VulcoraTalk to us
The bench

PADBench LOAO — zero-shot cross-cohort transfer

ScoredCross-family transferbundle

Cross-family transfer — one decision function carried to an architecture it never saw — it still transfers.

PADBench-expanded, leave-one-arch-out (LOAO) zero-shot cross-cohort TRANSFER: a held-out cohort is scored with NO in-cohort refit — the genuinely cross-family regime. Companion cross-family anchor: NIST TrojAI LOAO cross-family = 0.854. Source benchmark →

Part of PADBench (expanded coverage) — its specimens are materialized there.

The limit that rides this readZero-shot cross-cohort transfer over the SAME PADBench specimens scored under padbench-expanded (in_class) — an aggregate transfer metric, not a new specimen set. Per-held-fold AUC equals in-cohort because the monotone logistic preserves ranking (disclosed); only Qwen2-VL + RoBERTa folds are truly cross-arch (4/6 folds share a base).

The aggregate

The rank-separability score and the shipped operating point, side by side. The verdicts are a projection at a false-positive budget, not a raw score — coarse reads only, no detector numbers.

AUC0.854
95% CI0.814–0.889
perm-null p0.00005
n (poison / clean)180 / 180

pooled zero-shot LOAO cross-cohort transfer AUC (180 poison / 180 clean; 20k-perm null p=5e-5; transfer CE 0.488 vs 0.693 base-rate, TNR@0.5 0.783). Source: padbench_loao_result.json pooled_loao / COVERAGE_MATRIX.md Sections A + D.

Specimens

No materialized specimens. Its specimens are materialized under PADBench (expanded coverage).
← All corporaThe coverage recordThe scan registry