files: ./data/datasets/fineweb10B_sp8192/fineweb_val_*.bin
+[04:55:09]   val_loss_every: 4000
+[04:55:09]   vocab_size: 8192
+[04:55:09]   warmdown_frac: 0.72
+[04:55:09]   warmup_steps: 20
+[04:55:09]   world_size: 8
+[04:55:09]   xsa_last_n: 11
+[04:55:09] train_shards: 128
+[04:55:09] val_tokens: 40540160
+[04:55:10] model_params:35944537
+[04:55:11] gptq:reserving 12s, effective=588000ms
+[04:55:25] warmup_step: 1/20
+[04:55:35] warmup_step: 2/20
+[04:55:36] warmup_step: 3/20
+[04:55:36] warmup_step: 4/20
+[04:55:36] warmup_step: 5/20
+[04:55:36] warmup_step: 6/20
+[04:55:36] warmup_step: 10/20
+[04:55:37] warmup_step: 20/20
+[04:55:37] loop_warmup:enabled encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10]
+[04:55:48] loop_warmup_step: 1/20
+[04:55:48] loop_warmup_step: 2/20
+[04:55:48] loop_warmup_step: 3/20
+[04:55:48] loop_warmup_step: 4/20
+[04:55:49] loop_warmup_step: 5/20
+[04:55:49] loop_warmup_step: 6/20
+[04:55:49] loop_warmup_step: 10/20
+[04:55:51] loop_warmup_step: 20/20
+[04:55:57] 0/20000 val_loss: 9.0056 val_bpb: 3.4864
+[04:55:57] 1/20000 train_loss: 8.9990 train_time: 0.0m ms/step: 94 tok/s: 8382252
+[04:55:57] 2/20000 train_loss: 12.1311 train_time: 0.0m ms/step: 96 tok/s: 8191978
+[04:55:57] 3/20000 train_loss: 10.8970 train_time: 0.0m ms/step: 97 tok/s: 8089170
+[04:55:57] 4/20000 train_loss: 9.3165 train_time: 0.0m ms/step: 98 tok/s: 8026657
+[04:55:57] 5/20000 train_loss: 8.2513 train_time: 0.0m ms/step: 98 tok/s: 8005964
+[04:56:47] 500/20000 train_loss: 3.3531 train_time: 0.8m ms/step: 101 tok/s: 7784851
+[04:57:38] 1000/20000 train_loss: 3.2413 train_time: 1.7m ms/step: 101 tok/s: 7779617
+[04:58:28] 1500/20000 train_loss: 3.1927 train_time: 2.5m ms/step: 101 tok/s: 7786280
+[04:59:19] 2000/20000 train_loss: 3.0750 train_time: 3.4m ms/step: 101 tok/s: 7786337
+[04:59:23] layer_loop:enabled step:2038 frac:0.350 encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10]
+[05:00:31] 2500/20000 train_loss: 2.9983 train_time: 4.6m ms/step: 110 tok/s: 7160532
+[05:01:46] 3000/20000 train_loss: 3.0000 train_time: 5.8m ms/step: 116 tok/s: 6763256
+[05:03:00] 3500/20000 train_loss: 2.8969 train_time: 7.1m ms/step: 121 tok/s: 6506227
+[05:04:14] 4000/20000 train_loss: 3.0013 train_time: 8.3m ms/step: 124 tok/s: 6325949
+[05:04:20] 4000/20000 val_loss: 2.8729 val_bpb: 1.1122
+[05:05:34] 4500/20000 train_loss: 2.8262 train_time: 9.5m ms/step: 127 tok/s: 6193216
+[05:05:53] 4612/20000 val_loss: 2.8054 val_bpb: 1.0861
+[05:05:53] stopping_early: wallclock_cap train_time: 588046ms step: 4612/20000
+[05:05:53] peak memory allocated: 39716 MiB reserved: 39742 MiB
+[05:05:53] ema:applying EMA weights
+[05:06:00] pre-quantization post-ema val_loss:2.80218688 val_bpb:1.08483663 eval_time:6337ms
+[05:06:00] Serialized model: 135431355 bytes
+[05:06:00] Code size: 19846 bytes
+[05:06:00] GPTQ:collecting Hessians from calibration data...
