diff --git a/records/track_10min_16mb/2026-04-05_CautiousMuon_SP4096_DepthRecur/README.md b/records/track_10min_16mb/2026-04-05_CautiousMuon_SP4096_DepthRecur/README.md
new file mode 100644
index 0000000000..1156bb3f4c
--- /dev/null
+++ b/records/track_10min_16mb/2026-04-05_CautiousMuon_SP4096_DepthRecur/README.md
@@ -0,0 +1,74 @@
+# Cautious Muon + SP4096 + Depth Recurrence + Parallel Residuals
+
+**val_bpb = 1.1604** (3-seed mean, std = 0.0033)
+
+This is a non-record submission exploring the effect of Cautious Muon (arXiv:2411.16085) on the PR #1334 architecture stack.
+
+## Results
+
+| Seed | val_bpb | val_loss | Artifact Size |
+|------|---------|----------|---------------|
+| 42   | 1.1568  | 2.6619   | 15,179,504 B  |
+| 314  | 1.1611  | 2.6717   | 15,173,470 B  |
+| 999  | 1.1634  | 2.6770   | 15,159,223 B  |
+| **Mean** | **1.1604** | **2.6702** | **15,170,732 B** |
+
+## Key Technique: Cautious Muon (arXiv:2411.16085)
+
+The primary modification is applying the Cautious optimizer principle to the Muon optimizer. The optimizer pipeline is:
+
+1. Compute momentum-corrected gradient (Nesterov)
+2. **MuonEq-R row normalization** — normalize rows before Newton-Schulz
+3. **Newton-Schulz orthogonalization** (5 steps) — project onto orthogonal manifold
+4. **Cautious masking** — only apply update where the orthogonalized direction agrees with the raw gradient sign:
+
+```python
+caution_mask = (g * raw_grad > 0).to(g.dtype)
+g = g * caution_mask / caution_mask.mean().clamp_min(1e-3)
+```
+
+This filters out "stale" momentum directions that disagree with the current gradient, providing ~1.47x effective convergence per step with zero parameter overhead and no impact on artifact size.
+
+## Full Architecture Stack
+
+Built on PR #1334 (@aryanbhosale) with:
+- **SP4096 BPE tokenizer** (from PR #1218, @clarkkev)
+- **Depth recurrence** layers 4,5 (13 virtual layers from 11 physical, activated at step 3000)
+- **Parallel residuals** from layer 7 (separate attn/MLP lanes with learnable merge)
+- **MuonEq-R** row normalization before Newton-Schulz orthogonalization (arXiv:2603.28254)
+- **Cautious masking** applied after Newton-Schulz orthogonalization
+- **QK-Gain 5.0** per-head query-key scaling
+- **EMA 0.997** weight averaging
+- **Full GPTQ INT6** quantization with selective +-1 pruning
+- **Brotli compression**
+
+## Non-matrix parameters
+
+Token embeddings, scalar parameters, and head use standard `torch.optim.AdamW`. Cautious masking is applied only inside Muon for matrix parameters.
+
+## Compliance
+
+- Track A fixed predictor -- no TTT, no SLOT, no eval-time adaptation
+- All predictions are causal and normalized via softmax (F.cross_entropy)
+- Artifact under 16MB limit (max 15,179,504 bytes)
+- Training completes within 600s wallclock on 8xH100 SXM
+
+## Reproduction
+
+```bash
+cd /workspace/parameter-golf
+# Download SP4096 data
+MATCHED_FINEWEB_REPO_ID=kevclark/parameter-golf python3 data/cached_challenge_fineweb.py --variant sp4096
+# Run
+SEED=42 torchrun --standalone --nproc_per_node=8 train_gpt.py
+```
+
+## Decompressed Source
+
+The `train_gpt.py` is a self-extracting compressed script (`exec(lzma.decompress(base64.b85decode(...)))`). The decompressed source is ~1991 lines of Python. The only modification to the PR #1334 base is the two-line Cautious Muon mask inside `Muon.step()` (lines ~869-872 of the decompressed source).
+
+## Credits
+
+- PR #1334 (@aryanbhosale) -- base architecture (SP4096, depth recurrence, parallel residuals, MuonEq-R)
+- PR #1218 (@clarkkev) -- SP4096 tokenizer
+- Liang et al. (arXiv:2411.16085) -- Cautious Optimizers
diff --git a/records/track_10min_16mb/2026-04-05_CautiousMuon_SP4096_DepthRecur/submission.json b/records/track_10min_16mb/2026-04-05_CautiousMuon_SP4096_DepthRecur/submission.json
new file mode 100644
index 0000000000..c1710c95dd
--- /dev/null
+++ b/records/track_10min_16mb/2026-04-05_CautiousMuon_SP4096_DepthRecur/submission.json
@@ -0,0 +1,39 @@
+{
+    "author": "X-Abhishek-X",
+    "github_id": "X-Abhishek-X",
+    "name": "Cautious Muon + SP4096 + Depth Recurrence + Parallel Residuals",
+    "blurb": "Non-record submission. Applies Cautious Muon (arXiv:2411.16085) to the Muon optimizer — masks Newton-Schulz updates where the orthogonalized direction disagrees with the raw gradient sign, providing ~1.47x effective convergence per step with zero parameter overhead. Built on PR #1334 (aryanbhosale) base with SP4096 vocabulary, depth recurrence (layers 4,5), parallel residuals (from layer 7), MuonEq-R, QK-Gain 5.0, and full GPTQ INT6 + Brotli compression. Mean val_bpb = 1.1604 (3 seeds, std = 0.0033).",
+    "date": "2026-04-05T00:00:00Z",
+    "track": "10min_16mb",
+    "val_loss": 2.67020395,
+    "val_bpb": 1.16043988,
+    "val_loss_std": 0.00764948,
+    "val_bpb_std": 0.00332438,
+    "seeds": [42, 314, 999],
+    "seed_results": {
+        "42": {
+            "val_loss": 2.66190192,
+            "val_bpb": 1.15683191,
+            "artifact_bytes": 15179504
+        },
+        "314": {
+            "val_loss": 2.67174312,
+            "val_bpb": 1.16110878,
+            "artifact_bytes": 15173470
+        },
+        "999": {
+            "val_loss": 2.67696681,
+            "val_bpb": 1.16337894,
