Research Program
Each project is a named thread within the SNT CoreX framework — combining formal mathematical theory, computational implementation, and empirical validation across domains.
Lie Algebra · Pure Mathematics
The algebraic foundation of the entire SNT program. Establishes the seven-generator Lie algebra G₇ = G₄(φ) ⊕ su(2) and derives the CoreX Spark constant φ★ = (1+√5)/2 as its unique spectral invariant. Derivation rigidity proven: the outer derivation space is exactly one-dimensional, with no semidirect product decomposition possible. Submitted to Communications in Algebra, currently with the editor.
Machine Learning · LLM Inference
Stability-oriented KV-cache compression for autoregressive transformer inference. Implements the Memory Inertia operator with a frozen-spark threshold at φ★. LongBench-v2 validated on Qwen2-7B and Qwen2.5-14B; 2.33× decoding speedup over StreamingLLM with higher task accuracy. Mass metric 0.443 ± 0.013, 7/7 ablation consistency.
Machine Learning · LLM Inference
Performance-oriented companion to ShrikeSNT. Query-aware adaptive KV pruning with entropy-driven keep ratios and query-aware Value boosting via the Nexter operator. Cross-model evaluation across Qwen2-7B, Mistral-7B, and Qwen2.5-14B. Mistral-7B shows the strongest gains over the full-KV baseline. Submitted to SN Computer Science after reformatting.
Machine Learning · Long-Session Agents
Utility-governed memory orchestration for long-session LLM agents. The Styx Field (seven operators) drives dynamic salience estimation, paired with the Lethe Layer — an embedding-based controlled decay buffer. Strongest variant reaches RR@k = 84.1%, 97.2% compression, and ~99% prompt token reduction with O(1) memory scaling. Currently being reformatted to meet ARR page-length requirements ahead of the August 2026 resubmission window.
Connectomics · Neural Dynamics
Eight-operator behavioral phase dynamics framework for C. elegans, built on the Cook et al. (2019) connectome. A semi-Markov phase controller makes competing locomotor drives structurally exclusive, reproducing sharp behavioural transitions that biophysical and mean-field models miss. The pool-dispatch variant matches calcium-imaging and dwell-time references closely and predicts all six held-out generative test cases correctly. Submitted to Biological Cybernetics.
Computational Biology · Genomics
Operator-constrained genome reorganization framework, validated across six organisms from bacteriophage to human mitochondrial DNA. BLAST validation confirms a 45% ORF hit rate (33 of 74 tested). Submitted to Theory in Biosciences and entered peer review within a day of submission.
Quantum Error Correction · Physics
Adaptive pre-decoding error-shaping layer for surface-code quantum fault tolerance. Operators renamed to neutral terminology (ItrCorr, SubMap, RelRoute) for a general quantum-computing audience. Simulated on Stim + PyMatching across distance-3 and distance-5 surface codes, showing 20–35% logical error rate reduction over standard MWPM decoding. Withdrawn from EPJ Quantum Technology and resubmitted to Quantum Information Processing (Springer).
Machine Learning · Unified Inference Architecture
The long-term unifying architecture for the ML thread of the SNT program: a dual-layer compression system combining CharonSNT's session-level memory orchestration with JuggernautSNT's attention-level KV scoring. Held pending maturation of its component papers.