I work at the intersection of non-equilibrium control theory, mechanistic latent monitoring, and high-throughput systems engineering. My work focuses on eliminating diagnostic lag in AI alignment by shifting safety from post-hoc output filtering to in-memory residual stream kinematics during inference.

Current Focus:

Lead architect of Auditor OS / V50-S Sentinel Core, a dual-head thermodynamic governor for autoregressive transformers. The framework separates generative traversal from boundary preservation, evaluating latent acceleration ($a_\perp$) across early, mid, and late layer strata to enforce 0-token prefill evictions and 1–3 token closed-loop decode steering on malicious trajectory shifts.

Core Artifacts & Prior Art:

• Theoretical Specification: Universal Thermodynamic Homeostasis & Epistemic Multi-Anchoring (DOI: 10.5281/zenodo.23315812)

• Evidence Specification: RFC-001: The .sent Cryptographic Evidence Container Specification v1.0 (Zenodo)

• Reference Implementation: Open-source PyTorch reference engine with passing test suites and O(K . dmodel) Fused Top-K steering on GitHub (Regsiren/Auditor-OS-V50-Sentinel).

• Live Demo: Interactive observer portal at sentinel.thohatventures.com.

Why I'm on Manifund:

To secure focused compute and runway for empirical red-teaming sweeps across open-weight frontier models (Llama-3.3-70B on HarmBench / AdvGLUE) to benchmark the physical boundary limits of sub-token latent tripwires under adversarial stress.