This paper reports a planned negative result arising from a systematic audit of the DDIN (Devavāṇī-Derived Interpretable Network) research program. A labeling circularity identified in earlier work, where semantic axis labels were derived from the same initial consonant used to construct acoustic features, is corrected, and the key experiments are re-run under a rigorous gloss-based labeling regime drawn from the Monier-Williams Sanskrit-English Dictionary. Under corrected labels, an Adaptive Exponential Leaky Integrate-and-Fire (AdEx) spiking reservoir with per-neuron BCM homeostasis processes the full 2,000-root Pāṇinian Dhātupāṭha under three controlled conditions. Acoustic-only input yields ARI = 0.013 (chance level); prior-only input yields ARI = 0.70; full acoustic-semantic fusion yields ARI = 0.69, a negative marginal acoustic contribution. The earlier Speed-1 result (ARI = 0.204 at first phoneme) does not replicate under corrected labels (ARI = 0.016). A Phase 13 architectural validation experiment rules out reservoir failure as the cause: the reservoir successfully recovers the Pāṇinian articulatory locus taxonomy (five consonant place-of-articulation classes) at ARI = 0.0494 (p < 0.001). A direct semantic-locus correlation experiment finds ARI = 0.0095, quantifying the near-orthogonality of the two taxonomies. The failure is task-specific: first-order acoustic formants encode articulatory locus structure, which is nearly independent of the Monier-Williams semantic axis structure at first order. Two engineering contributions from the program, the Epileptiform Synchrony Limit characterization (doi:10.5281/zenodo.19602055) and the BCM homeostasis architecture that resolves it, are independent of the labeling regime and remain valid. This paper offers the corrected empirical findings, a precise mechanistic diagnosis, and a characterization of what would be required for a positive phonosemantic result.
Amit Kumar (Sun,) studied this question.