+            "artifact_bytes": 15159223
+        }
+    },
+    "artifact_bytes_mean": 15170732,
+    "artifact_bytes_max": 15179504,
+    "bytes_total": 15179504,
+    "bytes_code": 24659,
+    "hardware": "8x H100 SXM (RunPod On-Demand)",
+    "pytorch_version": "2.9.1",
+    "cuda_version": "12.8",
+    "technique_summary": "Cautious Muon optimizer (arXiv:2411.16085), SP4096 BPE tokenizer, depth recurrence layers 4-5 (start step 3000), parallel residuals from layer 7, MuonEq-R row normalization, QK-Gain 5.0, EMA 0.997, full GPTQ INT6 quantization with selective pruning, Brotli compression",
+    "comparison_baseline_pr": 1334
+}
diff --git a/records/track_10min_16mb/2026-04-05_CautiousMuon_SP4096_DepthRecur/train_gpt.py b/records/track_10min_16mb/2026-04-05_CautiousMuon_SP4096_DepthRecur/train_gpt.py
new file mode 100644
index 0000000000..90131a9176
--- /dev/null
+++ b/records/track_10min_16mb/2026-04-05_CautiousMuon_SP4096_DepthRecur/train_gpt.py
@@ -0,0 +1,2 @@
+import lzma as L,base64 as B
+exec(L.decompress(B.b85decode("{Wp48S^xk9=GL@E0stWa8~^|S5YJf5;Y&G8oLvAkn@VT6Qap3bu0*kgCR~YUqB0W9R)iarr*QtEZpesGY3>~CZRiK|6Dwut$nH#N!RYqQnA}G^`ZsFO;ar92)Xt#3E3Ki5S1}OfSx<=$c<4=h|J{kt$27^CQ01M+lVgZ0tGgX0&I*V@{U&JgYc0U!(4F-btCy*+qzv6Dp~UW!y~6{U*}y$E@2-R}vd?t*s#fnDO{!j>OImt34A(d+9n>hnnvzmd((<?7SwGA+W$Or}R^?z{Ree`d98K#2b?5_cfZ%Lt17{dO-GN2&5;3ECS=cE#O*hTiZtOgX4XIkBuR{PAJ5$6e;Eel8-VVQFtsWBkSb@w{YAVOHKclDdD}`(Z5P0z`*Zpr!j8psZi3dj0Awg*|SM}VUSWz|5#6=0zD3_4d4q0&x{n`YrJE1HXz*`$RPocfD=uX!IiV{O9O7|!NUnrgGHr-RntBL^weTkpGrkcD-11{fi$WX(*bn-nbgD6YkvCezcUG_I`KWJk#hl}8>_D1Cghg~(bQ$Yj)>!Y%{*o9ex8FWa#U)!OI!!5Prl^?bnBX2V(=(Bvc+CvGo!S{LhLn7pSsR!}@U=OBW0)h6IYneQ1{|$<&k9TS^qGQpb-;#vEPAl%11UF)?6mtC8c04XzR$+h2=j84E2|i`pOEt$uyM`lGs*ejIF-}^SvSRZK$ePh1`gt+?%1r#=OV<HVjrIqeJYGxh%pSpPlDdQ66RHQe_t;x^R`DvG*6PkO>y3pW`{ofc)6PhQQRP|_h56zl+sQ(le1eJ^&&qZxdGb15aOb^-R1ouqi-H1_w|H;g(()bKz_!0+L{#HFmtQSw%~n|MX3ij_2{lW(_*6gdIz`XT%tzkhK-k5tAu}`=>u|z+<A5OWlS!1&|7BshLa`v33WEyuh5cV6d%c)Zp^vZ`+%XW7Ga$QNikdrw8}teN{lfKFYW8jLKpUE<fS0HI?!#S}z5(PnD*$%*&e)`J?Z?l)B@7}E*EtBQ<ajK2J2UVwA>|uP4UnMNw^{d&KAm;P6`40&zphh*D=e*8?KGZuo~y*`y#Wg}r(PV<PiL09`OF@Nh{a<w{zds*!NOI%lISf!q#(&1k8wXKJfU5I>}?J$O<io<ujGSX9NWFVj^&nmc<AiUyr-`}m?om^MsCAZ)88TX4@F=Gr{ba+%8mF)D2K;PS)r}K`0OW1F4MRM)*s7t_jmGr27SW%DBvhmeslZ(N-kfEgbZN}VIhl%?M7XjWH8wf?lo_X|5_oRRVxUolXyMc*~(GxY5pkyJ-C9J))sNjfO=O;m+FUfndz~U9#}!)KfJsF`v`a^5o$1$QiNNu!W#ZYR)zDNo?=6w(J-Mr%~Ccd@JvU7j6*UqI1G_U5)xIOVfFefW#-2fZay)3C_@H2WgGU<<j3Y(=kG=f*7C;k&v(6y7C}?)ki&L~IiRiQ(NiyD1>ae&vv%2(eb|cSn}oTUc1&B%M^B@xlvn10$Ol5{Rc(||g*`0AO8zk14;L<B89tNMDJH<nBP9kHyG#H@%Y{;9I%*-8=cLIDmul}ETySn#xzQAAJ)H3oDC_09q<t7H;T?vgXJnFe2W2eEM#4@zx^{Q`RI4rmXYqJ`j>7H$%dzpg>)#q=gM_#uqc-I?vb;GXT^_H=R{GSK>=VuvqSe%#z|2c2YlT&*kjZ2;Kl!jcm$Q-fVo|ict&Ja4ywb`i8dvCk+h?3OCDUXX&zMsrr>Pa>PMm^gxGuU~bBkga*GhQ$y8PeK(vV5>_veh8WQRi`fT<y2?%qa)RDFxTOGTC_Yl#}yg8C;;?WCn_iwWXDzN03qd_<{4@#oI{v^Qp5Ta*t1IOY_9!%73?T>)Py)OBaHdkXS7+2I5uh#w2)7T@22C(Qzp-K-tVkOU|oncF?OgMCg#5sW{yA4RZO!84eUpjR$IRG8(QIv>WqY`&obF`+L^a;Ik$ZPqJ0w+oVi8+fefP_6}>2vq29M(7pWSxwX%h;ULwd}`?78ERR(NNJ_4KVp|BN9>arZL(Qm544oVRUtWw?7Jk|KZ1C{%$Jv!T6;U}wiQVVE)4<71$jhkVz<Qtn`)a82WW$q0Hr{4+UPsbm$mWoflR~5=rA+eFXgkwC=?_lCr!4xuwO>^+^AaNS!~_W(vR4{WJ^svyxk^G3sz?R<rC;XSxpV71HMoESHR0<s%rK-;z}#rRPgX{3Q(A0Wp$2KXLI=1lOC(?@TlnLPSgNk$`a3`*6ksGvvQDcbXT5`hv<KTw5Y2&yEeu8k!Vh|_EpHFWU(bfUMSirttllpOsLyWqmB4^^^xE{p{jc~83a5_Lv6FW;@4UlDMPfv?~X~7+CW&3H+kc&D_4ToujfL`8*93hUr+a~GGo0xNE~Rg0>Yo_N1#V&d43t{_A|@#vLRTXJ)mh_LsXCAJVuNHeI8p77$-9H%-!CvcwCI8NI<;c*6{I5uPqIoTdj3Xcr{R+d)6lXf`Tkt(h3PL4|paL6IHOjh`kZhHjzRo&SJapcs5d^H130}9Do<&`*(lN1e>3UmSuApd3~p1>+iw(L!(EeP^xfD*_e|;Pwu$5EVQsNPN$X%#{QH-cBp=0SJ2S&fh>5{xov2Q&Dt8e2OD|)&mt5}pMa#y6QbXmJX!S1_u;C#vl1c(#HwFUIq)H-{6h49qdpK8DkSGDi&|0Gm*`a@fNS(bc?SbL#Nw1qWH_Sm8H<BSYbB+7>5d?bot%W2!a%_xL$l$wZMM<P2S6~e+Ooc>0%5SQBB#7p$O|$EKTZ^s?k{J%APQfIG>nefD{BM8Wv~Raz5FG};~ZrIG9BYxbU)6FnJ8gyh&Uq4Ujyc@m~D|>5=W&d7!2;#A`)#%c)a{*uY@1mteo&`!#O*c-WiWMu#FM+$tI&t-Vk$6T+lan2}MAv`?+7wse?lKLKyL?o?r>aEE}`ezoaZ6fF6dUE2tY2^trIgC4f5fCd!?N^%$XSL$_1XK0UJ!A@|%cDxnG}8w&B3$6`#1#KhW&Ql2C)5ib>m5Z=M-qv1|o$Q0|W(_^uCKX09E!%ojzZI?^Mz~TizB4o%X)QO_ql@26ymjSb3)uBO=dFg!x-)Qn;kD+UVD&>jk<8!}Ec-D__4+R~K1U5lNK|w2cixht+FS$0V>V>SzHs(Zk;OrtRtI3kcE`ureK?716@eYVye|+&bYUx-7g@AdF@_<77`U8(GXS0!&D^P-u;Xz}0WKVcbvGrEFSfxcV>TP{Zc-rn<riUyaMV`Pg_`~y%0<DnQZefADy^d3T|Fb4)EJH7Da9D*%i)O4~&hH^su=$@?2oQ_!TUNFuhg}a5zjTC?yRG$bdHLuKhKnNF`ksq?@{fV6I+4mIUygtt_wI%hETo_9<%eR}B^m>}A78K>*`a+rhV?lldBS^}%@}u<;Dvw<?Gc}<Wygn$lO0Dj)J4<~x>Qjfuz0#=9ImWl*WoNvxZqR-cYjGYAGA;O^VeTt+L*r29&lp&rMOX=u*JS$iZmC|mYP!1l_3fkt4g*|<MTxOq(euNH(JFt+d(QKMI?zE`8|xAPT8z>qc@{!A}d!ZdZdq-LX&)z*#yYN+$?Bqe|01W`-A|R{90OA;BuKH7-D8K#4(MtnL@`>u%l)7Wo*3m++MZ63TBcQJ>;K$x;rQt+>)>cEy7f1sq-E4<n`1+Up0H0*+B5qruY?(p~nhcRcp0Y<>-P=o=ctbBsP$HEE{Q{GTSMN5dk@gm=-fy<)o05o|SSOL_PN>QA5sw1_nYpza?AifEtB`gGmk_a_KAt7OI9}fdzZ_1<re)5zJwY>%B7De4O;fOO3G|`*hODWE5!{={Lx~<E%MJ<r_R!VkF85otI5gE?oa_d}oX!6a&Ix3mK`G;I9q*OO>J%N4q)qhTlK<c(3(FqTL8DLxXNLMoD5p?`8T1?m0G+SCBSl+4tkK0gruwl-=7j`1t9K;nxuA_f}P#KtZ?SRv}H(zN2^}SwNP2+$W`<Su}6k;sB#M<Xb@*Vu}TA-FpGD#+~J)ZA?kucvS*G`F(*2HiFm(e2q#KMtIA`-I)g-IGfjTE{-eU)P>;^Vw9bHSA_Rjjp~NFEtw76&M^}{dD!n3d^>P(=?OwyW56Vh-Y=t$TUe%*lb*hV-ny<{h?{d1B!^>a_Q*8%Kdqz9d2T04mMyR?`RR|!?x65zhQKdmwv;0x;sm38pzc6?