+[05:06:13] GPTQ:collected 67 Hessians in 12.8s
+[05:06:19] GPTQ adaptive clip_sigmas (Hessian-sensitivity):
+[05:06:19]     tok_emb.weight: 20.00
+[05:06:19]     blocks.0.attn.c_q.weight: 11.13
+[05:06:19]     blocks.0.attn.c_k.weight: 10.94
+[05:06:19]     blocks.0.attn.c_v.weight: 10.71
+[05:06:19]     blocks.0.attn.proj.weight: 11.42
+[05:06:19]     blocks.0.mlp.fc.weight: 10.62
+[05:06:19]     blocks.0.mlp.proj.weight: 8.46
+[05:06:19]     blocks.1.attn.c_q.weight: 12.47
+[05:06:19]     blocks.1.attn.c_k.weight: 12.07
+[05:06:19]     blocks.1.attn.c_v.weight: 11.86
+[05:06:19]     blocks.1.attn.proj.weight: 13.01
+[05:06:19]     blocks.1.mlp.fc.weight: 11.69
+[05:06:19]     blocks.1.mlp.proj.weight: 10.34
+[05:06:19]     blocks.2.attn.c_q.weight: 13.20
+[05:06:19]     blocks.2.attn.c_k.weight: 12.85
+[05:06:19]     blocks.2.attn.c_v.weight: 12.66
+[05:06:19]     blocks.2.attn.proj.weight: 14.05
+[05:06:19]     blocks.2.mlp.fc.weight: 12.41
+[05:06:19]     blocks.2.mlp.proj.weight: 11.65
+[05:06:19]     blocks.3.attn.c_q.weight: 11.76
+[05:06:19]     blocks.3.attn.c_k.weight: 11.54
+[05:06:19]     blocks.3.attn.c_v.weight: 11.34
+[05:06:19]     blocks.3.attn.proj.weight: 11.17
+[05:06:19]     blocks.3.mlp.fc.weight: 11.03
+[05:06:19]     blocks.3.mlp.proj.weight: 10.98
+[05:06:19]     blocks.4.attn.c_q.weight: 12.31
+[05:06:19]     blocks.4.attn.c_k.weight: 11.63
+[05:06:19]     blocks.4.attn.c_v.weight: 11.63
+[05:06:19]     blocks.4.attn.proj.weight: 12.85
+[05:06:19]     blocks.4.mlp.fc.weight: 11.44
+[05:06:19]     blocks.4.mlp.proj.weight: 11.54
+[05:06:19]     blocks.5.attn.c_q.weight: 12.52
+[05:06:19]     blocks.5.attn.c_k.weight: 12.00
+[05:06:19]     blocks.5.attn.c_v.weight: 11.78
+[05:06:19]     blocks.5.attn.proj.weight: 13.64
+[05:06:19]     blocks.5.mlp.fc.weight: 11.71
+[05:06:19]     blocks.5.mlp.proj.weight: 11.44
+[05:06:19]     blocks.6.attn.c_q.weight: 15.04
+[05:06:19]     blocks.6.attn.c_k.weight: 14.46
+[05:06:19]     blocks.6.attn.c_v.weight: 14.37
+[05:06:19]     blocks.6.attn.proj.weight: 16.74
+[05:06:19]     blocks.6.mlp.fc.weight: 14.06
+[05:06:19]     blocks.6.mlp.proj.weight: 14.27
+[05:06:19]     blocks.7.attn.c_q.weight: 15.28
+[05:06:19]     blocks.7.attn.c_k.weight: 14.76
+[05:06:19]     blocks.7.attn.c_v.weight: 14.35
+[05:06:19]     blocks.7.attn.proj.weight: 14.00
+[05:06:19]     blocks.7.mlp.fc.weight: 14.35
+[05:06:19]     blocks.7.mlp.proj.weight: 15.62
+[05:06:19]     blocks.8.attn.c_q.weight: 16.27
+[05:06:19]     blocks.8.attn.c_k.weight: 15.32
+[05:06:19]     blocks.8.attn.c_v.weight: 14.08
+[05:06:19]     blocks.8.attn.proj.weight: 19.02
+[05:06:19]     blocks.8.mlp.fc.weight: 14.60
+[05:06:19]     blocks.8.mlp.proj.weight: 16.31
+[05:06:19]     blocks.9.attn.c_q.weight: 16.02
+[05:06:19]     blocks.9.attn.c_k.weight: 15.45
+[05:06:19]     blocks.9.attn.c_v.weight: 14.54
+[05:06:19]     blocks.9.attn.proj.weight: 17.17
+[05:06:19]     blocks.9.mlp.fc.weight: 14.80
+[05:06:19]     blocks.9.mlp.proj.weight: 16.58
+[05:06:19]     blocks.10.attn.c_q.weight: 16.05
+[05:06:19]     blocks.10.attn.c_k.weight: 15.59
+[05:06:19]     blocks.10.attn.c_v.weight: 15.60
+[05:06:19]     blocks.10.attn.proj.weight: 17.37
+[05:06:19]     blocks.10.mlp.fc.weight: 14.95
+[05:06:19]     blocks.10.mlp.proj.weight: 15.94
+[05:06:19] Quantized weights:
+[05:06:19]   gptq (int6): blocks.attn.c_k.weight, blocks.attn.c_q.weight, blocks.attn.c_v.weight, blocks.attn.proj.weight, blocks.mlp.fc.weight, blocks.mlp.proj.weight
+[05:06:19]   gptq (int8): tok_emb.weight
+[05:06:19]   passthrough (float16): blocks.attn.q_gain, blocks.attn_scale, blocks.mlp_scale, blocks.resid_mix, lane_merge, skip_gates, skip_weights
+[05:07:10] Serialized model quantized+brotli: 15887082 bytes
+[05:07:10] Total submission size quantized+brotli: 15906928 bytes
+[05:07:20] quantized val_loss:2.83669059 val_bpb:1.09819437 eval_time:7829ms
+[05:08:50] quantized_sliding_window val_loss:2.79264397 val_bpb:1.08114219 eval_time:90736ms
diff --git a/records/track_10min_16mb/2026-04-17_SP8192_AdaptiveHessianClip/train_seed42.log b/records/track_10min_16mb/2026-04-17_SP8192_AdaptiveHessianClip/train_seed42.log
new file mode 100644
index 0000000000..a463ad1a0c
--- /dev/null
+++ b/records/track_10min_16mb/2026-04-17_SP8192_AdaptiveHessianClip/train_seed42.log
@@ -0,0 +1,205 @@
+W0417 05:10:53.314000 42736 torch/distributed/run.py:803] 
+W0417 05:10:53.314000 42736 torch/distributed/run.py:803] *****************************************
+W0417 05:10:53.314000 42736 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. 
