gfx skill FINAL: masked runtime gate -5.1% vs backbone (3.974/4.187), replica 3.923; trained masters in q8_row; apply: cortiq skill apply nanbeige42-3b-q4t.cmf nanbeige42-gfx.skill.cmf -o gfx.cmf

coding skill FINAL: masked runtime gate -1.1% vs backbone (9.633/9.743); trained masters in q8_row; apply: cortiq skill apply nanbeige42-3b-q4t.cmf nanbeige42-code.skill.cmf -o code.cmf

coding skill (557 MB delta): masked replica -2.2% vs backbone (9.529/9.746), bare runtime gate +7.7% (mask not yet applied on quantized storage); phase A bottom uncaptured at 150 steps - a longer A will deepen it. apply: cortiq skill apply nanbeige42-3b-q4t.cmf nanbeige42-code.skill.cmf -o code.cmf

gfx specialist v2: strict-f32 phase A, bottom 4.225@4.4%, gate -1.5% vs backbone

standalone gfx skill (557 MB delta): cortiq skill apply nanbeige42-3b-q4t.cmf nanbeige42-gfx.skill.cmf -o specialist.cmf

Upload experiments/nanbeige42-gfx-pervisit-skill.cmf with huggingface_hub

Upload experiments/nanbeige42-gfx-pervisit-skill.cmf with huggingface_hub

card: name the library that reads this file

docs: discrete-card row — 94.2 tok/s on 0.5.49 (pipelined submission)

model card: measured speed with the hardware it was measured on (needs cortiq >= 0.5.35)

Upload nanbeige42-3b-q4t.cmf with huggingface_hub

Upload README.md with huggingface_hub

Upload README.md with huggingface_hub

initial commit