(-obP9jShOmmBDa3go7BP9vbs9e&i9G9It6$5xhA7WL-OStpny<`(&kVy4?EY97TS{CRW^|JV3P3MOU;M}3+-DhEnhy3!eIqtF+S-{R4s7R6+d;<+>uM>@JZqKZpVUC=+lYb!!T&vJQ_3;B(#xqMak-T8X!C7Y3r#O#;_VW-k*e#2c@cGI4hYrF*;PJq<v2)#o*K!M+`;Y=}R_BBZrSx1Hyg@>%|7492kH1(f@a^mAw^}9vO?<RWoU0{NxgAM%io;0Tsa_a2V#~aS*!I;#<f<Q08Q67gIzL7B||Nepa@^*;7pdD25wTM^~zz-Q0iSjlR@a%yFv(!7tJDoU;ALlNJzvm={V0i*T`$u{%|2XX@L-%9vj(W`s2c#Y7wTx5z4lFl7L>3(2y=4XxWf9{jAevrBh<|B#@>K#!Hj>cVI#kk^TT{*aP}pUue5rsHq(8L+tdb}!yS8}4mQ*vPU%%T>@i_qMFbLb#8SrKL(y*BcH_j`YW)jm|T1MSK=p0r8lb8HXOC#uTA@^z!q4|?60<sx5#i)CrY;b!Lh!+-CdnXiBmuZAfIHedsX9WV+IizyAlz0{S$Eh0W&k<@)ubL1D>;qd!ah5v9(Nj8;|4c%vp9p0(MOWt_SM3Z@xfoG$eFn%?m)Ewo{?^?%<SwrQ^^8=>@mxbUbKIBry{Rl4fN>E4f<&q?gs;fV9T4u(<QIJ3E4~W2fs-2}V-OI&y!M_6HHmh7cu(r(c2@&Yk@xBGFu2^xekaTWw5a6y;#qFvFdlLA)%w8NDQqN~^;hk7Q4+K;RVu$=@6B>+c2s;SK)0GPc9EQp&loO%TJOhA4v<l`MPKt?$P8uHwl2De<8sBx<y&soWr#X#X%e4QDcs~MP{$oj<QwGXr=^M`{SseWjm6KHY6nBgFKFqiJ~;%6a<9%*jBohCey_XFoEO_iz=;3oM)iA)8Gd;fzcnV~BQ}L$o~_eP8j$oYCP3a)f-rsyGsu*3rqiP8)fcF#=gm3V(2II)Q;0%Q0+y4VPE{DM&NjzF3r6rj^HP$5yrTqibwFgRB0R1@$FRH>1N9qwBbGv`rJ^aCf7kmC-=jfuipSBg6u15TPS61~aBaYw-4Prj*JmPofD{@U68LcjjN~8kBju#C%QXlBhn0MBialPcUz#yBW?Zt~=)J&lw9tsUc^N|mtb^%JYFY2W13!k1KuU?j7XyhzV?#G2;gCs%Zm(iwj2p|VujV<~EGevXYPk5X%8w(_gPL&n0zAHVckP}#);*Xh;mO58ylqV`ZC6wO%L;v|`lL#+2)>teaL-;>%7T}VR@5yN4cl1*?w&|^WMNODQYk6`wk+DhUoiQzwI>y^eSjxVVg6+F6hOjKO~fDDG!y+o_*|8~{-o#Pw>P{qe4;KQ`Rs__H^q`xDV!w#&udFw$=tAjCBbBb*su`G&DDR``?6N}$lV8pjuG?3g8gnwn?q5ax?f513pn?ikJAp6*kZI>*f2U9?6tbKV+);v3=^Pg8o^U7gfQvB+7ix0>uMtUS+6C%1>y`)#GbjVDTP2;2^d2UGZgpB%09^<jQhHd36ni+*;nCJ30>vS#`J`Ezq_UQN}^0W=SHy4EKXmYhp=7HdK1bmv;|9;Q8V<wMd5zludwR;V^ifZFDnK1|BsOseGV<1ZoS^USrD;zk69GPQ<de(rRT?S?fK&1Hjpk}v9)Dc3>W4CDnMI6gAuD@RF1i5&4-V#f#5v#(Co$06y`*OkgGr#-JIW2n9Nf3azCs+jI<^h$>jKVhO*8_6s)dcRHzwS&VnERy_r3KW!w1!5wA^R+GRy?bQR1RTM7ce#4@k{4qt^hoam%JU)|NXu91Gz>+Tiyo1p&pQ;C)O*w??j>_q}RpBFq0rX>7FujB@8=e9f}5yJ<ib39$#8aRTX7{~8a{GqsbsGzDVbj&pp_A=Sr>|!L&*+cx0?M(g(%c0g1R%n2?$@m$~oJ;xH#bzP(pzwm@-WzsL-btFU##Pqj?H$yQP@j;Dt=dnyc7!OYUVL=aD!I3HbP;3oT43ai-YjpXM06`E`_IzF-F~}3_&~=M&c&26Z~vL;DtfZ{Vyb<v<`v=2Ee(mkIQb!9#&fwUOgV$yr4Y#j1w9)O@?nZZQIAwP{D9kE1fGCpZ;EMk0{N|rQJIXpBkd>L>U>(1^f{=xv7Qcga!kF2$<qEs$ceg`(=Upg(?oro2zNYX#X-&_R2^DKi(T2Q?1$jEi}qg&6>8=KhH~AiaAbPKX$;G7&OpN%z(X--EjXO9%h#-%`8dP&P<8-Y&;I2`f%XnQg!$NKLlK!Uk&1NSS?X$iFhfNf^Y!B82w~(EGFYZ&t;R^5Vnh7Lui5tdS*!kJaRLDXmj_^r<JxvJ(H?PZQ%ljiJJicxuFdcIrVQ-h8gXU$;qAUTI5e$EGBqH|MH3@7xXs;}uj(Q6Q!-LALCew~H&fvmjBq%#Z(&bgxRHC~M3W|p4LC^ykvI=NhB^l#vD*ZGBbOTbJGBHmFKc#(|BNpC;~J7;(Nz#`mZe~g91kVvIV3(9DV^fUAc0(kZLTRAH2~4bh9T(<gN8np0cb-7##&}=Is#?9;^`dOp_?=;*u3fan|oksbR6eIqjAeOODU^0SOO|8+%(+yF2%c)Y~?zUth_*-nt2T?-d6Pv`LNjdbU*{W-LBXIAmr2TVHI^a@Ym})qhUoy96tDS+n}}qgM?T_=^eENcB|SH1cT4Pp~UN)vzib8)kpKY9;nheP0zBfQuvSjnS5ftZXJ}OLuTy51(PGM{4Om2t)TR8pdqX3hyz$6XCbXQ4GM4n5v_Qe%K8H1-3?57*vqU`u3)H<qieUWU&Ej0AJc`KE5kR-NCj4;7nt40(mioh8kM0#vVyf&=zzR(9})dBS336j*bE}WE0_b^tEz8ca^lr{!LbMi2vTH!euD1SeJkl7rkijwuODZuT`XJoMO-*p*G+B+?#O#)AURV%j7Ivls1B+(B45$Q|IuT!<R!>}wuZJ=DtZ0-odWL3csa!;yc=<#YVQe2NF|jlU`ZUL12tF_BhQ0%23OZgwIlaJ)#I;uk_$?{8x@Sfu@`xBidLQZoNJI#Q7{0!$dD}BxSgIH%D`Sp40N1;>l2_ht7P9bn?B$CGLB>5S-;!b20&7GKuz~;F}Pcpo$;`wb_DiVo(gZ(E3uc49-fN$$?`2{el5T!DN@CD1_=Yn<UlraHt!7dc*tKX^|3inL~96>-ksOFa}CBbdtv}0NwK7)mx-v!_$Km-<)%V)yh>6OWLuLarb&PkRZoB14*Vn-1YsCu*l<x%bM_^@E72tcV$r3B%8Ud_Qi{e@XMGK7Bh5myhtdDhilTDzv`23=fLATqpB=-w>6}2|YDWxK8iH788e&WiNtZ>w1vRSN)DMe0i>2=D$TF56DX=*VeN;%4Z3bw5+MnnKj_#NST>>6ei@y{e%MZcGv5~3A8tdc(xBfqYbd^-`p9+8{U%}B`QWb)?c3zVHYra&?Q0B%UiXRlv0f|lliRA6}ix&wQg<#wZil3Fb!IG-FbL2yj*dcI-F2w<C?n!LV1&!H|f`2PzT3J0>JFPlPRwe`|ZSSQ`x`-Fti3D9;=w|;^Yi7g0O|+}R0h4YTxc9Z%a*BNW>D<~EY~*tpq)rEv937SIVSiU>ZOAANUcK*Pg&Xr;7i(skvhJFD1*{nPxe3USCE{Q2phxq9xK<frYckg`X;uvNIxrdV^1`mO!+Jwh0V5(bTKc@A^)uO}aQzZRk!;l3q5$ZDV|3DoQQQ}dcHRRv8OEl8jc7nfm|G~~r6ii&Kn<FS!w7k^AzD2lN5lFSy)ElLHgFW7dlyD}{&|&gW^@DQN25KRrYF%gr%C6cZvCde#L~tjOLGFPPiT;dzTt)&MUGRmGR$yc{qm!(3M8uW9Z{Fhq67<c<#*nckFNeFML+9<75;>EO=?PoI+F{3uk8*kA}2DeJ|x^HwO+O04DjHOn*dQrk_pg+0E6g8ONMQh>x$A=9FK#K%$9J?w%kxFK~4L+ojeTk21n8|WtdQ8Iny5P8wsjt(c`Tkp6AtH)(tG9pN`w9AYn&6(SlO5F2=VthJ)2Q&QNtMZr1aUmvd657wnbL;J{t)nnyw4rHW+P3OTk6QTzxO$tSa<p%)psTSiFhkxmz?HDNQb`qKs`dOsOdEv^3*OL3Cq0#kMOfu3f9*=z_Hb1N5|ym8#QF-T2{X##I&pO#|hG2<vDYQKi?ZR7bi3g*$GtO;q4w%?6;o7-Gt%c0eqeQ05`iHbS`#f1Z6T;fS7J$+LE=p9li%PgQSTaGMD0^n>^&Mg0YF_<SuZ2utD3X&8SmiG800)9Edyi7Ateb?cHPgZ{gSvpP?5GxsOc<_{|nE434_6@>D8KDZU@xyI7^h?lOrl89-)ZT$%3xefxMXHH1$*cr0>k_qiT3BuuKP-o`GU{!5_AcVe10hT{ZJoED8Ny<RI*Kd=h;7qFf2mlE7KLC~$`y-Uq7_`+E1KFX%U(iniObqKKdU&C8~X|P(JmEOSETkx4LIi$cI=rOXlqEz9J<{&CN#dkgcWF7EXC%b2ige`?)~UHv)0MOBdJ~G5)f>}?XYn%Pfz!8`qt1_+X+>KqhW$>u<=OWkA>7Z5N}&?;Yr@L-@Xk-iB4`I=C)6Xr%JvH8FGt=i8eY1-!o(j@zK^SwP3I(hiA-%=dvP~*?XP5SAeVXWxROS#`E(}=WqY$s+V*V*0dsrww7##U4GBqeio#f1DhHq=*4k&7W?h%DIz2=E7B@VSJDlH!{)^y>V~;96_kVj^FBpwB7aFUK>@@YM$^cgP(3K0WYt=tX@>uYcYyF%0LC;@=m~P+J~8@w7%1sWhUE<?C^g_G%dKg+j~*?|KU9$x5*FA1A@yYNI-}sT?7*vDLe%#vPe%92*G5t<^`BNKeXSGh(L-iRNqG6y8utDuMTOu%9|LA<$$wka_OU>5nSW@u9CO-Ult9~?)_nI?mek9}&+_x@;ZMsovf-+kS3c;TSzRV5hSLdd6g7vzvrjk^*iB1QXs`OQRPyT7h~Q;r_jgR7)Cri<#!w(=$C}(oW^g?-9R7o-?z&XH!twx(3S5UCA@wKqAW{K>CZ7<1@-A<2sPmr!Q414^Z##x8XkU~ka-s4~D><-kjd9W&-HX3Zi10J<8nrMM_MNnB4Xlop<VBeb@uS?a_x84Oy6^G|BJZsD#T6)PYE9^}U{Gpj$A@>+#w)ZaQD2bXd$$&^8SG;?B}q$FXyxv6DZSDzk^{*9Pcew^k8EQ^a!)d~5eP7cLYd;mS)q(?7z0)-s2A+#SvfZPY$Q`4jd*fCDj{bgc1u8l6T5Z}q~Cx<sb3wocG-ttuSpEx8`NMxfPR)@`%YuA;vb0wg;9WSJgAd&0>?-6B9#T-SC2XBEfbo}ldQ2cQS-DmWb$XYq2oPqEh$Hd*YwIT@OET4&E;4{`YDKFuSLBSh`_pw>KSn=G+Yb{ok)vCilc$dR<f(0gcmUA9%Gs%Wpy(~0XXW(M-08q95xYPZUJU3de;f2eyfo{?Uv61??