+W0417 05:10:53.314000 42736 torch/distributed/run.py:803] *****************************************
+[05:11:06] Hyperparameters:
+[05:11:06]   adam_eps: 1e-08
+[05:11:06]   adam_wd: 0.02
+[05:11:06]   beta1: 0.9
+[05:11:06]   beta2: 0.95
+[05:11:06]   compressor: brotli
+[05:11:06]   data_dir: ./data/
+[05:11:06]   datasets_dir: ./data/datasets/fineweb10B_sp8192
+[05:11:06]   distributed: True
+[05:11:06]   ema_decay: 0.9965
+[05:11:06]   embed_bits: 8
+[05:11:06]   embed_clip_sigmas: 20.0
+[05:11:06]   embed_lr: 0.6
+[05:11:06]   embed_wd: 0.085
+[05:11:06]   embedding_dim: 512
+[05:11:06]   enable_looping_at: 0.35
+[05:11:06]   eval_seq_len: 2048
+[05:11:06]   eval_stride: 64
+[05:11:06]   gptq_calibration_batches: 64
+[05:11:06]   gptq_reserve_seconds: 12.0
+[05:11:06]   grad_accum_steps: 1
+[05:11:06]   grad_clip_norm: 0.3
+[05:11:06]   head_lr: 0.008
+[05:11:06]   is_main_process: True
+[05:11:06]   iterations: 20000
+[05:11:06]   ln_scale: True
+[05:11:06]   local_rank: 0
+[05:11:06]   logfile: logs/92dfe3a5-f98b-4b29-837c-1a108f536d7a.txt
+[05:11:06]   logit_softcap: 30.0
+[05:11:06]   loop_end: 5
+[05:11:06]   loop_start: 3
+[05:11:06]   matrix_bits: 6
+[05:11:06]   matrix_clip_sigmas: 12.85
+[05:11:06]   matrix_lr: 0.022
+[05:11:06]   max_wallclock_seconds: 600.0
+[05:11:06]   min_lr: 0.0
+[05:11:06]   mlp_mult: 4.0
+[05:11:06]   model_dim: 512
+[05:11:06]   model_path: final_model.pt
+[05:11:06]   muon_backend_steps: 5
+[05:11:06]   muon_beta2: 0.95
+[05:11:06]   muon_momentum: 0.99
+[05:11:06]   muon_momentum_warmup_start: 0.92
+[05:11:06]   muon_momentum_warmup_steps: 1500
+[05:11:06]   muon_row_normalize: True
+[05:11:06]   muon_wd: 0.095
+[05:11:06]   num_heads: 8
+[05:11:06]   num_kv_heads: 4
+[05:11:06]   num_layers: 11
+[05:11:06]   num_loops: 2
+[05:11:06]   parallel_start_layer: 7
+[05:11:06]   qk_gain_init: 5.25
+[05:11:06]   quantized_model_path: final_model.int6.ptz
+[05:11:06]   rank: 0
+[05:11:06]   rope_base: 10000.0
+[05:11:06]   rope_dims: 16
+[05:11:06]   rope_train_seq_len: 2048
+[05:11:06]   run_id: 92dfe3a5-f98b-4b29-837c-1a108f536d7a
+[05:11:06]   scalar_lr: 0.02
+[05:11:06]   seed: 42
+[05:11:06]   skip_gates_enabled: True
+[05:11:06]   sliding_window_enabled: True
+[05:11:06]   tie_embeddings: True
+[05:11:06]   tied_embed_init_std: 0.005
+[05:11:06]   tied_embed_lr: 0.03
+[05:11:06]   tokenizer_path: ./data/tokenizers/fineweb_8192_bpe.model
+[05:11:06]   train_batch_tokens: 786432
+[05:11:06]   train_files: ./data/datasets/fineweb10B_sp8192/fineweb_train_*.bin
+[05:11:06]   train_log_every: 500
+[05:11:06]   train_seq_len: 2048
+[05:11:06]   val_batch_tokens: 524288
+[05:11:06]   val_files: ./data/datasets/fineweb10B_sp8192/fineweb_val_*.bin
+[05:11:06]   val_loss_every: 4000
+[05:11:06]   vocab_size: 8192
