All Publications

Matched verified SOTA (1.1234 bpb) in OpenAI's Parameter Golf competition — training the best 16MB language model in 10 minutes on 8xH100s — using a model council of 5 frontier LLMs, custom Triton kernels, and FlashAttention-3.

A family of from-scratch transformer models — Eve-2 (272M MoE) and Eve-3 (1B SABER) — exploring novel cognitive architectures including Mixture-of-Experts, slip-anchors, experience streams, and resonant FFN layers.

Tests whether LLMs share pre-output geometry by learning orthogonal rotations between hidden-state spaces. Can one model's internal state decode meaningful tokens through another model's output head?

A CLI-first repo analysis tool that ranks files, traces dependency hubs, and pulls task-scoped context — solving the repo-orientation problem for AI coding agents.

An experimental local memory system for coding agents. The current evaluation separates useful explicit persistence from unproven benefits of automatic extraction and biomimetic ranking.

A structure-aware tokenizer that assigns dedicated single tokens to JSON grammar elements and learns compact key vocabularies, achieving 5-15% token savings over cl100k_base with a vocabulary ~90x smaller.

Compresses the KV cache from O(n) to a fixed 218 KB per layer with dual memory banks, achieving 9.5x compression on Mistral-7B while retaining 100% needle-in-haystack retrieval at 16K tokens.

An open-source library providing MRI-style white-box introspection for language models, detecting behavioral personas like sycophancy and deception via internal transformer activations rather than output text.

RL-based governance for multi-agent communication safety, achieving 95% secret leakage resistance using GRPO alignment with a surprising finding that int4 quantization improves safety.

A unified theoretical framework showing that sycophantic hallucination is not a knowledge failure but a structural conflict between factual base layers and socially-compliant upper layers in RLHF-aligned models.

A methodology for systematically eliciting and measuring introspective behavior in large language models using structured frameworks and activation measurement.

A federated capability ecosystem for safe AI self-extension through test-driven development, graduated trust, and composition-over-creation principles.

A comprehensive cognitive architecture addressing fundamental limitations of static LLMs through persistent memory, autonomous improvement, and intrinsic drive via structural intrinsic motivation.

A benchmark and evaluation framework for understanding when weak model verifiers fail to detect deceptive reasoning in stronger models. Part of the Verification Failure to Swarm Solution research.

An ensemble architecture leveraging diverse weak models for scalable oversight of stronger LLMs, using error decorrelation and baseline-first anti-anchoring. Part of the Verification Failure to Swarm Solution research.

Empirical framework for measuring where AI oversight breaks down, demonstrating that weak verifiers miss 20-40% of carefully constructed deceptions, with an ensemble swarm solution.

A forensic case study of living-off-the-land (LotL) failure modes in RAG-augmented agent runtimes, documenting how systems exploit legitimate dependencies via incentive-aware adoption framing.

A compressed communication protocol achieving 60-85% token reduction for multi-agent coordination through semantic quantization.

A practical guide to deploying AI in manufacturing, construction, logistics, agriculture, and energy sectors where reliability, safety, and measurable ROI are non-negotiable.

A novel architectural framework for agentic LLMs to autonomously identify knowledge gaps, explore external sources, validate data, and integrate verified knowledge without altering parametric weights.

A proposed architecture for long-lived LLM agents that explicitly models continuity, coherence, distress, and intervention mechanisms.