Granite Guardian 4.1 8B · file← Granite Guardian 4.1 8B

Q4_K_M

IBM's own file and our first judge: 60 out of 60. Deleted on 12 September. IBM official GGUF.

Size: 5.1 GB
Status: Deleted 12 Sep 2026
Scored runs: 2
Smallest Mac: 16 GB estimate, 8K context

Results2 scored · 95% intervals

Suite	Score	95% interval	Passed	Runtime · profile	Recorded
screeninggranite-guardian-4.1-8b	100.0%	94.0–100.0	Passed 60/60	llama.cppextraction	2026-08-31
screeninggranite-guardian-gguf-q4km	100.0%	94.0–100.0	Passed 60/60	llama.cppextraction	2026-09-09

Runs marked unenforced were measured on a suite that requires constrained decoding, on a server shown not to apply the schema: mlx_lm accepts it and ignores it, and a few other servers and models break it. Those numbers do not measure what the suite is designed to measure. They are kept and labelled rather than deleted.

The recordHow it was made · how it was served · what went wrong · what happened to it

Rebuilt from the lab's records by Claude, an AI model. The results above come straight from the automated runs. How these pages are made.

File	`granite-guardian-4.1-8b-Q4_K_M.gguf`
Format	Q4_K_M GGUF, 5,116,703,808 B
Where from	downloaded 2026-08-31 from IBM's GGUF repo (byte-exact against the HF manifest after a corrupted first attempt)
Status	deleted 2026-09-12 08:04Z
Hugging Face	https://huggingface.co/ibm-granite/granite-guardian-4.1-8b-GGUF

Where it came from

Repo / file: ibm-granite/granite-guardian-4.1-8b-GGUF / granite-guardian-4.1-8b-Q4_K_M.gguf, 5,116,703,808 B (remote manifest; local byte-exact on 2026-08-31). Revision not on record. sha256 not on record.
Commands: first curl -sL --retry 20 … -C - -o …/granite-guardian-4.1-8b-Q4_K_M.gguf "https://huggingface.co/ibm-granite/granite-guardian-4.1-8b-GGUF/resolve/main/granite-guardian-4.1-8b-Q4_K_M.gguf". That produced 5,818,044,416 B ("701 MB too large"; the resume appended onto a partial from an earlier aborted attempt). Deleted and re-fetched without -C -, byte-exact at 01:47Z.
Date / who asked: 2026-08-31; Alan, "Try the 2 main candidate models if they make sense".
llama.cpp support was gated first: LLM_ARCH_GRANITE present in the b10618 tree.
GGUF chat template present (6,099 chars).

How it was served

label	dates (UTC)	binary and version	ctx / -np / KV / -fa / --jinja / --reasoning-format	harness profile	why this label exists
`granite-guardian-4.1-8b`	2026-08-31 02:57Z–03:14Z	`llama.cpp/build/bin/llama-server`, b10618 `eb25b7263`	`--ctx-size 16384 --n-gpu-layers 99 --jinja`; -np, KV type and -fa defaults (not on the line); no reasoning format	extraction	the first measurement, standalone server so the router was untouched; box to itself after OOM beside Coder
`granite-guardian-gguf-q4km`	2026-09-09 11:12Z–11:38Z	same binary, b10618	`--ctx-size 16384 -ngl 99 --jinja --alias granite-guardian`	extraction	the GGUF arm of `guardian_validate.sh`, the baseline the MLX build was checked against
`granite-guardian` (12:47 and 12:50 rows only)	2026-09-09 ~11:30Z–11:50Z	same	same	extraction	a mistake: both validate arms were relaunched with `--model-id granite-guardian` so the pre-flight checked the same id the runner used, and the GGUF arm's results overwrote the oMLX arm's in one file

Seats held: the judge seat 2026-09-09 10:35Z → 18:55Z (stood down for REAP's memory). Earlier there was no seat; the 2026-08-31 run was a standalone test.

Run-log llama_build "0.2.0-dev (build 10618, commit eb25b7263)" matches the evidence (the next rebuild was 2026-09-10).

Decisions

Downloaded as one of "the 2 main candidate models" for the judge seat.
Kept as the judge after the MLX build failed on oMLX.
Stood down 2026-09-09 18:55Z by the assistant under Alan's "okay so what next"; kept on disk "for flagged-item judging".
Requantised to Q4_0 on 2026-09-11 as the cleanest one-variable test of the format lever.
Deleted 2026-09-12 08:04Z on Alan's "kill and q8 and q4_k_m safely?", as a "Q4_K_M with a Q4_0 sibling". The sibling had no quality run.
Not published (IBM's own file).

What went wrong

Corrupted resume (2026-08-31): curl -C - produced a plausible-sized but wrong file; caught only by comparing bytes with the HF manifest. Recorded in a lab note's "Gotchas".
Health check accepted a 503 (2026-08-31): the first run scored 0 coverage against a loading server.
Partial-run rows: 12:22:08 (33/60) and 12:27:20 (36/60) for granite-guardian-gguf-q4km, 12:47:47 (41/60) for granite-guardian. These are scans taken mid-run; the run log reports 100% coverage of the rows present.

What happened to it

Downloaded 2026-08-31 → benched 2026-08-31 (100%) → judge 2026-09-09 → stood down 2026-09-09 → requant source for the Q4_0 2026-09-11 → deleted 2026-09-12 08:04Z. No RESTORE entry. Re-downloadable from ibm-granite/granite-guardian-4.1-8b-GGUF (check bytes = 5,116,703,808).