+[05:11:06]   warmdown_frac: 0.72
+[05:11:06]   warmup_steps: 20
+[05:11:06]   world_size: 8
+[05:11:06]   xsa_last_n: 11
+[05:11:06] train_shards: 128
+[05:11:06] val_tokens: 40540160
+[05:11:07] model_params:35944537
+[05:11:07] gptq:reserving 12s, effective=588000ms
+[05:11:22] warmup_step: 1/20
+[05:11:30] warmup_step: 2/20
+[05:11:33] warmup_step: 3/20
+[05:11:33] warmup_step: 4/20
+[05:11:33] warmup_step: 5/20
+[05:11:33] warmup_step: 6/20
+[05:11:34] warmup_step: 10/20
+[05:11:35] warmup_step: 20/20
+[05:11:35] loop_warmup:enabled encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10]
+[05:11:45] loop_warmup_step: 1/20
+[05:11:53] loop_warmup_step: 2/20
+[05:11:53] loop_warmup_step: 3/20
+[05:11:53] loop_warmup_step: 4/20
+[05:11:54] loop_warmup_step: 5/20
+[05:11:54] loop_warmup_step: 6/20
+[05:11:54] loop_warmup_step: 10/20
+[05:11:56] loop_warmup_step: 20/20
+[05:12:01] 0/20000 val_loss: 9.0094 val_bpb: 3.4879
+[05:12:01] 1/20000 train_loss: 9.0044 train_time: 0.0m ms/step: 94 tok/s: 8388768
+[05:12:01] 2/20000 train_loss: 12.1695 train_time: 0.0m ms/step: 96 tok/s: 8178614
+[05:12:01] 3/20000 train_loss: 10.9234 train_time: 0.0m ms/step: 98 tok/s: 8060992
+[05:12:01] 4/20000 train_loss: 9.2871 train_time: 0.0m ms/step: 98 tok/s: 8018310
+[05:12:01] 5/20000 train_loss: 8.2345 train_time: 0.0m ms/step: 98 tok/s: 7994414
+[05:12:51] 500/20000 train_loss: 3.3404 train_time: 0.8m ms/step: 101 tok/s: 7777469
+[05:13:42] 1000/20000 train_loss: 3.2332 train_time: 1.7m ms/step: 101 tok/s: 7764576
+[05:14:32] 1500/20000 train_loss: 3.1976 train_time: 2.5m ms/step: 101 tok/s: 7778013
+[05:15:23] 2000/20000 train_loss: 3.0725 train_time: 3.4m ms/step: 101 tok/s: 7781869
+[05:15:26] layer_loop:enabled step:2037 frac:0.350 encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10]
+[05:16:35] 2500/20000 train_loss: 3.0001 train_time: 4.6m ms/step: 110 tok/s: 7160223
+[05:17:49] 3000/20000 train_loss: 3.0005 train_time: 5.8m ms/step: 116 tok/s: 6764119
+[05:19:04] 3500/20000 train_loss: 2.8995 train_time: 7.1m ms/step: 121 tok/s: 6507092
+[05:20:18] 4000/20000 train_loss: 3.0029 train_time: 8.3m ms/step: 124 tok/s: 6325841
+[05:20:24] 4000/20000 val_loss: 2.8745 val_bpb: 1.1128
+[05:21:39] 4500/20000 train_loss: 2.8266 train_time: 9.5m ms/step: 127 tok/s: 6192265
+[05:21:58] 4612/20000 val_loss: 2.8068 val_bpb: 1.0866
+[05:21:58] stopping_early: wallclock_cap train_time: 588125ms step: 4612/20000
+[05:21:58] peak memory allocated: 39717 MiB reserved: 39742 MiB
+[05:21:58] ema:applying EMA weights
+[05:22:04] pre-quantization post-ema val_loss:2.80359756 val_bpb:1.08538276 eval_time:6896ms
+[05:22:05] Serialized model: 135431355 bytes
+[05:22:05] Code size: 19846 bytes
+[05:22:05] GPTQ:collecting Hessians from calibration data...
