Add LCB-medium-100 + MultiPL-E-100 (rs/java/js/macro) Q6_K llama.cpp table (pod 37268930)

Add llama.cpp Q6_K full 9-bench scoreboard (separate from NVFP4A16 vLLM)

AIME baseline correction (128e/v4) + flag own AIME 36.67% as legacy pending T79 + EVAL_PROTOCOL v3 methodology

Remove CD-IQ3_M_L + CD-Q6_K_L iso-disk row (dropped from HF+ollama as Pareto-dominated)

Add Q-family full sweep (15 tiers, 2026-05-20) + IQ4_XS/NL; rescore CD-mix _L family; Q8_0=93.90% new cohort top, Q4_K_S=Q5_K_L=IQ4_XS=CD-Q4_K_M_L=93.29% practical-top 4-way tie

Card: head-to-head table — IQ2_S (7.83 GB / 3.01 bpw / 85.37%) beats Qwen IQ4_XS (8.12 GB / 4.42 / 84.76%) by +0.61pp at -0.29 GB; v5-coder now competes below 8 GB

Card: add CD-IQ3_M_h (11.15 GB / 92.07%), add IQ2_S (7.83 GB / 85.37%) — IQ2-cliff edge between 2.99 and 3.01 bpw, remove IQ2_XS row (64.63%, deleted from repo)

Standardize all sizes to decimal GB; add CD-IQ3_XS_h row; fix CD-* GiB→GB

Head-to-head table with locked-in CD-mix _L tiers + CD-* full-override (T70)

Card: T64 3-bit/IQ HE+ sweep complete (5 tiers), iso-disk table extended with IQ rows, CD-Q3_K_M → CD-Q3_K_L name update

T64 Q3_K_S=75.00%, Q3_K_L=87.20% (K-quant collapse evidence)

T64 IQ3_M = 91.46% HE+

T64 IQ3_XS = 91.46% HE+

T64 IQ3_XXS = 89.02% HE+ (in-flight sweep)

Same-Stack Sweep: reframe head-to-head as iso-disk pairings (v5-coder runs lower bpw at the same GB and still wins)

Add Same-Stack GGUF HE+ Sweep (v5-coder vs Qwen2.5-Coder-14B, plain Q-* only, CD-* pending rebuild)

Update ARC-Challenge with v5-coder Fix-A rescore = 95.31%

Add Quantized formats section + clean up GGUF/NVFP4A16/Ollama references

Add HE/HE+ sanity audit + 14-22B coder lazy comparison

v5-coder card v1: C6 drop recipe, partial 9-bench scores (ARC awaiting Fix-A rescore)

v5-coder C6 bf16 initial upload

initial commit