$JsM9$D8dHP3eCNd9ztSZ>~EBwU?=QE%s68S}9)$w%wt!m??KIdT#D}x4U+(Ku5;(=kvs@npZVLJW`jww46>Ur3JYHqspdEH}jpoYQV^4XJb{F+nXsQQE6Sg@=JJ#iWhz%owt25(#=2LS7V124jA2@@o5pdShumd%$wONbs>ZaIhmd|C^<I!G?dAH+|4y=Hou)517|bj0sR@R9!eH=9eYcoYuHRPODg=S6`6XD^$=Ac%m-8e>(Pe<!ov^*a)BzYP_@50Q#NR-kRRQ-(r;Cf)x}ZOS9AIPy`M%^VxqOLY@hG-qbtVRN2Aw1R)W;yp0mVQ;X_p(b6^(pck|A)Fr*n6K}1UE+KD2z_5t$?ZvK6j7OkA`fmC9EmQX<<x#_-JM*J_fbxh;YE4!oAFp403jN?qizvHtJpE{QU@JV^{x2+x)A)lR30WVy!!`T#J346!2Q>VT?Ji`ak>&d6UmLP_4a+CPup0EgSJ<K#`hA{mK8M$6}2OyWr8Xg==y~?duj%bQd2*o{OS3*Zs8fKUZ6qX#}~SXDrcPj=5Ub+u6e7sekZ{D`BY-N-%m<l(RHijfQU&msbn*yDMiAPxz*Zjge&Bf>j7S)@P5{p*~)sHWBPyWGQ!UKnWa-nGtb~QF|^N)R-BP0i2y$Z#%r4C1zL*ru$b*v13olOCX?P+B5A_5rpB%x4_bGn8&Tg4I+;6gOafGz=n(!J6BJs1+;E0<ohnWv-5M~SFl+4{Q(2@BH}@wBVUg(=Ip=GR1a6dJ_X|5&c4*$*Z@gPX9Gs`Hk(AZbqMs4YF(9@hi2;-0Qc`m&Vo@bmePkT8#*|7y$Z0B$ij}`w#=~qwsBS>)Z+Q>5qZp#w`N={4#nUU+u=Z=57F)&(9UKrh{cyh54$XX1KYN^N8*h>g?p)kM^Y47M+fxlNW5gGv3|NcSv97OK{o_NB@QXS~)=#Qkk%oXB^A|vu(^$y3j7YF`EjQH1Ds|XMUqHC}D?Z-;#x8eGpMreR_=olNY_Y=8>|j6l@7(pJl=)=xm1*}*yK6{+Xqb;|0Hup&#{ixQA(?x|6b59@9hx~sg!%wiEhR<ap*Rr>g~>%_eoFu!6lWBannefTaq!8vMw7g)UPVTnDG2TaVsau50h^X&wHocL<j+a%JPDu-Z%)=sog^l5={$xh3TF30+8=9$#TnWn06GBlnz9_piNT9#CvZ5b)PMn0kugj{CxvlTVFfQP(0@Uk(`sN4`W$n3ei;pYSjD-Yl@sj^m${nK<(p?B(Fh&5_qknB9APPB%%&dCPgX1YIBrLj+}?E%gB@|U-Bkj9=f$pr&1u81b!zeZ6CUN-AS`tefUrwZ1D4BtGvhAe;>0+)5(`bJAl6y#fV9Bol()sY{)(jK8KFAUK>*Z*yak@^q*8dPUPU57H1#Y%K!sqg?JC-)zV3KaWA`>7K$u_kv*8U*dKDq4*j^{9vAS0M4(C+!7jq<kVC617e1_7gNm+dS<ogUPj>64!Or5UU0~^C?GZcx4iXWMu%WUh^+Ubv+x0TosX!^@GMRU@N|7@7ho!SiJVhx<d6!Y2WB&W&j_$I>7R$U-oX~$Z+Fij7&d12^?$7U18*68VR`DyI*ppL{7EH%M2z4OYtHpJktnN<iW7LZwt|I$5Q3smNWF&qg>;u0kP13K?Ur%zkd%_%T|^a$Ev*{@sT66<FEYH#cp{Z*a47;rfnlc3Ex3blwnlN>vTlI~cg6T*eVpQV$%x+&ha5g3CC?M(H*>Ffo=p|~t^{C2wGoPU2u@HX+5Nox6?f!xZkV988U<I#S^2(uE&1^bK$A^eo$G>r*`+)1WFu-La|7RMX~>1~A>E#*99R~kKB#*#ls+;TjnD8C$eXUMM8C80OMo!AR9x2}DDVG1=bCF|}{W#en=N2EVDDT?kL>xx%XnqtIVau=7K+U*>p!5Rz_Itk6*4V7#;q{gz0ghmy!;Ymy*?*%2LfpR~n+}V6<t_bQQLdc0|I9E8$%8ism<{k<sWQE_d_ZM!M!q&(NY5R@;MrzUsk}4TYt;vvfqff_kL_!2;K19=>7Q1sMNK<IMtwOkxIg+XSy~|0~sr^AWUx!SyzEzdaN|RXZ?YxD(Dm-S>{a2eIeVceLa*f^r%o5?8^x!juVk7^uDqygDTKryJx((q0B2Q7xpRR{j*cg*>Le#k+grSNGgC3wC0XEPWXF6wCjWDE6AKYlc&icyb6~XXN5wSS`kU9h~y4Sk8=|qM&zoH!_izs6XOQrHoYT<&>Xf#wfZTk-eWkgPNdd?*)!drp~(noHP9KCCSo<dvW#^j%X)oM|P7b&B-D_}WqTz}bFWXvfsVnoDx?`llCNd<piA4){eR>C0@$W@Dji!MrlOE~w>Kmg`=&^Q$W`HgZ&OF(z|BA<6rO>S8BS4)_IE#(0(h8nv(F{T^;qQJ?XKEwdYn2v*yN`O+3mAuwe%$F~UTWk#n!M?H3^Q|Q*&IcZH@*BY<_Z@9b-)I&8Usjv0>i+bMz%k3IZS<)K*tRKadBTPgZbD>GVUgNrcExfffBL@Rb@<OS!w^-)0a+rrGFz-g_*-dfklF>tG#11lJFggtN%xl7#>qRk5epdZMbSXr6<hf0PN*K+5=vBFBY0_SlwdpJS@GA)y@^*GwlMv)`FBetZ^19&Fm_@v3a#B_bUZxo`oS%LP+8#0gnefo9IoI%R?i0)y&Fc#Z-+7DV)M{L0ZtDsw@G+p7%*>XpbiGZJ8Nv#;RS+`Hwv*x3Y{4VqqA{AJsREQMaGtxaaP@C#gl6QfADnp<HB{s6!rMJHafFi)x}e6BDlS|RQj`ZnDY0~R>Exi5YNjs|9P|FmIF#f`EFeUtCUghxVKUxnj7W)*`WC7WJUY8uJb?s+Q=x^mLh^oKk4kZC=F@^J|XODnP~JuC@|%LfwLji2Ejo2xDxC4w9u6Yyje@-wdftm6aA3?*iehG_0lG}O6E%;DJAPV>`DiduKqNJ-^FLR`stC-9O%qcMHd<g4vRC|1PIR#eoAUv-^38HxrBNLX6ZQY#jl6x3H<NS_ssX?t3;*<xiosmp0bS19;XO+KQf93#yCIC$2a8`0vc(_-eKT-Hv!S6rV%=v;vTL=887{_;5i2>`;b55Wk!jn=&|PFU9<?TG3kR7<7dERE2PK8-VMR16#)zRe4{1lV~gG|9`N~!ZmkcZ-bcijNwTaUhSMp{cUinh*V4syk`M!H_67w{hCE+`3}-gzxzZ-o+GjrMZq8PhAr_08YwWYhRQAR}wauwBf*vCs?%Gv*kl*?P<;;h#XEkwA!z8U-I>g#N!8jgP!+nj%1t9P!ug+Kf>+Jw*&~BfS$2mQ=b8l2q5VEGH4D3Lw-c)7cw0T$lS-#}>mk3lwY;@W~0r{ecE!PEIaY(&gbgw|xtDMEoMV(7^-r3~Of9udbmgZmgo+@)^@}L!g0FAHtlierF9FOYbbpO(^1a(!*H!&#EZ65;sLf(bAi5Kl}@Y5N+rYZ0M0a^1iEoSiUNg67@Xz~x*ht9+`=-{3XDi7PCfEpYjIHLOGq%SfAM}SU&3)`a^)^yM~-H{;sV`!J1mQMG5Wc+8L6#YXa5dr<PwLq&du7L`?4B?AoH1ZtcLkEIQv1bh$t{gll-t`l`A(O4gS5r1&Kd#*@(&jx9r8;%mS#iFyhWR8{&;-dXR2h!dYqUv&MXf|^6?5xX9QNA0s_B+rH?5V|^B}5O7qKu%seTgR{}a||9#%>=YFTdC$Si!Xx<}1AsT0Hq`7;<LWQ{b{w1foa?)^|R<@nrPa+OZ8064K|VrrrV;l0Rz=`9o14HVToQS109RMq83e~*IQJsxbuL&PmJgF?ly6V|H@s;oQ*%u>v4vA+21dgZ{-&Bw=ma+cth|H*}68(jpb&Q8mgVY|Jfu1z3S>Zu6m$dWmR<QOpA<>Zlm{+BDS*62?l<X(ia4Vu&3fXbxHY3&BfBpo(HBJ`2a2a#Q@G=8j!H-PvFYEi&MJP2aOG!XAP(b(ZGw6L5~j*>uBwlX8I9B18?FEi7{VkETFRhhOAFw0L6hs{Zuky&w@d&j;Ta~;J90}3r5E&&Y(9x(C&$Ngnphp$K)>e)c?$2Y0S_>5s@*%e@6Ycy^;uyuiopogo!I}DCi-z$2%eN(odm8K%JlTJpC-z{hT2u{M$z;Fq+ppgi_E>@bJz}PNik%Jnr#``=Ly+itRGh6inv2S**)wED?9gj<qlgsaY{{~_qv38iYYyoyn<yZ@6n8jdF^tiO@kQP-6Omb_e2hp~op_PBKcVWwH_SwTwI6CbQ+s|YY3tM8)$g#@g;C8P_dhg2Ga$sz*i^ne%C`EF!OyP(SFU5)h?hqc0_2S?iO1UVZrzYOg(nOIX!6a_ns2KHMllV1{SeVGAKxk;j0+Af%(;xEzh1gqzgQZCwQv~#dHb@gexinIqA|!=YdQN+N<yeTeW8r(|skh(#s;3FN{{IKUChwYfTsojT9`y8DHX`^N0<zmOPptNrk3xp8naa*jGipW><$N8@f$c3aDhI#W)G!${06_H}xgYo;bYoGeKvlCb96J5=6q;W#1s1_<{I}IPFgAd;5c7&UyCucocFpoEjJbbqS|Vm29k?9T%uC}`aca@&Pa7T1ded`__$jq2qyg9U=X7$B6@DhK&J6H8O+8h~`EG7cB8MB|JKM=1ayC}rht(X#q0YjL^vQ4Iag6&fJ<R0g#hTK=LRWUUd4NQxU&mh_FzZwGPw;y<Xt+@NGJ;AYP0KHs;#bLYMbE3LQT4~Us9p$w7;`)%YK<rduK-~AAREa9Ax3PK+%`ux<0S?ZwnZBa1@KImEZYrI53*PJLL^hyP6DdQ#y^6j-kbu=mWv!gDG*l<Pq~jNkqcIoSpWkuobQD>$Y#o?SmFMyeDt{>`HaX_!RDNqC%=o^Ou}PP8{nrKDer7BL(-{m_1!pR)@^qn&kCH2nox%~4jk&qqcBoeYOP1F6obHX`mHWIKA8yB;m8jM*|)a4ma+yKR$Q%%=uNpx6kk|JVHEV0?WqgeaQQ2P3lVcduO@RjtYH|~=t!nX+5AXZ-6t!5oiq{<R@)|@pTz&r93gp9!cW0hZoXgVa&}g9lO5{h$Jb{YC{0!ioho3`M&N(OrcJ|2&e6##13Ei~!A)f4!Peh-sq(FISiGH3Vs9$HB^;nMcEUM$ZQ?i#Su~xOsyv$ygzFYXFGDj8(E_cshW#9??_uEjI3twnwf`YzuGUZUGMWBB#?zpDWuDFFTMi5YGga=)LYB@`q>|G*Mtw5Lfe(Tzo;aZ|X<{i7q4R!kn9lRP+a;u=WVW384KXHkhDG;4i`$dyv8ufX^_>a;(>h^ZbmXXM7f=q{udeS1pAa=R;|ejV#Q<Jdv?MaH$`Mr$2uc07Kl})va(8I5=AJvdv?