+[05:22:17] GPTQ:collected 67 Hessians in 12.7s
+[05:22:25] GPTQ adaptive clip_sigmas (Hessian-sensitivity):
+[05:22:25]     tok_emb.weight: 20.00
+[05:22:25]     blocks.0.attn.c_q.weight: 11.09
+[05:22:25]     blocks.0.attn.c_k.weight: 11.02
+[05:22:25]     blocks.0.attn.c_v.weight: 10.76
+[05:22:25]     blocks.0.attn.proj.weight: 11.28
+[05:22:25]     blocks.0.mlp.fc.weight: 10.59
+[05:22:25]     blocks.0.mlp.proj.weight: 8.04
+[05:22:25]     blocks.1.attn.c_q.weight: 12.56
+[05:22:25]     blocks.1.attn.c_k.weight: 12.19
+[05:22:25]     blocks.1.attn.c_v.weight: 11.96
+[05:22:25]     blocks.1.attn.proj.weight: 12.11
+[05:22:25]     blocks.1.mlp.fc.weight: 11.68
+[05:22:25]     blocks.1.mlp.proj.weight: 10.79
+[05:22:25]     blocks.2.attn.c_q.weight: 13.35
+[05:22:25]     blocks.2.attn.c_k.weight: 12.70
+[05:22:25]     blocks.2.attn.c_v.weight: 12.76
+[05:22:25]     blocks.2.attn.proj.weight: 14.16
+[05:22:25]     blocks.2.mlp.fc.weight: 12.40
+[05:22:25]     blocks.2.mlp.proj.weight: 11.64
+[05:22:25]     blocks.3.attn.c_q.weight: 11.90
+[05:22:25]     blocks.3.attn.c_k.weight: 11.45
+[05:22:25]     blocks.3.attn.c_v.weight: 11.25
+[05:22:25]     blocks.3.attn.proj.weight: 11.64
+[05:22:25]     blocks.3.mlp.fc.weight: 11.06
+[05:22:25]     blocks.3.mlp.proj.weight: 10.88
+[05:22:25]     blocks.4.attn.c_q.weight: 12.24
+[05:22:25]     blocks.4.attn.c_k.weight: 11.74
+[05:22:25]     blocks.4.attn.c_v.weight: 11.63
+[05:22:25]     blocks.4.attn.proj.weight: 12.72
+[05:22:25]     blocks.4.mlp.fc.weight: 11.45
+[05:22:25]     blocks.4.mlp.proj.weight: 11.60
+[05:22:25]     blocks.5.attn.c_q.weight: 12.56
+[05:22:25]     blocks.5.attn.c_k.weight: 12.13
+[05:22:25]     blocks.5.attn.c_v.weight: 11.95
+[05:22:25]     blocks.5.attn.proj.weight: 12.89
+[05:22:25]     blocks.5.mlp.fc.weight: 11.73
+[05:22:25]     blocks.5.mlp.proj.weight: 11.77
+[05:22:25]     blocks.6.attn.c_q.weight: 15.00
+[05:22:25]     blocks.6.attn.c_k.weight: 14.57
+[05:22:25]     blocks.6.attn.c_v.weight: 14.43
+[05:22:25]     blocks.6.attn.proj.weight: 16.67
+[05:22:25]     blocks.6.mlp.fc.weight: 14.09
+[05:22:25]     blocks.6.mlp.proj.weight: 13.95
+[05:22:25]     blocks.7.attn.c_q.weight: 15.43
+[05:22:25]     blocks.7.attn.c_k.weight: 14.80
+[05:22:25]     blocks.7.attn.c_v.weight: 14.27
+[05:22:25]     blocks.7.attn.proj.weight: 14.93
+[05:22:25]     blocks.7.mlp.fc.weight: 14.38
+[05:22:25]     blocks.7.mlp.proj.weight: 15.57
+[05:22:25]     blocks.8.attn.c_q.weight: 15.87
+[05:22:25]     blocks.8.attn.c_k.weight: 15.18
+[05:22:25]     blocks.8.attn.c_v.weight: 14.34
+[05:22:25]     blocks.8.attn.proj.weight: 19.01
+[05:22:25]     blocks.8.mlp.fc.weight: 14.63
+[05:22:25]     blocks.8.mlp.proj.weight: 16.29
+[05:22:25]     blocks.9.attn.c_q.weight: 16.25
+[05:22:25]     blocks.9.attn.c_k.weight: 15.60
+[05:22:25]     blocks.9.attn.c_v.weight: 14.57
+[05:22:25]     blocks.9.attn.proj.weight: 16.95
+[05:22:25]     blocks.9.mlp.fc.weight: 14.82
+[05:22:25]     blocks.9.mlp.proj.weight: 16.76
+[05:22:25]     blocks.10.attn.c_q.weight: 16.06
+[05:22:25]     blocks.10.attn.c_k.weight: 15.62
+[05:22:25]     blocks.10.attn.c_v.weight: 15.49
+[05:22:25]     blocks.10.attn.proj.weight: 17.92
+[05:22:25]     blocks.10.mlp.fc.weight: 14.96
+[05:22:25]     blocks.10.mlp.proj.weight: 15.71
+[05:22:25] Quantized weights:
+[05:22:25]   gptq (int6): blocks.attn.c_k.weight, blocks.attn.c_q.weight, blocks.attn.c_v.weight, blocks.attn.proj.weight, blocks.mlp.fc.weight, blocks.mlp.proj.weight
+[05:22:25]   gptq (int8): tok_emb.weight
+[05:22:25]   passthrough (float16): blocks.attn.q_gain, blocks.attn_scale, blocks.mlp_scale, blocks.resid_mix, lane_merge, skip_gates, skip_weights
+[05:23:17] Serialized model quantized+brotli: 15889177 bytes
+[05:23:17] Total submission size quantized+brotli: 15909023 bytes
+[05:23:33] quantized val_loss:2.84044015 val_bpb:1.09964597 eval_time:8133ms
+[05:25:04] quantized_sliding_window val_loss:2.79653431 val_bpb:1.08264830 eval_time:90581ms
diff --git a/records/track_10min_16mb/2026-04-17_SP8192_AdaptiveHessianClip/train_seed999.log b/records/track_10min_16mb/2026-04-17_SP8192_AdaptiveHessianClip/train_seed999.log
new file mode 100644
index 0000000000..7b5c40c5d0
--- /dev/null
+++ b/records/track_10min_16mb/2026-04-17_SP8192_AdaptiveHessianClip/train_seed999.log
@@ -0,0 +1,205 @@
+W0417 05:26:27.035000 43583 torch/distributed/run.py:803] 
+W0417 05:26:27.035000 43583 torch/distributed/run.py:803] *****************************************
+W0417 05:26:27.035000 43583 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. 
