Instruments Before Invariants: Fail-Closed Assay Calibration and Out-of-Sample Relational Steering in Qwen2.5-7B

Activation steering usually treats the residual stream as a passive Euclidean space in which intervention effect is governed by how far a vector is moved. That assumption is rarely tested as an instrument. This preprint reports a fail-closed sequence of frozen protocols on open-weight Qwen2.5 models that first diagnosed invalid event worlds, broken answer interfaces, and insufficient target-only prediction, then isolated one bounded confirmatory result. On Qwen2.5-7B-Instruct at decoder layer 20, a convex residual blend of a target state with a donor state (mixing weight 0.25) moved content-scored arithmetic margins by a mean of −0.988 nats without inducing syntactic recurrence. A subsequent target-only predictor associated with finite displacement on 160 held-out items (partial Spearman ρ = 0.288) but missed its frozen effect-size and donor-specificity gates. Replacing that scalar score with a pre-intervention, answer-independent dyadic derivative — the inner product of the clean top-1 versus top-2 decision gradient with the target-to-donor residual — produced a confirmatory pass on 192 fresh targets and 768 causal edges. Mean within-target Kendall τₐ was 0.440 (permutation p = 5.0 × 10⁻⁵; stratified bootstrap 95% CI [0.368, 0.509]). Directional concordance was 0.720. After residualizing ranks on target and donor fixed effects, Spearman ρ = 0.371. A direction-norm baseline was uninformative (τₐ = −0.037). Post-patch accuracy remained 97.7%. A follow-on second-order assay asked whether a mixed directional Hessian could predict nonadditive composition of two donor directions. It stopped on development data: primary τₐ = 0.101, self-curvature positive control τₐ = 0.095 against a 0.30 gate, and low/high-dose rank stability τₐ = −0.010. A post-hoc audit showed that bfloat16 quantized self-margin shifts to 0.125-nat steps while median low-dose Taylor predictions were 0.027 nats. The first-order relational ranking therefore stands; a stable local quadratic regime does not. The contribution is dual. Substantively, finite intervention response in this world is a property of a target–donor pair, not of raw vector energy and not of target susceptibility alone. Methodologically, competence and potency must be shown before geometry is interpreted, and a significant p-value cannot rescue a failed effect-size, specificity, positive-control, or stability gate. No result is offered as evidence of a global residual-stream manifold, a general cognitive architecture, or artificial superintelligence.