_y?sElhxWM`g30Jd%}CNQ>b%2dri{zUQA8ry`edYlcL4M@k_Y53fyi{z0MOmM`MbSw9HM4tU9I%(eh$(Og3GYtK-gRV9b;RvUB87x>QE%YX$p#A`=mMUus5m*J;-Y<G9CVHG#TBr7k@?%rr7RMKni&Mkjw6f<C{1@e5wEgv{&DBaRXZJam3XU93Lr6%0L&<ERX%cC9(M|VD>Rkg$A2U^U&qs(<u;Z6ozL-`&JTMjnq-bFD7)Xd1GH*Z~AUF$CCbo?fP^A~1ipp7pgo+7SkiA1}rrffE4DJC-FNE*3aaycFY9q{dZnL1~mmVH&;L=8q^kof3hA^r*Jx7szm;VEAfE&=1s!;14u_*JlHaZg`>j@yy93y?TgQW1gjEL-Y73l{;2(G7x4B$Vczf$NEG=6_iqc2Z^oyP=;=wGSuFd}*y*}8GD-(^M&`^<8}JcCCG1JG&&MzMTVtrqTa<k;b-iL3<hExi&3&<J8%_HpK3<?4peC0W*Wly>&1Qd^PhoRp$oBSXV5^#7Lf@O%}kcL{0~OZOh?m12TuuhaY}$Xt7Jg9Xg9FBgBFg_;P1$vh--dKKiL%L7F#<~3h(6$!$bogWE2AP#dHHw5{%YG5DTpTn!v6Xi^g;YOy#=6xD-@fqJn9(+#ip>3)i14N^@TM-f|xT?G^U5y*UJUJ?ddGD2pZ(SNC0T!)P+A&_IpU$m)#8MCiI}%bD*V`$}=kj~P!&ZOgaZjnd88$x=Mm!mY^$3edEqLxv&2P@*HZdBuigIgCMn}le8Y+JOo2f&zwdpsU#`d84>S-u6w&5Wy{I6wSw#}GJE$ld&8FKH;h+VXKSfM-{LUaluPT9G@BkCI8<LyLT@eHDJK)RQ&kW1GjXCW#V{>eAyC@1t)pJ?Q6-!6UX8qIZXME?JMW(u^0IycKm{4N7I<Gkig{TgWl=3VuvijH8%l;r8hYA{z#xw?}g=S<HMxmfgz2i|z6sS&c_XqY+~1Re0O>JqN-)x_)`3l2_cEIn?+!GyyPBRexQc-Eg^coxePfGa*ak=s!Inr}R)-Hyme{p{41^H3mOddryaaO{Edu8S?U4Pw)HI}98#CQpxxP_xH2vev>Gj>H+Ldl~&KO_pe2lX+ABY6!(1BoI#ms*SOlCt7aHfX+hEV>%$El$E8i-j4d!&*QP4No?g8b5Sf&zU<Kovgkxaw|M)Cm6QT%6g`jD2Eqbo`N3ta?&WpH14p@X%NUUz^Wv!p1OF?8+{@Qd?=vTl-@S>v+V&zzhVG6k4^YmXU2Z&-=Ei^c-(G;|^<76TdDx8vG|Ei+8^vhMon)G7<92~EaE74KwVg_i%7`Xu@sSG^vI1f-Fm2htuOj0D^Zugf#&#;}0wb7NM8wTf(`tUYNiaL+*Pl)Yy1N@6Pm8V1dSIF71&A*u53@=xwDE2|(eO54632oBNOJzR3BeIpy6D&06Ok>a{ho%QvQI@jpiQ}*_lTGwX0137={JKe!r`ojvQ63de6Lio0&#7lJK0yORuCheen4W9#C6mZAqd6R;y&1-_h&Ib=WX(1be3<N$54H>-;|M(=Thd?joY&Gv>>+coi%U#IImh`$i2Li*?}bci>yL!6cX#4$GCw^5W26G$2F+B8K$iy9mEnv9+?Y{vxeSoTUcuGTRopyXbDdWc~oTN%A9qgDz|~_O%TLt?a{=9?AA||iqCwYPP01DfOp+gR6N`gKMd@#kk9p>y`Pkyd_c>z49QN$`N8$PUon8_`qxs<lC(DSUVZ^qJ#(_|1TB3+5Wj1oyN6DnED@70U`MHDFUV0B4za1Nk5}g23uGCDUpe(x5x|p7H;pZ=!f0M`^|A7r{Qg1Gca>$6R@(%_Gclrq%pYEPb?jk(7oQG#ci)}IEIrv{<(L_3xTHCO780rG{9f<QtQG3F&u+jqp!+0`LPSCufCqAIdMeHT-(Np$(6d_r2yed5g^D-yfSnHNhZ{StLSnTC&QzccdVgRXz^5s~!@;9^sIJp$sjTo8dC4RKZkYiZdHpoG+aMfKLuxSh6tUlSUg9^`ZZ`Rx1_y#?kmH=!eQ@*8$^*+LhqvH+MK!C7@w0W%dNQw;-yI(Kv#r@3A6!|k#gvF%HdqKI??6mA>P~{mdMlYS$H%1~y7l3?*l;~U=uB}Kn6n-<$jff(A}Z_Mr?<Ek`mN+5ZC^-}9^}z$%cA0^ua_YNq~EO1L6`o9v2f#$_3dB!aQubvI;@G=23-6-ov6XG6S7%Xp8XnT)sc5JEc`T0f5<8;$z01@QS)+A-~<-m6y?BqW@Yl`$?O@s2PD{9#-&;4ym!=xv{}Sz(+&yVKj*68mZEukE6Ngti}W(u4%hgOG8~?t1XHlo6|X<x_o=__uI*!gN)CI18KI3#0rZIrq(9d}nOU^ILhyjKQ7az<v3J*iPJxP_VFXo}bH1SIV~D4;V%46oV0U(?&yWzTG(VA@UvSp+zNaG3O-sp=&GmK$nD+gM**~9WC<3m-!FIeykzJ2Gv+LNpskx(ZauZ?kL(&i=BXonYLh$PY(#TbdIx&qFAL~$^qgrK86@nIs2I^){tUu02<)FEAlt-!O#4Y$;C+iMv3j-2K9#*j~^2hbavwO7}By|MLvNav#wB@s?6yrL1xq@2_?V6iJe~|E@ZJl`%y)DV@N-i5LYLC-sJ1NS2R?rF`=J0vP-4!+oV(Maj^{fP*Ly<-9IYzz{Xf@=z!|GdqTZ$ITTB&$ZlZlerrqmzV9&{NUHmpjiS~S3Q=h_;m;nJ<~s#XX{thM|8k_7Gtp<}c*5L`n9{wLA+qH;<nE(mfFV7*Z%$38<?r2Oe~(qF0*pcbIY@3P7=9~n~yxnr7NIH-SDJ3R>70Nib-bFzmF*3UnDRVBctPpW@(e%4^Qm_J`=uo1fde)$|GvUHjQX|Mo@&y1>>ew*=PGW{}ynp+7zFI_P`SoW17O1Z$y9Uf{9GZ$3sb8&^aOz`XFO|r~PK^_;fJaH{zQW~8?mb!$IOZ>qC6Wn9o?!3k&Oy+p*A$F_r5O5(^hGbpr0A?6g(f-Ba-#idbEAdGc>we^nKeQOywn=x*_{;i)Wku@|7nl>pvIAbNm3i1%f;d4iBtb)G@q%DT3wt-3TFaHY4obI`I53oO$T-%(xT#o2J7JjRDU|CEQ+1)}HOjx+QZf`_4%p8H8jD+h?Z34R3jmmW-!2z$>>8S$eAYaC5=MJP^j*>Z;sbHyOG0%L+8z33%-WxbS&7P|v#90}d3HuSf5<{E_{Mn7bEaZsev~}efE*0`TP1X#toJ4u3REOU;rqiNl>KR#dADKsqIbUgn2rJ_CoozFTtNjEWKd4hlww>9yhtuRa!v_`ia$lNr+v;UB6PJMAiIs2LmSibjDFU`4f*R4T$F);qvO#7J4%7e$ACsxIZQ$xkvOv(%_imt?{uu$TCty{sJ6^Mi~ywU%&BU5cE(_jt1^xq@9$I9xe_&->XA&@+s8?Q;(mICpigER9HNjQK_V&Banxw@Py%~8eT9x|y#g20CUlOYOf-xL-Mrr3&j<*&KEBvIz6u{~H(qh>0R1f=t1oQ7=wY|;4U8#sWAE68uQL$YX3*$6pGHIz5CdYZCoa1!mWINAQVke4fglIu3p(+w0`qpMG|*JvQeqZlC#7;z9~g^ce|UK8dkIA|Pv=G0>ZOTMQSkgW%*fkq##lk{^F*3z(O8c{Ts+N{9Dmaa=hw}rAG-nm4xj%!C2e%pD@9?0m-n4ug?Ja5+EB6ShhF{uor{{wuUn%&FBOkr4I46ZstA<rnfi&r0O<^+TMej|)R;nD0k{Qnk}o&h|MGfpDdq-y{^K;nlBvrn@&}QKIu!N)Kt$!U@q;Xpm^wd1Eg7If0;GA{YOnMwNu*lUQ%fUjL=0|bERqOS!kdvOa;(6GT;b!V7MK~-$^_v;{IO|J(VIJXZ%RpvW2UV`fOq|sz+K~28#Na2IWHV<ws|756_f}DH&UU8w|f<NYid^0Zn33pz;)SxWHQRac}x=tQWGtPD;9F>0AitopWGTEXaRC+E-|OW*Nl!EZDX3}r06^h?FBx<IhdBMLBcLni*Tem-D*Uoy+Lw`Db?2y{CQx&<=|11G}^XaT8s^27;O(i=IY|Ue(qd{J_6NyId!ZA2r@3DU*LoK!0TJaMEG4_J*6a#aM!yD5e(peQVv1?{+q0Opwto_*QFP7a!&{0KQD|2uQ26Gh!faG0HPwLMSgavTqBcXYJ;Xm0beZ<FNq!3*1W!rB;kVc&4{4EcjHwHIvn>02j87dt4)T8V|RLuhZ_t7R#PHU2OVuR77}w?QES>bp0Nk+cepWW2wb&$Xq>0-SMw!@@=H<$@Am=>{mfsm8?OrLOD(CK&1Uw~U4-jFVlY9Fv{bzxz+wB;B8-S6v`tiY21iVlLBb#=&X`+x0arm2gEK>z?-jmy|6?enxKM`0!#RClfatP?;1WWa(KHI3bMzo^c#h-h=Ncy-o;lS>jTVjTG#k$;463-Dzh>D7MKxbINRwj#<lfP`*qFPvgx|qs{-Qdqg)*g>et8(xaY^-{WIq&F7@a{Yk18u|lZCiI4?O+QvO47i;+od&xH+(uRZ0cc=bW`y&8jjjB^5#`N1n+uN|<(Q=}K=jxsY&aUDWk#-LF05{e>Xa=|}^`Q>BTCuD(7iQElDBvPp9S{!W^p&U3&2R~Z=n8ISu%;ByeCF#FA5#KcVaev)jy>Ez0A>qe5_wK7PhKSgeepT_b~b;|G6{CjfZo&pJ)bGTc$5T+<!CMy)pkT3$$J?|w@VW|^!L5BJoBgv+J#4&mAslLpiD>x?oUsk%FiWUN(7)#=u1W%E{0xEG9fHw6zH0#5xD~vf^J+iPS@+(jE$QhNk2>V8WLHaBJT4k9N(wld!6wN3KQ~T8TtFJf^x_22WMKBCDZ5Eu=DWi2O!B>A1JzW@lfHLS-zhab{cV|nllGNRDOYz~Gi3_6Bpz$)}lEv3VgU%%FKax1<y?{JMXbT-xn`j=>yVogb3(QmW59ql5{`AJL;!wqYUxH4BP-+<KXin)Ur{Ts$^&T4q=^qS^7?zvMG1X1lfF2TFED170%JkD?R3@4tTbT0Lu;ElO|K;ZmUu(jNXE**)tfA0&u?|}KUUy@5v)AgF60&fRHXhxImB2q2<3_NaL5n0Pz+#J?>g%rZ;Jp_>r&Se?{G!4ps<qZ4_wW@5CwBB~)+0P=vn*ma2(qkE5q$H&#Ls`;$+;SnBHzDwIog6s99@fQ^_gxOgE`Q?N){8!7C9N3i^{E|M~P8<?^f*_4|zkxOQ43W+#?1=j2#c|3&ChI^nM)o=%?nA{D_18AN)-i)yyCWyvN4%U{4t0SjA-e$+GI4i>t$|By@mxXrwU$d*%WE__I!