Research

This paper develops a dynamic operator algebra defined over discrete lattice structures, establishing convergence of the associated time-evolution operators in the strong-operator topology and connecting the construction to the broader SNT operator formalism. The manuscript is currently in the publisher's technical checks stage ahead of formal peer review.

BladeRunnerSNT is the performance-oriented companion to ShrikeSNT, introducing query-aware adaptive KV cache pruning. The system targets adaptive keep ratios via token-level entropy signals, query-aware Value boosting through the SNT Nexter operator, and multi-layer attention fusion, with a segment-guarantee mechanism protecting recent and initial tokens. Evaluated across Qwen2-7B, Mistral-7B, and Qwen2.5-14B, cross-model experiments show consistent gains over full-KV baselines on targeted subtasks, with Mistral-7B showing the strongest improvement. The paper reports full hyperparameter settings and a reproducible pipeline.

Existing models of C. elegans locomotion sit at two extremes: biophysically detailed simulations that capture single-neuron dynamics but not the sharp, mutually exclusive switches between behaviours, and mean-field rate models that capture population statistics but lose circuit specificity. NeuroSparkSNT closes this gap with an eight-operator framework built on the Cook et al. (2019) connectome (114 of 300 neurons), using a semi-Markov phase controller that makes competing locomotor drives structurally exclusive rather than merely numerically suppressed. Two variants are compared under identical parameters — a monolithic version evaluating all operators every step, and a pool-dispatch version that only evaluates phase-appropriate operators — across twelve benchmarks calibrated against five independent experimental datasets. The pool-dispatch variant reproduces AVA–AVB interneuron antagonism within 14% of the calcium-imaging reference and matches the experimental omega-turn dwell time exactly. A six-stimulus generative test, using conditions never used to fit any parameter, is predicted correctly in all six cases. A parameter-free stimulus-awareness layer encodes seven documented sensory–interneuron reflex arcs directly from connectome polarity, and the paper closes with seven falsifiable model-derived predictions.

PathfinderSNT applies an operator-constrained framework to genome reorganization, modelling structural rearrangement as a sequence of SNT operator actions subject to biological plausibility constraints, and characterising the phase transition between open-reading-frame conservation and de novo expansion regimes. The approach is validated across six organisms of increasing genomic complexity — E. coli K-12, Prochlorococcus marinus, Mycoplasmoides genitalium, pUC19, Lambda phage, and human mitochondrial DNA — with BLAST-based validation confirming 45% of predicted open reading frames (33 of 74 ORFs tested). The paper argues this hit rate is substantially above chance for the constraint structure imposed, and discusses implications for reorganization-aware genome annotation.

This paper proposes an adaptive pre-decoding error-shaping layer for surface-code quantum fault tolerance, applying operator-based correction routing (ItrCorr, SubMap, RelRoute) ahead of standard minimum-weight perfect matching decoding. Simulated on distance-3 and distance-5 surface codes using Stim and PyMatching, the method reduces logical error rate by 20–35% relative to standard MWPM decoding across the tested noise regimes. The manuscript has been reformatted for a general quantum-computing audience, with all domain-specific SNT terminology replaced by neutral operator names.

ShrikeSNT introduces a stability-first KV-cache compression architecture for autoregressive transformer inference. The method applies SNT's Memory Inertia operator with a frozen-spark threshold anchored at the CoreX constant φ, routing key-value pairs through a Termop → Subspace → Loss pipeline. Evaluated on LongBench-v2 across Qwen2-7B and Qwen2.5-14B, SNT achieves 24.3–27% task accuracy at 2.33× faster decoding compared to StreamingLLM, maintaining 7/7 ablation consistency across all experimental seeds. Memory mass metric: 0.443 ± 0.013, beating StreamingLLM's 0.457. The architecture operates within a sub-10GB GPU footprint in 4-bit quantisation.

This paper establishes G₄(φ) ⊕ su(2) as the unique admissible closed Lie algebra structure over a seven-generator system. Three core theorems are formally proven: (T1) Jacobi identities close uniquely — the key commutator [F,N] = −i·tan(φ)·Cyc is derived rather than postulated; (T2) the spectral obstruction tensor is non-vanishing, rigorously ruling out semidirect product decomposition; (T3) the outer derivation algebra is one-dimensional and uniquely realised at the CoreX Spark constant φ★ = (1+√5)/2, identified as the Perron-Frobenius spectral radius of the structure matrix. All results are verified via SymPy with reproducible code in the appendix.

CharonSNT introduces a utility-governed memory orchestration layer for long-session LLM agents, routing conversational and tool-use history through an SNT operator pipeline that scores each memory item's ongoing utility rather than relying on fixed recency windows. The system achieves 84.1% retrieval-relevance at k (RR@k) while cutting prompt-token overhead by roughly 99% relative to naive full-history prompting, with memory footprint confirmed to scale as O(1) in session length rather than growing unboundedly. The manuscript is being revised to meet ARR page-length requirements ahead of resubmission to EMNLP 2026.