+W0417 05:26:27.035000 43583 torch/distributed/run.py:803] *****************************************
+[05:26:39] Hyperparameters:
+[05:26:39]   adam_eps: 1e-08
+[05:26:39]   adam_wd: 0.02
+[05:26:39]   beta1: 0.9
+[05:26:39]   beta2: 0.95
+[05:26:39]   compressor: brotli
+[05:26:39]   data_dir: ./data/
+[05:26:39]   datasets_dir: ./data/datasets/fineweb10B_sp8192
+[05:26:39]   distributed: True
+[05:26:39]   ema_decay: 0.9965
+[05:26:39]   embed_bits: 8
+[05:26:39]   embed_clip_sigmas: 20.0
+[05:26:39]   embed_lr: 0.6
+[05:26:39]   embed_wd: 0.085
+[05:26:39]   embedding_dim: 512
+[05:26:39]   enable_looping_at: 0.35
+[05:26:39]   eval_seq_len: 2048
+[05:26:39]   eval_stride: 64
+[05:26:39]   gptq_calibration_batches: 64
+[05:26:39]   gptq_reserve_seconds: 12.0
+[05:26:39]   grad_accum_steps: 1
+[05:26:39]   grad_clip_norm: 0.3
+[05:26:39]   head_lr: 0.008
+[05:26:39]   is_main_process: True
+[05:26:39]   iterations: 20000
+[05:26:39]   ln_scale: True
+[05:26:39]   local_rank: 0
+[05:26:39]   logfile: logs/3778f8e1-66f0-4f01-8365-f8ccc9f22d81.txt
+[05:26:39]   logit_softcap: 30.0
+[05:26:39]   loop_end: 5
+[05:26:39]   loop_start: 3
+[05:26:39]   matrix_bits: 6
+[05:26:39]   matrix_clip_sigmas: 12.85
+[05:26:39]   matrix_lr: 0.022
+[05:26:39]   max_wallclock_seconds: 600.0
+[05:26:39]   min_lr: 0.0
+[05:26:39]   mlp_mult: 4.0
+[05:26:39]   model_dim: 512
+[05:26:39]   model_path: final_model.pt
+[05:26:39]   muon_backend_steps: 5
+[05:26:39]   muon_beta2: 0.95
+[05:26:39]   muon_momentum: 0.99
+[05:26:39]   muon_momentum_warmup_start: 0.92
+[05:26:39]   muon_momentum_warmup_steps: 1500
+[05:26:39]   muon_row_normalize: True
+[05:26:39]   muon_wd: 0.095
+[05:26:39]   num_heads: 8
+[05:26:39]   num_kv_heads: 4
+[05:26:39]   num_layers: 11
+[05:26:39]   num_loops: 2
+[05:26:39]   parallel_start_layer: 7
+[05:26:39]   qk_gain_init: 5.25
+[05:26:39]   quantized_model_path: final_model.int6.ptz
+[05:26:39]   rank: 0
+[05:26:39]   rope_base: 10000.0
+[05:26:39]   rope_dims: 16
+[05:26:39]   rope_train_seq_len: 2048
+[05:26:39]   run_id: 3778f8e1-66f0-4f01-8365-f8ccc9f22d81
+[05:26:39]   scalar_lr: 0.02
+[05:26:39]   seed: 999
+[05:26:39]   skip_gates_enabled: True
+[05:26:39]   sliding_window_enabled: True
+[05:26:39]   tie_embeddings: True
+[05:26:39]   tied_embed_init_std: 0.005
+[05:26:39]   tied_embed_lr: 0.03
+[05:26:39]   tokenizer_path: ./data/tokenizers/fineweb_8192_bpe.model
+[05:26:39]   train_batch_tokens: 786432
+[05:26:39]   train_files: ./data/datasets/fineweb10B_sp8192/fineweb_train_*.bin
+[05:26:39]   train_log_every: 500
+[05:26:39]   train_seq_len: 2048
+[05:26:39]   val_batch_tokens: 524288
+[05:26:39]   val_files: ./data/datasets/fineweb10B_sp8192/fineweb_val_*.bin
+[05:26:39]   val_loss_every: 4000
+[05:26:39]   vocab_size: 8192