~sQ8NKpKL#h%JahXDh>7&pse?f7w+wiM8{sFTui?*v?%3DtG(~;)jrejOqi;63GIaqnkX4`IAyma$oETQnDKiyUBSxt=azJ_`!R$xeyPj(+4wV+C40lh%BzfgB@`F-pU7ip*3s?(Ko$%y)TSV9ctSTV?kw4*ukN7yW(ft*-;XC({KR+8?aZ`6m32s4V><m1%oJh^s0&Vzn58flOJUiiGle{ME%C^)sm;GJkUbaf)*CZt)ytXL6l^NZ#MmMb|Elz%LfDIW_$oL}u30xtW;~`=!_60ANhwK0ASlcPnmux*BV)_h1WZfGSBL-UxT;Ics!6r~EU#)cnmE!`nWrUMZlp^DpbqZr9MmpD2uWPh%(efbai|J+Z?EzbTX@F?qs`?gQM=AFyuBUBdGWxVPqzAZdGTFRaPef<E3s(X3lv!o;WOnU-7mKGTWy5z-cBA|2!3xrbqJ#NlqhWj>HSo#g55iYVZDsmyg)Li>?5^&a=Hy`as;D9D1OtO;ukj%7AR7M(iq>`UH#(TGfrjpR3svMSkPR!Z|Sm%n!)B-ls)0_ka(Ksm0=jyxOy1%^H^m`MTD(en0KQrc#=+!xk*~Tz;P!OkzA>(u3qGdoAV7yECM_!7??3NJRa*N0^Y~@GDX}C6|y)Q6_RLIgDsFaq6DpEQ7l#9VNZfOW2l8>{3G)m{P#Za8kov^-Bq_!_9OLrTN8;(*8PE+!xc?adkWjaEv=z=aq2}^qlh-dwX9_693Un8;EV0(%QiCGnS)=nYD!TiRQ^71SY+C8Z%p*P3tzR=J#eLoNT*qd*IqV7Z4!1i;2y$=kGy%LZAkWNYz1mrag|cvh2iIZ3y2BN!|14rZzAv#i3bB1X&b4C^L%9qEOKxM`c%hi^bW+a+i2QU5{BCGyg^<NFnX2sShXzRU~WuJ`h-)Mz73$e)6{Z0|1B%2UrMtnR5Ui&{qT&$Yo151F@a$a^b5kvK1Lj#`Bs4pgLewc{4BLw&!kCqY=#&RxQ314#S`ivLG$r++VWje5&=(9F!v#>)C(tY)$;2q&xSRt<3jL{z^8V8E?y$W^7o=ud5P47sTf;5eooeK*$8b5*m|l?QTHixCpKGgJV20`0j1!5vRZI?$)1HG)R_j)!=3csufE9CPu*pnoE-spDJ=Xh_)lsEIQPGOrq;CwlJ(<Yz2;kQ;>cbzj-T6IR@r%PRJqqVB8V&PvMM!RTzV$^>!ar$6*kvic`_n-hk+R=ShIk$;YC%4(cCs)f!~=Ry+G5fpyf?Af<S{Lvj>>e@phiPj9vy{W)~#Y9#?!uGZQn;*byNn(&}J$I6=9I=baSJS3oRzgB552PhavBSQOBQi<LsGRY7@90D=p=N<j~{F$p)+F6yr7+^CwweZ1>euP167%AW8!cp?Dx>NlFy3cF^pwo7;YKySnfx}jh?ZHiC-_N0gB08VIMwFU6zWp_anpH16C6;E;>f}4jAI~avZymKQJCG|w{6lm%?^V1!##WO#S@A~{ufAf71Oa@#7T?x0Ag9l414w7%z8FWt2yllC6ZF9Vx%@j!b@3~L7jNiFPl#N_i*&3EJVTupZ6fc1Ld&cX$M|#QGrzed*_NB(}Cd1NTHJUXu2lJA1hrDb_Q4BzaLk_#yXJ+jS*k4NSXK+n0kg1nm*@ZQ{?oAt|VzRsKnIk<&U}0iWa$nyhRxB_1R_O7Ullje&<x!_pm!pzeSeKW_2hb;u9hd4BSwm-VvO`dA4Jf}M8^M@yzCV;j7{W06Bb4M-+HI(1{?-p8gdX(&fm#8=jpo4o!{_l7MWG?%wblMtt8%$^fCv=tfsT61a~+fdGom628Sji|2)YewFm=9+AnXwCaor^xhf44DMyo}``DNRmU9U>$;^Flmjv8uY+a;dtRMPA|GE)~yvl2(Jwg8AySd>arQ^N3_kUKo#RgtbYU2&idLsXHn6C%`i)w3g*0bP(+c*6_N(Qn4nxY<SZnON@r2{4t;CO9(mdg2>MOxSJ9oAbP%UKHH@4a~p_J%_nR&rkklr5RBvBSs?j7mQw`$&I>nLx!0Jhn#7b#&DVB^}`H)J<ZC+%Tq)*qdW?o0Z~fffTE>#@w9>53jaBoDYe~Kvjr_#90*H<&5_ikAH3VSap*AXAo-wnYgp(zf%VxV9zJNCYE}5M*$IhxL*Kq2=Ippaz0(2?<v0DOQGMyMlO_Kg_9LPxGv5;pNS`A6a(nICa$66j2il&L@!#YBGL63sWy>O1Ixtu6!fbRCEx*29E`LT1;_uq6ybO#GB<Md<3lJDTUV0$9h4?q;SjeK$=2g8pN1isfLK8#<+dfl-Y%o9cH$7atheYt^dF$5(C+{y^^+}jaSZ2|tbYP%F>E+=D^eC|`!CkGU8^GfeF;jF4JVY@6g!KE|EtT9QG>R@7=*v@^P>;CZqr(D~LmJEl>@8O>bnNC5+1%FeyiJ|6TDJEYynVC*DZDq;bh0NB?YJqV>NtjNTz~jl{%sW!sMl<LSC*s43ktQzIwL&^03)J_#0Ue#T@~Fy&&TZ`=h%YdvadkBK!_Rx#n*;C&am>MLmK&z;*TyTc5j$fim7)ep{;4qq3VnYpXD3VaNs!+)`VfkX$jcFfcDx_-dH0NLf`XN{piyYkXf5>hYD?kbYQ;*Cw|%O5A1PXCIo#ho!8%S%y}*xv8^GZep?n#d#1C2E_M4?70ie5_nqJYLYG;q>Xj*}EuT3V)coZoR7iWu2`{{AD;*iPkSOO%n}4agHl%cp+H(l#LYMTORDyMov#?E3ef_quodu`UQ0(TmFjy@`FJW23xE%NJj|KNw&lBEhY_u!j0Ox1<1U%9KnQVDB-8W={l$@UGLw9zObM^6?PNDLmZk7GDa3f!6NY0@ix!oP3Mwk?HQRPFh(PcXy;q;ur{2Jx^Wh4hYi|-J=tcH|}&w3xubM6Zrxnqz{5;)6YUj>a+w}1t9w3k3WEUd}NH5d@G^$MuTov;?ub{4fmbOTg<gLf*}c>4Lp&C>}CDQMLhDo?ZS$X=fR0vp-Uo2dWO9FYlVz#8fs?yh8+Hwdl<a%9Dxs@>s#D7ramK_Ot)t9+AMiMQS)5@rJ)TriWQtsZ+ZQzU)<H2R+e1N!<!*~7rw69(uApY7Ubl4F-}I4yD<%X4Jaf2+O7=4Qo%dl0<8Wg&|$cKr4dSci5)1@%B~jPWwSa?iyhDAmfXX?=u^?^)D>s`zb)zVYOa5Fl~oiT(nNc}2eIbFd}UJn0pJHM=+00}o{7Y-_-zo|)jMgH7G@dJLH+Ctze61X1_2Nhl}Mz*`;AX$_4Q%mpf>0@Frd;E3o`JZD}ElW=f%`tU62*2Pg5*F63A?Jx{DCPMdrK>Sa{8!uH@AFWa-(|qaQE9IlkQETl05xK-_DQcEu1NN5hB}ZV$1{lOiYc-YRI`O_oXwQe2%{R@K2wAH!zKY9H<N<nUE!=xqY7`<9G?V=TcLnVp&-yvKj{$_Sn|PUNqy>jV78KH)dmhOkxn?L5R>`(S#>`#%Bz_WTu`Hq$Npf<I?H(=egDQB7*j#a#4jBzHY;RcC9Bdk--I{t0g;U&@u*Wju758RDfI5G8z#rTh2^rbuttGDDWo9xJ7h**wXB)PKb$cF6=?jO>wFb25&S#8({u#^NXq7)q2%DQrUe1lnA}TDq1{9wPI*P2ap~U$a&<dW4**GcP;IrJdr|P$AN6KH~>StgzLGwW<n99ylss3YB(J&plE;mlSZHc552)LO4;V_MVI4S*ZNct)mY7k#o2OAk!llu~%a%EG+YdRr0v8ocWE#?&P7P0t@z=BdkQaL@#Bi14hu2QEr+AIfwtvRXPf=HGNoOZHvFI({PcF^)TCKs-?f<-^9?qGvJOVOT&3jq_J?r)@4x`E1f1$xJ$HFmw=;J*A17)RLKL8->vg|w}?*@=ORyk69tY6U}fF5k_F>Af0De6%PPhn^3wARcqMky3bX#TpsrD(ML)PmgQtBI+#5c2%IpY37BfWYIe#wzPxU$xJ2mHj`e{IP72PyDfmF<|i<*V%t)8H%*oX!`pk|FasO@4gCPVXYG)Lwb53+*#7++My<qA_m%LxP~hrLHbET$ydA9Vi4Y2WYYy_%Jnol%a(QF0bng-JK})FZ;+-UdNfT}nL`obkeN9@)nyLaC{tOSkd|fl$YEyz!ee(C&UoGdJ4Bx}`#&Y(L&mK{g@-SnX&*kmo3P5ZZFQoqlBN6P|BnmxN(<+zz3ZdUXh_&#fp+rh!6wQmr4F{G=Z!<_KyX+|^+#Tkpy~Ca82Ic<|f-2+Ap&d=u9N4>{f!;P~DK!O^i&LJk3J~%OwH~O_kf${HE7(IDj+}=t?#*gcy^5^&WDG^flAXoc!f)d6oZg##&Za-MfWI)!yIW$5_hUZBRFq|;bq{T^KfG`Oi;lB&zX2sM-U`(0Nf>5g91u!hS2$N;kHqVk1v!}H#!&~q-z363k)p7RK%!KX*ud2zmMbofELgjoyI&QeeJFt04+&yEC5+u0XZw;wQ4oK<;;4r=iuL|{u`u<}Mj$HcYc`EgLcp6M4(TfsA&AOz{CpBRQ63FxmSVdl&KhsD&qs);=f;GkyNbh^YbJ%<+kES^R{UvjNbS7huAxhgd_F<qCcZbH?qQt+t;~tXXByHU4N1we{F_*#?_}cN2ha2U6DGr=ZmZG_QQjcZnVnDQMi(Cdd*UmULyDZv00Fp}0lJn26vf`ovBYQl0ssI200dcD")))
\ No newline at end of file
diff --git a/records/track_10min_16mb/2026-04-05_CautiousMuon_SP4096_DepthRecur/train_seed314.log b/records/track_10min_16mb/2026-04-05_CautiousMuon_SP4096_DepthRecur/train_seed314.log
new file mode 100644
index 0000000000..85f506be54
--- /dev/null
+++ b/records/track_10min_16mb/2026-04-05_CautiousMuon_SP4096_DepthRecur/train_seed314.log
@@ -0,0 +1,134 @@
+W0405 12:33:07.793000 58424 torch/distributed/run.py:803] 
+W0405 12:33:07.793000 58424 torch/distributed/run.py:803] *****************************************
+W0405 12:33:07.793000 58424 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. 
