Path A: PR openai#1019 base + legal score-first TTT, no SLOT — pending H100 validation

Author: sahiee-dev

records/track_10min_16mb/2026-04-01_PR1019_TTT_Clean/README.md

@@ -0,0 +1,32 @@
+# PR #1019 Base + Legal TTT (Score-First SGD)
+
+## Architecture
+Built on PR #1019 (AR Self-Gen Full Hessian GPTQ + XSA-all 11L + BigramHash 3072x112).
+Added legal score-first TTT (SGD, lr=0.002, 3 epochs, chunk_tokens=32768, freeze_blocks=0).
+No SLOT. No n-gram caches.
+
+## TTT Legality
+Score-first: each chunk is evaluated under torch.inference_mode() before any weight update.
+Chunk N is scored, then model adapts on chunk N for future chunks.
+No multi-epoch pre-eval training. No future token access.
+
+## Results (pending)
+| Seed | val_bpb | val_loss | eval_time |
+|------|---------|----------|-----------|
+| 314  | TBD     | TBD      | TBD       |
+| 42   | TBD     | TBD      | TBD       |
+| 999  | TBD     | TBD      | TBD       |
+| mean | TBD     | TBD      | TBD       |
+
+## Run Command
+```bash
+TTT_ENABLED=1 TTT_LR=0.002 TTT_EPOCHS=3 TTT_CHUNK_TOKENS=32768 \
+TTT_FREEZE_BLOCKS=0 TTT_MOMENTUM=0.9 TTT_BATCH_SEQS=32 TTT_GRAD_CLIP=1.0 \
+SEED=314 \
+DATA_PATH=./data/datasets/fineweb10B_sp1024/ \
+TOKENIZER_PATH=./data/tokenizers/fineweb_1024_bpe.model \
+VOCAB_SIZE=1024 \
+torchrun --standalone --nproc_per_node=8 \
+records/track_10min_16mb/2026-04-01_PR1019_TTT_Clean/train_gpt.py \
+2>&1 | tee seed314_ttt.log
+```

records/track_10min_16mb/2026-04-01_PR1019_TTT_Clean/submission.json

@@ -0,0 +1,10 @@
+{
+  "val_bpb": 0.0,
+  "val_loss": 0.0,
+  "artifact_bytes": 0,
+  "seeds": [314, 42, 999],
+  "seed_bpbs": [0.0, 0.0, 0.0],
+  "eval_stride": 64,
+  "eval_times_ms": [0, 0, 0],
+  "status": "pending_h100_run"
+}