+[05:26:39]   warmdown_frac: 0.72
+[05:26:39]   warmup_steps: 20
+[05:26:39]   world_size: 8
+[05:26:39]   xsa_last_n: 11
+[05:26:39] train_shards: 128
+[05:26:39] val_tokens: 40540160
+[05:26:41] model_params:35944537
+[05:26:41] gptq:reserving 12s, effective=588000ms
+[05:26:55] warmup_step: 1/20
+[05:27:02] warmup_step: 2/20
+[05:27:05] warmup_step: 3/20
+[05:27:05] warmup_step: 4/20
+[05:27:05] warmup_step: 5/20
+[05:27:05] warmup_step: 6/20
+[05:27:06] warmup_step: 10/20
+[05:27:07] warmup_step: 20/20
+[05:27:07] loop_warmup:enabled encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10]
+[05:27:17] loop_warmup_step: 1/20
+[05:27:17] loop_warmup_step: 2/20
+[05:27:18] loop_warmup_step: 3/20
+[05:27:18] loop_warmup_step: 4/20
+[05:27:18] loop_warmup_step: 5/20
+[05:27:18] loop_warmup_step: 6/20
+[05:27:19] loop_warmup_step: 10/20
+[05:27:20] loop_warmup_step: 20/20
+[05:27:26] 0/20000 val_loss: 9.0075 val_bpb: 3.4872
+[05:27:26] 1/20000 train_loss: 9.0020 train_time: 0.0m ms/step: 94 tok/s: 8387944
+[05:27:26] 2/20000 train_loss: 12.1542 train_time: 0.0m ms/step: 96 tok/s: 8225299
+[05:27:26] 3/20000 train_loss: 10.9154 train_time: 0.0m ms/step: 97 tok/s: 8109904
+[05:27:26] 4/20000 train_loss: 9.3420 train_time: 0.0m ms/step: 98 tok/s: 8057998
+[05:27:26] 5/20000 train_loss: 8.2516 train_time: 0.0m ms/step: 98 tok/s: 8027543
+[05:28:16] 500/20000 train_loss: 3.3440 train_time: 0.8m ms/step: 101 tok/s: 7783905
+[05:29:07] 1000/20000 train_loss: 3.2346 train_time: 1.7m ms/step: 101 tok/s: 7793096
+[05:29:57] 1500/20000 train_loss: 3.1938 train_time: 2.5m ms/step: 101 tok/s: 7793722
+[05:30:48] 2000/20000 train_loss: 3.0811 train_time: 3.4m ms/step: 101 tok/s: 7788322
+[05:30:52] layer_loop:enabled step:2039 frac:0.350 encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10]
+[05:32:00] 2500/20000 train_loss: 3.0019 train_time: 4.6m ms/step: 110 tok/s: 7165990
+[05:33:14] 3000/20000 train_loss: 3.0045 train_time: 5.8m ms/step: 116 tok/s: 6767854
+[05:34:29] 3500/20000 train_loss: 2.8986 train_time: 7.0m ms/step: 121 tok/s: 6509839
+[05:35:43] 4000/20000 train_loss: 3.0045 train_time: 8.3m ms/step: 124 tok/s: 6329113
+[05:35:50] 4000/20000 val_loss: 2.8751 val_bpb: 1.1131
+[05:37:04] 4500/20000 train_loss: 2.8267 train_time: 9.5m ms/step: 127 tok/s: 6195906
+[05:37:23] 4614/20000 val_loss: 2.8073 val_bpb: 1.0868
+[05:37:23] stopping_early: wallclock_cap train_time: 588078ms step: 4614/20000
+[05:37:23] peak memory allocated: 39716 MiB reserved: 39742 MiB
+[05:37:23] ema:applying EMA weights
+[05:37:30] pre-quantization post-ema val_loss:2.80418200 val_bpb:1.08560902 eval_time:6679ms
+[05:37:30] Serialized model: 135431355 bytes
+[05:37:30] Code size: 19846 bytes
+[05:37:30] GPTQ:collecting Hessians from calibration data...