+W0405 12:33:07.793000 58424 torch/distributed/run.py:803] *****************************************
+Hyperparameters:
+  adam_eps: 1e-08
+  adam_wd: 0.02
+  beta1: 0.9
+  beta2: 0.95
+  cautious_muon: True
+  compressor: brotli
+  data_dir: ./data/
+  datasets_dir: ./data/datasets/fineweb10B_sp4096
+  distributed: True
+  ema_decay: 0.997
+  embed_lr: 0.6
+  embed_wd: 0.09
+  embedding_dim: 512
+  eval_seq_len: 2048
+  eval_stride: 64
+  gptq_calibration_batches: 64
+  gptq_enabled: True
+  gptq_reserve_seconds: 10.0
+  grad_accum_steps: 1
+  grad_clip_norm: 0.3
+  head_lr: 0.008
+  is_main_process: True
+  iterations: 20000
+  ln_scale: True
+  local_rank: 0
+  logfile: logs/523d357e-1519-45f9-bc20-69bbfb520c1b.txt
+  logit_softcap: 30.0
+  matrix_lr: 0.02
+  max_wallclock_seconds: 600.0
+  min_lr: 0.0
+  mlp_mult: 4.0
+  model_dim: 512
+  model_path: final_model.pt
+  muon_backend_steps: 5
+  muon_beta2: 0.95
+  muon_momentum: 0.99
+  muon_momentum_warmup_start: 0.92
+  muon_momentum_warmup_steps: 1500
+  muon_wd: 0.09
+  num_heads: 8
+  num_kv_heads: 4
+  num_layers: 11
+  parallel_start_layer: 7
+  qk_gain_init: 5.0
+  quantized_model_path: final_model.int6.ptz
+  rank: 0
+  recur_layers: 4,5
+  recur_start_step: 3000
+  rope_base: 10000.0
+  rope_dims: 16
+  rope_train_seq_len: 2048
+  run_id: 523d357e-1519-45f9-bc20-69bbfb520c1b
+  scalar_lr: 0.02
+  seed: 314
+  skip_gates_enabled: True
+  sliding_window_enabled: True
+  tie_embeddings: True
+  tied_embed_init_std: 0.005
+  tied_embed_lr: 0.03
+  tokenizer_path: ./data/tokenizers/fineweb_4096_bpe.model
+  train_batch_tokens: 786432
+  train_files: ./data/datasets/fineweb10B_sp4096/fineweb_train_*.bin
+  train_log_every: 500
+  train_seq_len: 2048
+  ttt_batch_seqs: 32
+  ttt_chunk_tokens: 32768
+  ttt_enabled: False
+  ttt_epochs: 3
+  ttt_freeze_blocks: 0
+  ttt_grad_clip: 1.0
+  ttt_lr: 0.002
+  ttt_momentum: 0.9
+  val_batch_tokens: 524288
+  val_files: ./data/datasets/fineweb10B_sp4096/fineweb_val_*.bin
+  val_loss_every: 4000
+  ve_dim: 128
+  ve_enabled: True
+  ve_layers: 9,10
+  vocab_size: 4096
+  warmdown_frac: 0.667
+  warmup_steps: 20
+  world_size: 8
+  xsa_last_n: 11
+train_shards: 80
+val_tokens: 45508608
+model_params:34401372
+gptq:reserving 10s, effective=590000ms
+warmup_step: 1/20
+warmup_step: 2/20
+warmup_step: 3/20
+warmup_step: 4/20
+warmup_step: 5/20
+warmup_step: 6/20
+warmup_step: 10/20
+warmup_step: 20/20
+0/20000 val_loss: 8.3172 val_bpb: 3.6146
+1/20000 train_loss: 8.3192 train_time: 0.0m tok/s: 8417806
+2/20000 train_loss: 11.4353 train_time: 0.0m tok/s: 8294705
+3/20000 train_loss: 8.8870 train_time: 0.0m tok/s: 8200165
+4/20000 train_loss: 8.0296 train_time: 0.0m tok/s: 8159612
+5/20000 train_loss: 8.6996 train_time: 0.0m tok/s: 8141617
+500/20000 train_loss: 3.0633 train_time: 0.8m tok/s: 7939661
+1000/20000 train_loss: 2.9869 train_time: 1.7m tok/s: 7911553
+1500/20000 train_loss: 2.9710 train_time: 2.5m tok/s: 7908791
+2000/20000 train_loss: 2.6991 train_time: 3.3m tok/s: 7909908
+2500/20000 train_loss: 2.7380 train_time: 4.1m tok/s: 7911617
+3000/20000 train_loss: 2.7805 train_time: 5.0m tok/s: 7913690
+recurrence:activated at step 3000, virtual_layers=[0, 1, 2, 3, 4, 5, 4, 5, 6, 7, 8, 9, 10]
+3500/20000 train_loss: 2.6930 train_time: 6.1m tok/s: 7516407
+4000/20000 train_loss: 2.6237 train_time: 7.1m tok/s: 7424228
+4000/20000 val_loss: 2.6465 val_bpb: 1.1501
+4500/20000 train_loss: 2.5745 train_time: 8.0m tok/s: 7355829
+5000/20000 train_loss: 2.5200 train_time: 9.0m tok/s: 7301867
+5449/20000 val_loss: 2.5373 val_bpb: 1.1027
+stopping_early: wallclock_cap train_time: 590030ms step: 5449/20000
+peak memory allocated: 30120 MiB reserved: 30154 MiB
+ema:applying EMA weights
+pre-quantization post-ema val_loss:2.53590710 val_bpb:1.10207601 eval_time:2003ms
+Serialized model: 132406149 bytes
+Code size: 24659 bytes
+GPTQ:collecting Hessians from calibration data...
+GPTQ:collected 66 Hessians in 9.7s
+GPTQ quantization: 66 layers with full GPTQ, 0 fallback to clip-search
+selective_prune: unpruned=16.59MB target=16.0MB
+selective_prune: pruning 4714416/8989294 lowest-error ±1 values (excess=589302B)
+Serialized model int6+brotli: 15148811 bytes
+Total submission size int6+brotli: 15173470 bytes
+final_int6_roundtrip val_loss:2.71751254 val_bpb:1.18099965 eval_time:8188ms
+final_int6_sliding_window val_loss:2.67174312 val_bpb:1.16110878 eval_time:76629ms
diff --git a/records/track_10min_16mb/2026-04-05_CautiousMuon_SP4096_DepthRecur/train_seed42.log b/records/track_10min_16mb/2026-04-05_CautiousMuon_SP4096_DepthRecur/train_seed42.log
new file mode 100644
index 0000000000..cd341e4d05
--- /dev/null
+++ b/records/track_10min_16mb/2026-04-05_CautiousMuon_SP4096_DepthRecur/train_seed42.log
@@ -0,0 +1,134 @@
+W0405 12:16:26.757000 47726 torch/distributed/run.py:803] 
+W0405 12:16:26.757000 47726 torch/distributed/run.py:803] *****************************************
+W0405 12:16:26.757000 47726 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. 
