From Internal Readouts to Governed Legal Reliability: Causal Intervention, Cross-Model Replication, and Harness Ablation under Affiliation Pressure

Mechanistic interpretability can reveal internal representations that never appear in a model's answer, but legal reliability is an end-to-end property of a system rather than a property of a readable activation alone. This paper reports Phase 2 of the Sub-Propositional Affiliation Signals experiment. A prospectively frozen intervention on Qwen2.5-0.5B moved internal and behavioral margins in the predicted directions but failed its competence gate. A separately frozen exploratory TRUE/FALSE repair produced a coherent within-model effect. Fresh legal cases then tested Qwen2.5-0.5B and SmolLM2-135M against matched random controls: the proposed mechanism was not confirmed across models. Behavioral tests with Sonnet 4.6 showed no affiliation-induced legal errors. A final preregistered comparison between bare Sonnet and configurations of an author-developed legal-AI product preserved appropriate uncertainty but exposed four opening-label contradictions on one temporal case. The study supports a layered evaluation rule: readout is not causation; causation in one model is not a general mechanism; behavioral robustness is not evidence of harness benefit without a same-model comparator; and correct reasoning is not operationally reliable when the machine-consumed decision field contradicts it. Interpretive boundary. The experiments do not establish consciousness, intention, mens rea, political identity, a universal affiliation mechanism, or general legal reliability. Mechanistic interpretability is used here to study internal model complexity, not to justify prior ideological screening. Reproducibility. Public protocols, synthetic cases, code, sanitized outputs, aggregate results, and hashes are available in the linked repository.