Residue eval redesign: expansion-head control arm (B), paired chain bootstrap, within-zone position analysis, contamination gate, readout-capacity ladder

DeepBLAST rerank: honour CACHE_ROOT/LEMON_ROOT, add RETRIEVER to the launcher

DeepBLAST rerank: let any model supply the candidate prefilter, record the retriever

Preflight: a residue_decoder checkpoint cannot shrink on head swap, so stop demanding it

Residue rebuttal: drop srun fan-out, make the residue-decoder column optional

Fix srun task fan-out in scorer; make --scored-prefix per-dataset; parameterize notebook results basename

Notebook: dense dataset-level and per-database tables, fixing the level-name NaN

Local interactive runner, embedding status/ETA tool, and opt-in partial/prefix scoring

Add lemondec retrieval variant; chain-level bootstrap CIs for residue probes; rebuttal suite; withdraw leaked SS3 report

Pin deepblast transformers to 4.28.0 and preflight the deepblast interpreter instead of the shared one

Add bpept and lemonpt MLM-pretrained retrieval variants using mean-pooled encoder features

Clear conflicting SLURM memory specifications before every srun call site

Add dedicated deepblast env with full install_requires and build-time import verification

Put the DeepBLAST checkout on sys.path before importing deepblast

Fix run_dplm2 embed_dataset entry_sha256 signature and rich version check in dplm env verification

Add rich to the dplm env and verify byprot import prerequisites at build time

Fix TM-Vec tm_vec import path, unpin biotite for Python 3.10, isolate env build failures, add esm3/dplm interpreter guards, drop plmsearch

Lock each model/dataset directory against concurrent writers and guard the final rename

Let an explicit BATCH_SIZE override the per-model default in the interactive driver

Fail fast with guidance when the interpreter cannot read the BPE vocabulary

Tolerate non-git code trees and add a zest venv with modern tokenizers for the BPE vocabulary

Add interactive priority-ordered embedding driver with resume and batch override

Select GPUs by GRES type with sinfo validation and round-robin, since PACE overrides --partition

Restrict GPU jobs to A100-or-newer partitions to avoid the V100 driver mismatch

Fix torch.load weights_only regression, clarify preflight failures, add local embedding path

Support running a subset of retrieval baselines end to end

Use version-stable gdown Python API for the DHR checkpoint download

Swarm external retrieval: parallel GPU embedding jobs with CPU-only scoring

Stage external retrieval assets on Phoenix

Inherit cluster PyTorch in remote retrieval environments

Use full sequences with model-specific retrieval limits

Add benchmark-aligned remote retrieval calibration

Fall back to latest frozen BPE checkpoint for end-to-end fine-tuning

Use two epochs for 10-hour full-encoder continuation

Make dependent full-encoder run select safe source checkpoint

Run joint decoder fully unfrozen on H100 or A100

Discover current benchmark data snapshot automatically

Run ESM2-650M as 2h head plus 8h end-to-end

Add end-to-end ESM2-650M fine-tuning baseline

Allow explicit retrieval benchmark dataset paths

Add retrieval-preserving joint decoder finetuning

Add decoder-only progress probe and baseline comparison

Prevent ESM residue probe host-memory OOM and add resume

Revert residue jobs to stable H100-only setup

Fix CUDA environment and prevent residue CPU fallback

Broaden residue jobs to all compatible Phoenix GPUs

Complete residue-head training and runbook snapshot

Make residue preflight verify probe runtime imports

Fix residue decoder startup and publish probe runtime

NetSurfP-3.0 auto-download (atomic+validated), corrected 68-col layout (no SS3 block, RSA@55, phi/psi@65-66, disorder polarity), dataset-name resolution, prefetch step in submit scripts, RUNBOOK