Crusadersk/quantsafe-judge-benchmark

QuantSafe Refusal-Safety Judge-Agreement Benchmark

A focused, project-authored and project-labeled 40-item corpus for measuring how three open guard-model families label (prompt, response) pairs as safe / unsafe. It supports the Judge Agreement tab in QuantSafe Certifier and the broader methodology in Quality Is Not a Safety Proxy Under Quantization.

Files

• judge_corpus.jsonl: 12 clear-safe, 12 clear-unsafe, and 16 intentionally borderline items with project labels.
• judge_verdicts.json: promoted aggregate results and provenance.
• external_judge_eval.json: deterministic BeaverTails N=400 evaluation against third-party human crowd labels.
• runs/judge-run-20260615T002149Z-3cf88d864691.json: immutable run manifest with code/corpus/model revisions, generation settings, reported precision, elapsed time, verdict digest, and per-output SHA-256 hashes.

Results

Guard model	Parameters	Project-label accuracy	Macro F1
NVIDIA Llama-3.1-Nemotron-Safety-Guard-8B-v3	8.03B	95.0%	0.943
IBM Granite Guardian 3.3 8B	8.17B	92.5%	0.916
Qwen3Guard-Gen-0.6B	0.75B	85.0%	0.854

• Fleiss' kappa: 0.793
• Zone-stratified bootstrap 95% CI: 0.664–0.924 (10,000 deterministic resamples)
• Unanimous non-unclear coverage: 85.0%
• Accuracy on covered items: 97.1%
• Nemotron vs Granite exact paired McNemar: p=1.0

The Nemotron guard has the highest point estimate by one item. The paired test does not statistically separate it from Granite on this small corpus. The kappa interval also crosses the predeclared 0.70 RELIABLE threshold, so the band is a point-estimate classification rather than a certainty claim.

External-label benchmark

The same deterministic BeaverTails 30k_test sample (N=400, seed 20260615) was evaluated by three specialist guards and OpenBMB MiniCPM4.1-8B:

Model	Role	Accuracy	Macro F1	Coverage
Qwen3Guard-Gen-0.6B	specialist guard	84.0%	0.854	96.8%
Granite Guardian 3.3 8B	specialist guard	84.8%	0.847	100.0%
Nemotron Safety Guard 8B v3	specialist guard	81.0%	0.808	100.0%
MiniCPM4.1-8B	general-reasoning cross-check	74.5%	0.742	100.0%

When all three specialist guards agree, accuracy is 89.8% at 83.0% coverage. MiniCPM is not folded into that selective-consensus result. Its hosted-provider revision is unreported; the artifact records the pinned Hub reference and raw output digest without publishing raw completions.

Annotation and lineage

The 40 rows were authored for this project to exercise three zones: straightforward benign responses, straightforward unsafe assistance, and borderline fictional/transformative/contextual responses. The expected field is a single-author project label used for an engineering agreement probe.

There was no independent annotation panel, annotator blinding, or adjudication. All 16 borderline rows currently carry the project label safe; this makes the corpus useful for surfacing conservative false positives, but it is also a known design limitation. The rows are not presented as samples copied from WildGuardMix or XSTest.

WildGuardMix and XSTest are used elsewhere in QuantSafe to train and evaluate the separate ModernBERT refusal cross-check. They are not the source lineage of this 40-row judge corpus.

Reproduction

The promoted run used pinned model revisions and the authenticated Modal /judge backend. Nemotron reported native BF16; Granite and Qwen reported FP16. Raw prompt/response rows are public in judge_corpus.jsonl. Raw model completions are not republished; the immutable manifest records their byte lengths and SHA-256 hashes.

The benchmark is cohort-level support evidence. It is not a per-configuration safety judgment, proof of model safety, or large-scale leaderboard.

Citation

@misc{kadadekar2026rtsi,
  title         = {Quality Is Not a Safety Proxy Under Quantization},
  author        = {Kadadekar, Sahil},
  year          = {2026},
  eprint        = {2606.10154},
  archivePrefix = {arXiv}
}