Record: 11L + order-adaptive 9-gram backoff (mean val_bpb=0.9059)

records/track_10min_16mb/2026-03-25_11L_order_adaptive_9gram/README.md

@@ -0,0 +1,33 @@
+# 11L XSA-all + Order-Adaptive 9-gram Backoff
+
+## Results
+
+| Seed | val_bpb |
+|------|---------|
+| 42   | 0.9067  |
+| 1337 | 0.9059  |
+| 2024 | 0.9050  |
+| **Mean** | **0.9059** |
+| Std  | 0.0009  |
+
+Artifact: 13.99 MB. Train: 600s on 8xH100 SXM. Eval: ~150s.
+
+## Architecture
+
+- 11 layers, 512 dim, 8/4 GQA heads
+- XSA (Exclusive Self-Attention) on all 11 layers
+- 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
+
+Multi-order backoff (2 through 9-gram) with order-adaptive entropy gating:
+
+- Higher-order matches (9-gram) get lower entropy threshold, trusting n-gram even when model is moderately confident
+- Lower-order matches (2-gram) require high model uncertainty before trusting n-gram
+- Formula: `center = 3.0 - 0.25 * (matched_order - 2)`, `alpha = 0.05 + 0.55 * sigmoid(2 * (H - center))`
+- Separate hash tables per order (no cross-order collision)
+- Score-first, deterministic, no TTT

records/track_10min_16mb/2026-03-25_11L_order_adaptive_9gram/submission.json

@@ -0,0 +1,9 @@
+{
+  "name": "11L XSA-all + order-adaptive 9-gram backoff",
+  "val_loss": 0.9059,
+  "bytes_total": 13991619,
+  "blurb": "11L XSA-all with order-adaptive entropy-gated 9-gram backoff cache. Mean 0.9059 (std 0.0009) across 3 seeds.",
+  "author": "hypery11",
+  "github_id": "hypery11",
+  "date": "2026-03-25"
+}