out-of-core rework: embed_cluster_acts streams resids to disk memmap + streaming multi-GPU k-means (--kmeans-sample); build_data reads unified emb.f32 + GPU-batched Newton/IRLS probes

embed_cluster_acts: cluster on Qwen3 mean-pooled layer-27 residuals (drop BGE); scalable torch kmeans

build_data: hard-negative pairing in Qwen3 residual space (--hard-neg-k; subtle probe directions)

build_data: hold out probe-fit docs from SFT targets (no direction<->text contamination)

pretrain: default pack-len 0 (per-example+compile = validated 57% MFU; packing needs Blackwell flash-attn)

pretrain: sequence packing (block-diag attn, per-marker injection) — no-leak + inject proofs pass; MFU bench pending

cluster: faiss-gpu has no Blackwell kernels — CPU faiss + torch-GPU recluster_torch.py (reuses cached emb)

pretrain: round padded L to mult-of-64 (bounds torch.compile recompiles to ~2 shapes)

pretrain: MFU meter + length-bucketing + big-batch + optional compile

sae: loader + build_sae_data + eval_sae (cross-uplift metric); weights_only=False for torch 2.13

rl: pin TRITON_ATTN backend (FLASHINFER silently breaks vllm-lens injection on Blackwell)

rl: true per-token entropy bonus (maximize r + beta*H, explicit diversity knob, no KL)

scripts: attn sdpa (no flash-attn sm_103 build on Blackwell)

maxact-fast: hardening fixes (6 bugs) + Dr.GRPO vllm-lens rl.py + sbatch; Probe = universal direction

build_data: real validated LR probes (held-out AUC + activation-margin filter)

maxact-fast: clean repo — DESIGN + pretraining pipeline (embed/cluster/probe-cache/build-data/pretrain), Dr. GRPO RL to follow