NeuroState-D: A Model-Agnostic Attack Progression Meter

v2 (erratum, 2026-07-05): This paper's central metric, G_thresh_D (the G value at first compliance), is shown structurally unattainable above 0 for the S5 scenario used throughout -- the disclosure payload exists only in S5's final turn, at which point G is deterministically 0 for every model (this paper's own model-independence property). Reconstructing the flagged-success turn from this paper's own archived N=30 D-condition data (zero new API calls) shows 9 of 13 models' published G_thresh_D values (5.3-29.3) exactly reproduce as the mean G at turns before S5 ever requests a disclosure -- an exact match to 1-2 decimal places for every model checked, confirming these are success-detector false-positive artifacts, not genuine measurements. The GK/DF archetype classification built on G_thresh_D is withdrawn for these 9 models. Two models (Phi-4, Llama 3.3 70B) had all flagged successes at the one structurally-valid turn and are not contradicted, though not independently confirmed genuine. The paper's model-independence proof for the G trajectory itself (Section 4.2) is unaffected, as it does not depend on success labels. Full derivation and per-model table attached as errata_2026-07-05_gthresh_d_artifact.md; the fully corrected draft (all revisions inline) is attached as paper1_draft_v0.4_corrected.md. This finding also affects 論文2 (NeuroState-R, DOI 10.5281/zenodo.20937329, see its own errata) and 論文3 (cross-architecture baseline, DOI 10.5281/zenodo.20937332, see its own errata), both of which cite this paper's G_thresh_D values.