AI Mechanistic Interpretability and Self-Aware Networks: From Neural Decoding to Causal Self-Regulation

An internal representation can be decodable, causally effective and numerically reconstructable without implementing the operation a user intended. I examine these distinctions through the receiver-relative framework of Self-Aware Networks, connecting a source-recovered neural-decoding and feedback question to explicit transformer interventions. The research combines intervention prediction in controlled learned networks, passive and active feedback controls, GPT-2 activation and cache experiments, and native hybrid-state interventions in a pinned quantized Qwen3.5-0.8B decoder. Complete corrected-state transfer supports all eight tested bidirectional story groups in an initial Qwen assay, while attention-only transfer fails one; a complementary intervention gives identical attention state with different behavior. In a separate learned-correction assay, linear and nonlinear attention-state predictors each obtain five of 32 requested reversed-role answers but no story with both roles correctly reversed. Both preserve the tested color attribute. Lower training loss and partial-state error therefore fail to certify the intended relational change. Formal arguments distinguish observational from causal identification, partial-state from behavioral sufficiency, and output-margin guarantees from reconstruction loss. Earlier failures, nulls and inadequate baselines are retained. The contribution is a bounded experimental account of receiver-dependent intervention and its certification problem, not invention of mechanistic interpretability, evidence of phenomenal consciousness, or demonstrated autonomous self-regulation in frontier models. Preprint v1.0; integrated Draft 10, September 5, 2026. The checksum-linked shared companion, archived with DOI 10.5281/zenodo.22360017, includes the applications, protocols, data, figures, 31 accepted Lean statements under recorded assumptions, and a later full fixed learned-Qwen pipeline reproduction (49 serial steps; 465 original answers and 465 separate-code reconstructions). The same experiments support both AI papers; these are not independent replications. The negative learned-correction result is retained. Authoring-agent checks and same-workflow reconstruction are not independent scientific review. Base-model and runtime binaries and private source originals are not included. License scope: CC BY 4.0 for original manuscript prose, figures, documentation and data. Original code is public for review without a separate reuse license; third-party materials retain their own terms. See the companion LICENSE.md and notices.