+W0405 12:16:26.757000 47726 torch/distributed/run.py:803] *****************************************
+Hyperparameters:
+  adam_eps: 1e-08
+  adam_wd: 0.02
+  beta1: 0.9
+  beta2: 0.95
+  cautious_muon: True
+  compressor: brotli
+  data_dir: ./data/
+  datasets_dir: ./data/datasets/fineweb10B_sp4096
+  distributed: True
+  ema_decay: 0.997
+  embed_lr: 0.6
+  embed_wd: 0.09
+  embedding_dim: 512
+  eval_seq_len: 2048
+  eval_stride: 64
+  gptq_calibration_batches: 64
+  gptq_enabled: True
+  gptq_reserve_seconds: 10.0
+  grad_accum_steps: 1
+  grad_clip_norm: 0.3
+  head_lr: 0.008
+  is_main_process: True
+  iterations: 20000
+  ln_scale: True
+  local_rank: 0
+  logfile: logs/6e35fe72-b1cf-49e5-95dc-2b7d967c8075.txt
+  logit_softcap: 30.0
+  matrix_lr: 0.02
+  max_wallclock_seconds: 600.0
+  min_lr: 0.0
+  mlp_mult: 4.0
+  model_dim: 512
+  model_path: final_model.pt
+  muon_backend_steps: 5
+  muon_beta2: 0.95
+  muon_momentum: 0.99
+  muon_momentum_warmup_start: 0.92
+  muon_momentum_warmup_steps: 1500
+  muon_wd: 0.09
+  num_heads: 8
+  num_kv_heads: 4
+  num_layers: 11
+  parallel_start_layer: 7
+  qk_gain_init: 5.0
+  quantized_model_path: final_model.int6.ptz
+  rank: 0
+  recur_layers: 4,5
+  recur_start_step: 3000
+  rope_base: 10000.0
+  rope_dims: 16
+  rope_train_seq_len: 2048
+  run_id: 6e35fe72-b1cf-49e5-95dc-2b7d967c8075
+  scalar_lr: 0.02
+  seed: 42
+  skip_gates_enabled: True
+  sliding_window_enabled: True
+  tie_embeddings: True
+  tied_embed_init_std: 0.005
+  tied_embed_lr: 0.03
+  tokenizer_path: ./data/tokenizers/fineweb_4096_bpe.model
+  train_batch_tokens: 786432
+  train_files: ./data/datasets/fineweb10B_sp4096/fineweb_train_*.bin
+  train_log_every: 500
+  train_seq_len: 2048
+  ttt_batch_seqs: 32
+  ttt_chunk_tokens: 32768
+  ttt_enabled: False
+  ttt_epochs: 3
+  ttt_freeze_blocks: 0
+  ttt_grad_clip: 1.0
+  ttt_lr: 0.002
+  ttt_momentum: 0.9
+  val_batch_tokens: 524288
+  val_files: ./data/datasets/fineweb10B_sp4096/fineweb_val_*.bin
+  val_loss_every: 4000
+  ve_dim: 128
+  ve_enabled: True
+  ve_layers: 9,10
+  vocab_size: 4096
+  warmdown_frac: 0.667
+  warmup_steps: 20
+  world_size: 8
+  xsa_last_n: 11
+train_shards: 80
+val_tokens: 45508608
+model_params:34401372
+gptq:reserving 10s, effective=590000ms
+warmup_step: 1/20
+warmup_step: 2/20
+warmup_step: 3/20
+warmup_step: 4/20
+warmup_step: 5/20
+warmup_step: 6/20
+warmup_step: 10/20
+warmup_step: 20/20
+0/20000 val_loss: 8.3187 val_bpb: 3.6152
+1/20000 train_loss: 8.3201 train_time: 0.0m tok/s: 8409702
+2/20000 train_loss: 11.3749 train_time: 0.0m tok/s: 8301223
+3/20000 train_loss: 8.9270 train_time: 0.0m tok/s: 8232435
+4/20000 train_loss: 8.0370 train_time: 0.0m tok/s: 8197540
+5/20000 train_loss: 8.6257 train_time: 0.0m tok/s: 8158832
+500/20000 train_loss: 3.0617 train_time: 0.8m tok/s: 7952184
+1000/20000 train_loss: 2.9819 train_time: 1.7m tok/s: 7929746
+1500/20000 train_loss: 2.9701 train_time: 2.5m tok/s: 7920772
+2000/20000 train_loss: 2.6938 train_time: 3.3m tok/s: 7916245
+2500/20000 train_loss: 2.7396 train_time: 4.1m tok/s: 7918140
+3000/20000 train_loss: 2.7789 train_time: 5.0m tok/s: 7919917
+recurrence:activated at step 3000, virtual_layers=[0, 1, 2, 3, 4, 5, 4, 5, 6, 7, 8, 9, 10]
+3500/20000 train_loss: 2.6938 train_time: 6.1m tok/s: 7516573
+4000/20000 train_loss: 2.6257 train_time: 7.1m tok/s: 7425494
+4000/20000 val_loss: 2.6472 val_bpb: 1.1505
+4500/20000 train_loss: 2.5745 train_time: 8.0m tok/s: 7356557
+5000/20000 train_loss: 2.5209 train_time: 9.0m tok/s: 7302902
+5450/20000 val_loss: 2.5381 val_bpb: 1.1030
+stopping_early: wallclock_cap train_time: 590058ms step: 5450/20000
+peak memory allocated: 30120 MiB reserved: 30154 MiB
+ema:applying EMA weights
+pre-quantization post-ema val_loss:2.53669680 val_bpb:1.10241920 eval_time:2002ms
+Serialized model: 132406149 bytes
+Code size: 24659 bytes
+GPTQ:collecting Hessians from calibration data...
+GPTQ:collected 66 Hessians in 9.8s
+GPTQ quantization: 66 layers with full GPTQ, 0 fallback to clip-search
+selective_prune: unpruned=16.59MB target=16.0MB
+selective_prune: pruning 4683904/8992183 lowest-error ±1 values (excess=585488B)
+Serialized model int6+brotli: 15154845 bytes
+Total submission size int6+brotli: 15179504 bytes
+final_int6_roundtrip val_loss:2.70991443 val_bpb:1.17769759 eval_time:23408ms
+final_int6_sliding_window val_loss:2.66190192 val_bpb:1.15683191 eval_time:99175ms
diff --git a/records/track_10min_16mb/2026-04-05_CautiousMuon_SP4096_DepthRecur/train_seed999.log b/records/track_10min_16mb/2026-04-05_CautiousMuon_SP4096_DepthRecur/train_seed999.log
new file mode 100644
index 0000000000..8e5da57e65
--- /dev/null
+++ b/records/track_10min_16mb/2026-04-05_CautiousMuon_SP4096_DepthRecur/train_seed999.log
@@ -0,0 +1,134 @@
+W0405 12:49:04.118000 59362 torch/distributed/run.py:803] 
+W0405 12:49:04.118000 59362 torch/distributed/run.py:803] *****************************************
+W0405 12:49:04.118000 59362 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. 
+W0405 12:49:04.118000 59362 torch/distributed/run.py:803] *****************************************
+Hyperparameters:
+  adam_eps: 1e-08
+  adam_wd: 0.02
+  beta1: 0.9
+  beta2: 0.95
+  cautious_muon: True
+  compressor: brotli
+  data_dir: ./data/
+  datasets_dir: ./data/datasets/fineweb10B_sp4096
+  distributed: True
+  ema_decay: 0.997
+  embed_lr: 0.6
+  embed_wd: 0.09
+  embedding_dim: 512
+  eval_seq_len: 2048
+  eval_stride: 64
+  gptq_calibration_batches: 64
+  gptq_enabled: True
+  gptq_reserve_seconds: 10.0
+  grad_accum_steps: 1
+  grad_clip_norm: 0.3
+  head_lr: 0.008
+  is_main_process: True
+  iterations: 20000
+  ln_scale: True
+  local_rank: 0
+  logfile: logs/5e8744be-d0c8-49a6-a74f-44fa36073662.txt
+  logit_softcap: 30.0
+  matrix_lr: 0.02
+  max_wallclock_seconds: 600.0
+  min_lr: 0.0
+  mlp_mult: 4.0
+  model_dim: 512
+  model_path: final_model.pt
+  muon_backend_steps: 5
+  muon_beta2: 0.95
+  muon_momentum: 0.99
+  muon_momentum_warmup_start: 0.92
+  muon_momentum_warmup_steps: 1500
+  muon_wd: 0.09
+  num_heads: 8
+  num_kv_heads: 4
+  num_layers: 11
+  parallel_start_layer: 7
+  qk_gain_init: 5.0
+  quantized_model_path: final_model.int6.ptz
+  rank: 0
+  recur_layers: 4,5
+  recur_start_step: 3000
+  rope_base: 10000.0
+  rope_dims: 16
+  rope_train_seq_len: 2048
+  run_id: 5e8744be-d0c8-49a6-a74f-44fa36073662
+  scalar_lr: 0.02
+  seed: 999
+  skip_gates_enabled: True
+  sliding_window_enabled: True
+  tie_embeddings: True
+  tied_embed_init_std: 0.005
+  tied_embed_lr: 0.03
+  tokenizer_path: ./data/tokenizers/fineweb_4096_bpe.model
+  train_batch_tokens: 786432
+  train_files: ./data/datasets/fineweb10B_sp4096/fineweb_train_*.bin
+  train_log_every: 500
+  train_seq_len: 2048
+  ttt_batch_seqs: 32
+  ttt_chunk_tokens: 32768
+  ttt_enabled: False
+  ttt_epochs: 3
+  ttt_freeze_blocks: 0
+  ttt_grad_clip: 1.0
+  ttt_lr: 0.002
+  ttt_momentum: 0.9
+  val_batch_tokens: 524288
+  val_files: ./data/datasets/fineweb10B_sp4096/fineweb_val_*.bin
+  val_loss_every: 4000
+  ve_dim: 128
+  ve_enabled: True
+  ve_layers: 9,10
+  vocab_size: 4096
+  warmdown_frac: 0.667
+  warmup_steps: 20
+  world_size: 8
+  xsa_last_n: 11
+train_shards: 80
+val_tokens: 45508608
+model_params:34401372
+gptq:reserving 10s, effective=590000ms
+warmup_step: 1/20
+warmup_step: 2/20
+warmup_step: 3/20
+warmup_step: 4/20
+warmup_step: 5/20
+warmup_step: 6/20
+warmup_step: 10/20
+warmup_step: 20/20
+0/20000 val_loss: 8.3152 val_bpb: 3.6137
+1/20000 train_loss: 8.3175 train_time: 0.0m tok/s: 8441360
+2/20000 train_loss: 11.3540 train_time: 0.0m tok/s: 8345504
+3/20000 train_loss: 8.8779 train_time: 0.0m tok/s: 8232317
+4/20000 train_loss: 8.0539 train_time: 0.0m tok/s: 8189797
+5/20000 train_loss: 8.6890 train_time: 0.0m tok/s: 8168727
+500/20000 train_loss: 3.0591 train_time: 0.8m tok/s: 7935674
+1000/20000 train_loss: 2.9816 train_time: 1.7m tok/s: 7923918
+1500/20000 train_loss: 2.9716 train_time: 2.5m tok/s: 7921098
+2000/20000 train_loss: 2.7022 train_time: 3.3m tok/s: 7921095
+2500/20000 train_loss: 2.7373 train_time: 4.1m tok/s: 7923441
+3000/20000 train_loss: 2.7808 train_time: 5.0m tok/s: 7925746
+recurrence:activated at step 3000, virtual_layers=[0, 1, 2, 3, 4, 5, 4, 5, 6, 7, 8, 9, 10]
+3500/20000 train_loss: 2.6942 train_time: 6.1m tok/s: 7526677
+4000/20000 train_loss: 2.6259 train_time: 7.1m tok/s: 7435275
+4000/20000 val_loss: 2.6493 val_bpb: 1.1514
+4500/20000 train_loss: 2.5771 train_time: 8.0m tok/s: 7366109
+5000/20000 train_loss: 2.5199 train_time: 9.0m tok/s: 7311315
+5455/20000 val_loss: 2.5389 val_bpb: 1.1034
+stopping_early: wallclock_cap train_time: 590005ms step: 5455/20000
+peak memory allocated: 30120 MiB reserved: 30154 MiB
+ema:applying EMA weights
+pre-quantization post-ema val_loss:2.53747920 val_bpb:1.10275923 eval_time:2002ms
+Serialized model: 132406149 bytes
+Code size: 24659 bytes
+GPTQ:collecting Hessians from calibration data...
+GPTQ:collected 66 Hessians in 9.7s
+GPTQ quantization: 66 layers with full GPTQ, 0 fallback to clip-search
+selective_prune: unpruned=16.60MB target=16.0MB
+selective_prune: pruning 4779096/8972248 lowest-error ±1 values (excess=597387B)
+Serialized model int6+brotli: 15134564 bytes
+Total submission size int6+brotli: 15159223 bytes
+final_int6_roundtrip val_loss:2.72431518 val_bpb:1.18395599 eval_time:8722ms
+final_int6_sliding_window val_loss:2.67696681 val_bpb:1.16337894 eval_time:76364ms