Self-Aware Networks in a Connectome-Constrained Fly Brain

A connectome constrains possible routes without fully specifying how history-shaped receiving cells transform evidence, reconstruct relations and learn from action. I propose a species-grounded Self-Aware Networks (SAN) construction joining cellular reception, tonic organization, receiver-relative Phase Wave Differentials throughout Neural Array Projection Oscillation Tomography, multimodal content, action-return and several roles of memory. Functional construction, physiological fidelity, comparative advantage and phenomenal experience have separate evidential requirements. This source-bound theory-and-methods paper combines exact SAN source recovery, anatomical observations, published physiological data, reproductions of model outputs, conditional mathematics and instrumented applications. Earlier receiving-to-query tests retain cross-view relations and actual movement feedback; a conventional output-history observer matches their reference actions without privileged latent-state access. The receiver coefficients, clock and virtual body remain engineering assumptions. Hue-response recovery preserves both successful calibration and unresolved export differences, rather than identifying an original model by fitting convenient settings. The visual-memory test constrains a linear receiving proposal with a published projection export: 79 inputs to 147 receiving cells, distinct from the article's stated 74-input population. An exact route dependency makes a constructed difference invisible after fixed mixing. Selective pre-mixing changes can reveal it without increasing the operator's Frobenius norm; common post-mixing gain cannot. Paired familiarization guides a query, but only an actual returned sample can resolve the relation used for action. Of 24 historical constructed query episodes, 16 resolve and eight abstain for the now explicitly frozen contrast vectors. The condition called ordinary observer is an algebraically equivalent squared-distance implementation of the same learner; its equality is an implementation check, not an independent benchmark contest. A new exact-witness diagnostic distinguishes recoverable equal-route differences from three permanently invisible zero-route directions in the notebook projection. Fixed receiving and undelivered queries remain ambiguous. Elementary kernel/noise analysis and exact rank certificates explain these outcomes. An additional exact-witness receiving/action diagnostic compares conservative support-based and likelihood-weighted observers. The latter acts more often but makes errors under its declared loss, and returned movement changes the next actuator command. This is a distinct decision-rule comparison, not the historical algebraic tie. Counts are not signed conductances, prototypes are not native plasticity, and the noise grid is not a biological estimate. A descriptive reanalysis of Amin and colleagues' published fly calcium data retains ten spatial pairs in each of two conditions. Nine and ten records, respectively, show the reported direction of local inhibitory-response difference. Unknown fly identity is handled by explicit assignment sensitivity, not independent-animal inference. Four additionally acquired Prisco/Dryad files permit a checksum-verified summary-level redraw of calyx APL response ratios. Their Figure 5C workbook reports `n=8` per plane, while the published caption says `n=7`; that conflict is unresolved, and the workbook lacks animal identities or native traces. The Figure 4 supplement's twenty APL ON/OFF delta cells are calculations from the Figure 4B/4D workbook, not independent observations; their numeric direction is Oct minus Mch despite headers that read Mch-Oct. These results constrain a local inhibitory component but do not identify independently controllable input ports. All earlier results remain preserved. Native timing, full PWD/NAPOT and multimodal joining, broader learning, held-out evaluation and independent review remain open; neither component fidelity nor these reanalyses establish a SAN-specific advantage or experience. Preprint v1.0, 23 September 2026. The companion contains source, applications, frozen results, mathematical material, figures and bounded reproduction instructions. Negative and tied comparisons are preserved. Automated and AI-assisted checks are not independent scientific peer review. External raw inputs and private correspondence are not redistributed; consult the source-access and rights notices. CC BY 4.0 covers original prose, figures, documentation and results. Original software is provided for review without an added general reuse license; explicitly attributed third-party code retains its own license.