+[05:37:43] GPTQ:collected 67 Hessians in 12.7s
+[05:37:50] GPTQ adaptive clip_sigmas (Hessian-sensitivity):
+[05:37:50]     tok_emb.weight: 20.00
+[05:37:50]     blocks.0.attn.c_q.weight: 11.11
+[05:37:50]     blocks.0.attn.c_k.weight: 10.98
+[05:37:50]     blocks.0.attn.c_v.weight: 10.77
+[05:37:50]     blocks.0.attn.proj.weight: 11.72
+[05:37:50]     blocks.0.mlp.fc.weight: 10.61
+[05:37:50]     blocks.0.mlp.proj.weight: 7.91
+[05:37:50]     blocks.1.attn.c_q.weight: 12.49
+[05:37:50]     blocks.1.attn.c_k.weight: 12.18
+[05:37:50]     blocks.1.attn.c_v.weight: 12.02
+[05:37:50]     blocks.1.attn.proj.weight: 12.08
+[05:37:50]     blocks.1.mlp.fc.weight: 11.70
+[05:37:50]     blocks.1.mlp.proj.weight: 10.48
+[05:37:50]     blocks.2.attn.c_q.weight: 13.23
+[05:37:50]     blocks.2.attn.c_k.weight: 12.73
+[05:37:50]     blocks.2.attn.c_v.weight: 12.63
+[05:37:50]     blocks.2.attn.proj.weight: 14.53
+[05:37:50]     blocks.2.mlp.fc.weight: 12.42
+[05:37:50]     blocks.2.mlp.proj.weight: 11.55
+[05:37:50]     blocks.3.attn.c_q.weight: 11.80
+[05:37:50]     blocks.3.attn.c_k.weight: 11.44
+[05:37:50]     blocks.3.attn.c_v.weight: 11.21
+[05:37:50]     blocks.3.attn.proj.weight: 12.20
+[05:37:50]     blocks.3.mlp.fc.weight: 11.09
+[05:37:50]     blocks.3.mlp.proj.weight: 10.61
+[05:37:50]     blocks.4.attn.c_q.weight: 12.21
+[05:37:50]     blocks.4.attn.c_k.weight: 11.86
+[05:37:50]     blocks.4.attn.c_v.weight: 11.54
+[05:37:50]     blocks.4.attn.proj.weight: 11.97
+[05:37:50]     blocks.4.mlp.fc.weight: 11.47
+[05:37:50]     blocks.4.mlp.proj.weight: 11.95
+[05:37:50]     blocks.5.attn.c_q.weight: 12.64
+[05:37:50]     blocks.5.attn.c_k.weight: 12.01
+[05:37:50]     blocks.5.attn.c_v.weight: 12.03
+[05:37:50]     blocks.5.attn.proj.weight: 13.40
+[05:37:50]     blocks.5.mlp.fc.weight: 11.76
+[05:37:50]     blocks.5.mlp.proj.weight: 11.45
+[05:37:50]     blocks.6.attn.c_q.weight: 15.04
+[05:37:50]     blocks.6.attn.c_k.weight: 14.59
+[05:37:50]     blocks.6.attn.c_v.weight: 14.41
+[05:37:50]     blocks.6.attn.proj.weight: 16.49
+[05:37:50]     blocks.6.mlp.fc.weight: 14.12
+[05:37:50]     blocks.6.mlp.proj.weight: 14.32
+[05:37:50]     blocks.7.attn.c_q.weight: 15.31
+[05:37:50]     blocks.7.attn.c_k.weight: 14.80
+[05:37:50]     blocks.7.attn.c_v.weight: 14.42
+[05:37:50]     blocks.7.attn.proj.weight: 14.85
+[05:37:50]     blocks.7.mlp.fc.weight: 14.40
+[05:37:50]     blocks.7.mlp.proj.weight: 15.69
+[05:37:50]     blocks.8.attn.c_q.weight: 16.01
+[05:37:50]     blocks.8.attn.c_k.weight: 15.21
+[05:37:50]     blocks.8.attn.c_v.weight: 14.26
+[05:37:50]     blocks.8.attn.proj.weight: 18.82
+[05:37:50]     blocks.8.mlp.fc.weight: 14.65
+[05:37:50]     blocks.8.mlp.proj.weight: 16.30
+[05:37:50]     blocks.9.attn.c_q.weight: 16.16
+[05:37:50]     blocks.9.attn.c_k.weight: 15.52
+[05:37:50]     blocks.9.attn.c_v.weight: 14.57
+[05:37:50]     blocks.9.attn.proj.weight: 17.74
+[05:37:50]     blocks.9.mlp.fc.weight: 14.85
+[05:37:50]     blocks.9.mlp.proj.weight: 16.75
+[05:37:50]     blocks.10.attn.c_q.weight: 16.09
+[05:37:50]     blocks.10.attn.c_k.weight: 15.59
+[05:37:50]     blocks.10.attn.c_v.weight: 15.52
+[05:37:50]     blocks.10.attn.proj.weight: 17.76
+[05:37:50]     blocks.10.mlp.fc.weight: 15.00
+[05:37:50]     blocks.10.mlp.proj.weight: 15.72
+[05:37:50] Quantized weights:
+[05:37:50]   gptq (int6): blocks.attn.c_k.weight, blocks.attn.c_q.weight, blocks.attn.c_v.weight, blocks.attn.proj.weight, blocks.mlp.fc.weight, blocks.mlp.proj.weight
+[05:37:50]   gptq (int8): tok_emb.weight
+[05:37:50]   passthrough (float16): blocks.attn.q_gain, blocks.attn_scale, blocks.mlp_scale, blocks.resid_mix, lane_merge, skip_gates, skip_weights
+[05:38:40] Serialized model quantized+brotli: 15891689 bytes
+[05:38:40] Total submission size quantized+brotli: 15911535 bytes
+[05:38:51] quantized val_loss:2.84091481 val_bpb:1.09982973 eval_time:8378ms
+[05:40:22] quantized_sliding_window val_loss:2.79680127 val_bpb:1.08275165 eval_time:90539ms