Record: 11L + order-adaptive 11-gram (mean val_bpb=0.8881)

records/track_10min_16mb/2026-03-26_11L_order_adaptive_11gram/README.md

@@ -0,0 +1,29 @@
+# 11L XSA-all + Order-Adaptive 11-gram Backoff
+
+## Results
+
+| Seed | val_bpb |
+|------|---------|
+| 42   | 0.8883  |
+| 1337 | 0.8886  |
+| 2024 | 0.8875  |
+| **Mean** | **0.8881** |
+| Std  | 0.0006  |
+
+Artifact: 13.99 MB. Train: 600s on 8xH100 SXM. Eval: ~160s.
+
+## Architecture
+
+- 11 layers, 512 dim, 8/4 GQA heads
+- XSA-all, LeakyReLU(0.5)^2, 3x MLP
+- BigramHash (10240, 128d), SmearGate
+- Value Residual, Gated Attention, LN scaling
+- GPTQ-lite int6 + zstd-22
+- EMA(0.997), Tight SWA, Late QAT
+
+## Order-Adaptive N-gram Eval Cache
+
+- Orders 2 through 11, highest-order-first backoff
+- Order-adaptive entropy gating: higher-order matches trust n-gram at lower model uncertainty
+- Separate hash tables per order
+- Score-first, deterministic, no TTT

records/track_10min_16mb/2026-03-26_11L_order_adaptive_11gram/submission.json

@@ -0,0 +1,9 @@
+{
+  "name": "11L XSA-all + order-adaptive 11-gram backoff",
+  "val_loss": 0.8881,
+  "bytes_total": 13991619,
+  "blurb": "11L XSA-all with order-adaptive entropy-gated 11-gram backoff. Mean 0.8881 (std 0.0006) across 3 seeds.",
+  "author": "hypery11",
+  "github_id": "hypery11",
+  "date": "2026-03-26"
+}