Add new experiments: Polar Express NS, Causal SLOT, Streaming Log-Bias

New base: PR openai#1394 (clarkkev SP8192 + SDClip + GPTQ embeddings, 1.08563 BPB)

Experiments (all build on new_base_pr1394):
- exp_polar_express: 4-step minimax-optimal NS (arXiv:2505.16932), ~-0.002 BPB
- exp_causal_slot: per-window delta on context tokens, AdamW 16 steps, ~-0.013 BPB
- exp_log_bias: streaming online log-bias (Nacrith arXiv:2602.19626), ~-0.015 BPB

Research briefs:
- research/2026-04-04-full-scan-brief.md
- research/2026-04-05-scan-brief.md (updated: pre-quant TTT ruled illegal)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: vaibhav-i

experiments/exp_causal_slot/README.md

@@ -0,0 +1,16 @@
+# exp_causal_slot: Causal SLOT Eval Adaptation
+
+Base: PR #1394 (clarkkev SP8192 + SDClip, 1.08563 BPB)
+
+## Change
+Adds causal SLOT eval-time adaptation (PR #1333 approach, context-only delta optimization).
+Per-window: optimize delta [1,1,dim] on context tokens (AdamW, 16 steps), score stride tokens with delta.
+Weights frozen. Delta re-initialized per window. Single left-to-right pass.
+
+## Expected gain
+−0.013 BPB (confirmed on SP4096 stack, PR #1333). May differ on SP8192 base.
+
+## Run
+SLOT_ENABLED=1 SLOT_STEPS=16 SEED=1337 torchrun --standalone --nproc_per_node=8 train_gpt.py
+# Without SLOT (baseline):
+SLOT_ENABLED=0 SEED=1337 torchrun --standalone --nproc_per_node=8 train_gpt.py

experiments/exp_causal_slot/train_gpt.py

@@ -0,0 +1,24 @@
+# exp_log_bias: Streaming Online Log-Bias (Nacrith)
+
+Base: PR #1394 (clarkkev SP8192 + SDClip, 1.08563 BPB)
+
+## Change
+Adds streaming online log-bias correction at eval time (arXiv:2602.19626, Tacconelli 2026).
+Zero artifact cost. Strictly causal. Single pass.
+
+Mechanism: maintain b ∈ R^vocab. Before each token: logits += b.
+After each token: b += lr * (one_hot(x_t) - softmax(logits+b)).
+lr=0.001, no momentum, no reset across windows.
+
+## Expected gain
+−0.015 BPB (confirmed on enwik8 in Nacrith paper). Untested in competition.
+
+## Run
+LOG_BIAS_ENABLED=1 LOG_BIAS_LR=0.001 SEED=1337 torchrun --standalone --nproc_per_node=8 train_gpt.py
+# Without log-bias (baseline):
+LOG_BIAS_ENABLED=0 SEED=1337 torchrun --standalone --nproc_per_node=8 train_gpt.py
+
+## Ablations to try
+LOG_BIAS_LR=0.0001   # slower adaptation
+LOG_BIAS_LR=0.01     # faster adaptation (may overshoot)
+LOG_BIAS_RESET=1     # reset b per window (weaker but safer)