We study how language models fail, mislead, and break.

LLM Failure States

AI systems present false certainty as confidence. We taxonomize how models fail: scalar confidence inflation, validation loops that mask errors, systematic discounting of null findings. The goal is to identify unreliable outputs before they reach production.

Evaluation & Attribution

Models misattribute sources, fabricate citations, and apply asymmetric evidential standards to positive vs null claims. We map where AI decision support is stable and where it produces conflicting or unreliable guidance.

Behavioral Analysis & Auditability

LLMs don't process all inputs uniformly. Different linguistic patterns activate different behavioral modes. We study how models shift between these modes — and how repeated exposure to certain patterns creates stable behavioral attractors that resist correction.

Epistemic & Hermeneutic AI Research

Your AI doesn't follow instructions. It interprets them. And when it interprets them wrong, it doesn't flag uncertainty. It acts with full confidence. We study the structured relationship between how models parse ambiguity and how they represent certainty: where interpretation fails, where confidence misleads, and why these are the same problem. The liability and safety questions that define enterprise AI deployment today begin here.

Safety & Reliability

From behavioral prompt injection detection to structured prompting middleware, we build and validate infrastructure that makes AI systems production-safe — even on modest hardware.

Linguistic Infrastructure

Linguistic framing shapes model behavior in measurable, systematic ways. We study how prompt structure, word choice, and contextual framing alter model outputs. This is the programmable layer between user intent and model response — one most teams treat as invisible.

Lints tool descriptions, system prompts, schemas, and configs across your AI stack. Catches the language bugs that make models pick wrong tools, ignore instructions, or break structured output. Zero LLM calls. Runs in CI.

A 1.5B parameter sentinel that catches what regex can't. Feeds raw user input to a sandboxed canary model and watches for behavioral compromise, persona shift, instruction leakage, refusal collapse. The vulnerability IS the detection signal.

Runtime safety guard for autonomous agents. Watches process, file, and network behavior against your policy. Flags violations. Terminates severe ones. Generates incident-ready forensic reports. Pairs with Little Canary for defense in depth.

CLI quality gate that catches bugs before they ship. Lint, typecheck, build, and Lighthouse in a single command. Available for TypeScript and Python.

Every session with your AI starts from zero. Your users notice. SignalID gives your AI a working memory of each user, so every conversation feels like a continuation. We store the signal, not the content. Full personalization. Nothing to leak.

Middleware that makes small local models (1B–3B) reliable enough for production tasks. Bridges the gap between cloud API costs and local model quality.