-
decode.py: remove the qtip cross-check oracle (release decoder is pure numpy only)
-
README: gold6 numbers (mean retention 95.1% -> 96.3%)
-
BOARD: describe the shipped gold6 plain encode (was a trained checkpoint's numbers)
-
MANIFEST: payload yqi24gold3full -> yqi31gold6full
-
decode.py: trained-container entry (decode_expert_v3t/_layer_v3t) with exact op order
-
codec discriminator for the trained container
-
golden fixture: L0 e0 fp32 (all projections), gated vs served
-
update expert payload -> i31
-
update expert payload
-
update weights
-
Mach-1-Additive-35B + bf16 base vision tower (byte-identical clone, adds vision.safetensors)
-
initial commit
