malaiwah/GLM-5.3-Flash-TR3-6bpw

Correct fidelity claims: preserve lane, panel and native-serving limits

Add snapshot-pinned QFS size–KL comparison plot

cards: corrected quantization scope (routed experts + MTP only; attention/dense-MLP were never quantized), excess_over_control rename (P1-05), document-level statistical correction (P1-15), power-arithmetic fix (CC-01). See quant-fidelity-suite docs/PUBLISHED-CORRECTIONS.md

cards: repo paths moved k6/ -> engines/ upstream; fix deep links (were 404)

disclose that materialization-receipt.json names a destroyed filesystem

card: scope disclosure - the published KLD is a panel25 number, and brandonmusic's calibration-clean scope moves it -14.91%

card: fidelity provenance annotation (model-index result + x_fidelity block) — body unchanged

docs: publish qualified K6/K8 serving evidence

card: parts-bin dataset published (548.5 GB, 223,389 files) — link added

docs: add live-qualified K6 serving recipe

card: BF16 floor removed — quantization-attributable error (K8 2.52x better than K6)

card: full-panel streaming numbers (top-1 96.56%, K8 sibling 0.012384) + streaming receipts

card: explicit Hub lineage note (two sibling roots), cross-links to the measured family, discovery tags

card: codec disclosure upgraded to verified-equivalent (120/120 byte-identical vs the now-published sealed R10 core)

SEALED: 0.013723 nats, five bitwise-identical cold runs, full panel — 1.5x better than FP8 at 77% size

card: honest serving section — vLLM/b12x reference compose (TP4, untested), status matrix, P2P wiki link

model card: provisional preview (run 1/5) — K6 beats FP8 on the shared window at 77% size

Add files using upload-large-folder tool

Add files using upload-large-folder tool

Add files using upload-large-folder tool

Add files using upload-large-folder tool

initial commit