Found first: a primary source the press has not covered yet.
A paper posted to arXiv on 24 August 2026 formalizes prefix invariance for hybrid sequence models and proposes a two-pass audit that requires no training and no gradients. Tested against 192 injected causality faults across eight model checkpoints, the audit located every fault to the exact layer. Attention-mask inspection, described in the paper as "the field's default check," detected 0 of 192.
What the source says
Authors Taebong Kim, Youngsik Hong, Minsik Kim, Sunyoung Choi, Jaewon Jang, and Minseo Kim (no institutional affiliation is listed on the arXiv page) define prefix invariance as the requirement that a model's representation at position t must not depend on any future input. Their audit produces a per-layer causality score and, in the paper's words, "fits on one page and runs in seconds." Beyond the synthetic fault injection, the method also found a real inter-chunk axis error in two shipping models, Zamba2 and Nemotron-H, subsequently addressed via their reference implementations. The paper notes that leaks can propagate through SSM scans, aggregations, or normalization steps even when attention masks are entirely correct, which is why mask inspection is structurally blind to this class of failure.
Why it matters
Mask inspection is the dominant practical method for verifying causality in transformer and hybrid LLMs, and this paper shows it is provably incomplete for that task. The 0/192 detection rate is not a marginal failure: mask inspection gives no signal at all when the leak originates outside the attention mechanism. Zamba2 and Nemotron-H are production models, not research checkpoints, so the confirmed inter-chunk error gives practitioners running hybrid or SSM architectures a concrete reason to re-run verification. The audit's two-forward-pass design means it can be applied to any existing checkpoint without retraining.