Answers you can’t trust.
Confident mistakes, inconsistent responses or information that doesn’t hold up.
AI RELIABILITY. DEEP TECHNICAL EXPERTISE.
What is your AI
doing wrong?
Specialists in AI behavior.
Focused on your business.
Unreliable answers. Inconsistent agents. AI that falls short in practice. We use specialized tools and hands-on research to find what is going wrong with a model and test what helps. Start with a free conversation.
Let’s have a conversationChoose a time on our calendarYour first call is hosted by Logan; Nicholas and the team may join when relevant.
Free 30-minute call · No preparation needed
Years of independent AI research
Beyond the answer.Hidden-state research · Measured control
AI & LLM evaluationHidden-state expertiseControlled steeringPractical next steps
THE RESEARCH, IN VIEW
Look deeper.
See the decision.
An answer can look certain and still be wrong. Follow one real recorded question through the model’s output, internal measurements and the decision that followed.
Edited replay of one fixed-choice research experiment. The operator screen illustrates a review workflow; the displayed measurements and answer selection come from the recorded run.
Bring us a real AI problem.Let’s have a conversation
START WITH WHAT MATTERS
Your business has a problem.
Let’s understand it.
You don’t need to know why it happens.
That’s where the conversation starts.
Agents that lose their way.
Missed instructions, repeated work or unexpected behavior in a real workflow.
AI that’s hard to explain.
You can see the result. You need a clearer understanding of what to investigate next.
RESEARCH BUILT OVER YEARS
A deeper view
of model behavior.
Hidden states and AI interpretability are our specialty, built on years of independent AI research.
We investigate the internal signals a model produces as it processes information, how those signals relate to behavior, and how controlled changes affect the result.
Our toolbox includes the CYGNUS research system, diagnostic probes, evaluation harnesses, model adapters and research-enhancement methods. When model access allows, we can examine weights and hidden-state signals, then test candidate changes against a baseline.
We’re passionate about understanding how AI works—and turning that understanding into useful work for the teams building with it.
Explore the research behind our workTHE PROPRIOCEPTIVE APPROACH
01
Observe
Understand the behavior that matters.
02
Look inside
Investigate relevant internal model signals.
03
Test the change
Measure interventions against a baseline.
Research-led. Evidence-driven. Focused on your problem.
TALK WITH OUR TEAM
Your AI.
Our shared focus.
Book a conversation with Logan and our team.
Tell us what your AI is doing and what you wish it did instead. We’ll discuss whether evaluation, probes, adapters, a CYGNUS integration or a focused research engagement fits your situation.
No dataset to prepare. No technical brief required.
Just a conversation about your business.
EXPERTISE WITH A HUMAN APPROACH
Deep expertise.
A human approach.
Years of independent AI research. A lasting curiosity about how models work. And a team that cares about the people putting them to use.
Our work connects rigorous investigation with a straightforward conversation about your business.
Explore our public research
NEWS & RESEARCH FEATURES
Our work in the public conversation.
Company announcements, profiles and research features.
Yahoo FinanceCompany announcement
AP NewsCompany announcement
KTLA 5 Los AngelesCompany announcement
PR Newswire by CisionCompany announcement
ADVFNCompany announcement
Castle PlacementCompany profile
YouTubeResearch feature by Vinh Nguyen
BenzingaCompany announcement
ABC4 · Salt Lake CityCompany announcement
WGN9 · ChicagoCompany announcement
FOX8 · ClevelandCompany announcement
WFLA · TampaCompany announcement
PIX11 · New YorkCompany announcement
WKRN · NashvilleCompany announcement
FOX59 · IndianapolisCompany announcement
FOX8 · GreensboroCompany announcement
WAVY10 · PortsmouthCompany announcement
ABC27 · HarrisburgCompany announcement
FOX4 · Kansas CityCompany announcement
FOX31 · DenverCompany announcement
KXAN · AustinCompany announcement
KOIN6 · PortlandCompany announcement
WOOD8 · Grand RapidsCompany announcement
KRON4 · San FranciscoCompany announcement
WTNH · New HavenCompany announcement
FOX5 · San DiegoCompany announcement
NBC4 · ColumbusCompany announcement
8 News Now · Las VegasCompany announcement
CBS17 · RaleighCompany announcement
WRIC · RichmondCompany announcement
KRQE · AlbuquerqueCompany announcement
WWLP22 · SpringfieldCompany announcement
KFOR · Oklahoma CityCompany announcement
Arizona Daily Star · TucsonCompany announcement
WPRI12 · ProvidenceCompany announcement
NEWS10 · AlbanyCompany announcement
WREG · MemphisCompany announcement
WJHL · Tri-CitiesCompany announcement
KSN · WichitaCompany announcement
WANE15 · Fort WayneCompany announcement
WATE6 · KnoxvilleCompany announcement
WKBN27 · YoungstownCompany announcement
KARK4 · Little RockCompany announcement
KHON2 · HonoluluCompany announcement
Rochester First · RochesterCompany announcement
KHQ · SpokaneCompany announcement
FOX40 · SacramentoCompany announcement
A LITTLE MORE CONTEXT
Good questions.
Straight answers.
Do I need to be technical?
No. Explain what your AI is doing, where it causes problems and what a better result would look like. We’ll help with the technical questions.
Can we talk if we use a third-party AI platform?
Yes. Start with a conversation about your setup. We can discuss evaluation and workflow problems; deeper internal-state work depends on the access available to the model.
What happens after the first call?
We’ll talk through the next steps together. If our team can help, we’ll explain how and let you decide whether it’s a fit.
Who will I speak with?
Logan hosts the first call. Nicholas and other team members may join when relevant. Together, we bring research, engineering and business perspectives to your AI workflow and discuss the right next step.
For technical teams: what can the work include?
Depending on scope and model access, work can include probe-based evaluation, case-level failure analysis, reusable test harnesses, adapter experiments, weight and hidden-state inspection, controlled steering and CYGNUS research workflows. We compare proposed changes with baselines and relevant controls. Tool access, use and engagement scope are agreed after the introduction.
See an example deliverable →LET’S TALK ABOUT YOUR AI
You see the problem.
Let’s find the next step.
Start with a free, 30-minute conversation with our team.