July 15, 2026 default-brain audit: parent BF16 passed 33/33 fresh neutral reasoning probes and 33/33 Grug-clean probes, including tool-enabled agent systems, with original instruction-free template. Existing rocks already derive from that intrinsic checkpoint. grug not churn identical rock; filenames, weights, and hashes below remain same. This note records verification.

grug honest note: old rocks replaced after dialect repair. same filenames, new weights and new hashes. old rock knew club skill but sometimes thought polished tiny English. new rock comes from corrected merged bird. redownload file.

full bird: ProCreations/grug-v2-9b. pre-repair GGUF repo state preserved on branch pre-dialect-fix-2026-07-13.

choose rock

file	size
`grug-v2-9b-Q4_K_M.gguf`	5.24 GiB
`grug-v2-9b-Q5_K_M.gguf`	6.02 GiB
`grug-v2-9b-Q6_K.gguf`	6.85 GiB
`grug-v2-9b-Q8_0.gguf`	8.87 GiB

Q4_K_M: best default small rock
Q5_K_M: more accuracy meat, still compact
Q6_K: high-quality local rock
Q8_0: biggest popular rock, closest quantized weight

grug verify

all four rock converted fresh from same gated BF16 candidate. no quant-of-quant. all loaded and generated on GPU with llama.cpp commit 14d3ba45f3369e75a308212399cfada5d349883b before upload. exact hashes in SHA256SUMS; smoke logs summarized in smoke-results.json.

full merged bird exact greedy harness:

test	score %
HumanEval pass@1	82.9
MBPP pass@1	77.0
card valid / strict / right tool	100.0 / 100.0 / 100.0
broad valid / strict / right tool	100.0 / 100.0 / 94.1
held-out dialect-clean trace	100.0

quant-specific full benchmark not claimed. grug not invent number.

run

llama-cli -hf ProCreations/grug-v2-9b-gguf:Q4_K_M -n 512 -c 8192

need recent llama.cpp Qwen3.5 support. runtime must honor embedded chat template. grug reasoning stays in <think>...</think>. native XML tool club stays sacred.