Phase 1A: quantize LoRA guardian for swarm serving

Granite Phase-3a re-run on clean protocol (apples-to-apples with Phi-4)

Qwen-2.5-14B drift-inversion transfer (Direction 1, dose-response test)

Phi-4 drift-inversion transfer test (Phase-3a methodology, Direction 1)

bf16 diagnostic v2: capture per-window PPL for diffuse-vs-concentrated read

Add bf16-on-9 diagnostic per web-Claude (bound the gap)

Add alpha-sweep script: α∈{0.5,0.75,1.0}, full-set eval, α=1.0 control

Path B v3: full test set + same-protocol refs (6.5482 / 7.3814)

Validation: switch from n=40 to full wikitext-2 test set per web-Claude review

Path A fix: bump transformers>=4.51 for AutoAWQ qwen3 compat

Path B fix: handle dotless top-level layer names

Path A fix: stage manually to put model on cuda:0 before profile

AWQ POC supporting code + runners 2026-05-20