We find where your AI hallucinates, fix it, and prove it.

Proprioception is your body's ability to sense its own position and movement without looking. When you close your eyes and touch your nose, that's proprioception. Language models lack this—they have no awareness of their own behavioral state.

We built small neural networks (probes) that read the hidden states of language models and detect behavioral patterns—hedging, repetition, sycophancy, shallow reasoning—before those behaviors manifest in the output. The model gains "self-awareness" of its behavioral tendencies.

Cognitive probes are tiny neural networks that attach to the hidden states of any language model. They read the model's internal representations and detect behavioral problems — like hedging, hallucination, shallow reasoning, or repetition — before they manifest in the output.

RLHF (Reinforcement Learning from Human Feedback) modifies the model's weights. It's expensive, requires human labelers, and often degrades capabilities. Our approach leaves the model frozen—we just read hidden states and intervene at decode time.

Better yet: our probes can replace human labelers for RLHF. Instead of paying humans to rate outputs, use probe scores as the reward signal. We call this Probe-Guided Reward Modeling. It's patented.

Yes. Detection happens inside the model, before output — so a response can be held before it's shown, and an agent's tool call can be gated until internal evidence checks pass. The same probes that measure a behavior can stop an unauthorized action from executing.

Separation measures how well probes distinguish between desired and undesired behavior. Prior published research achieves 2–5×. We achieve 125×–1,376×. That's the difference between a lab curiosity and a production-grade system.

We do not claim ownership of all AI architecture, all latent manifolds, every steering method, Anthropic's exact J-lens, or every autonomous research agent.

The stronger and more credible position is that our January–February 2026 record appears to place Proprioceptive AI early across a substantial emerging layer: pre-output internal-state sensing, low-dimensional operational state, streaming control, architecture-portable normalization, reversible intervention, post-action verification, agentic gating and evidence-governed improvement.

Patent priority is determined limitation by limitation from the actual as-filed specifications. The portfolio's practical strength therefore depends on continuity, written-description support, prior-art distinction and successful prosecution—not on the volume of filings alone.

Yes. The technology is architecture-agnostic — validated across 40 architectures from 410M to 405B parameters, including transformers, Mamba state-space models, mixture-of-experts, and attention-free designs like RWKV. You need access to hidden states during inference (standard in most frameworks). Training a probe for a new behavior takes about 20 minutes on a consumer GPU.

We're preparing enterprise licensing. Contact us at logan@proprioceptiveai.com for early access and partnership opportunities.

The internal portfolio reports 55 provisional filings and coordinated utility work spanning hidden-state monitoring, inference-time adjustment, per-token control, cross-architecture transfer, state-space embodiments, predictive control and governed improvement.

The strongest candidate dates are January 27, 2026 for the broad operational-manifold and read/control genus and February 4, 2026 for the more specific label-free cross-architecture normalization and transfer mechanism. Exact effective dates and enforceable scope require counsel to map every claim limitation to the official as-filed record.

Pre-revenue. Validated technology, 55 provisional patents, architecture-independent proof. First commercial deployments targeted Q3–Q4 2026 in clinical AI.

Research is published and archived on Zenodo:

Proprioceptive AI: Self-Compounding Behavioral Probes for Autonomous Model Improvement and Probe-Guided Compression — 82 pages

Consistency Is All You Need — 43 pages

Unified Behavioral Modulation in Large Language Models: Cross-Architecture Validation of Geometric Behavioral Subspaces — 32 pages

Controlled Language Models Inference-Time Control, Tokenization Engineering, and Reversible Optimization — 69 pages

Mathematics Is All You Need — 459 pages

Semantic Tube Prediction independently shows that hidden-state trajectory geometry can be decomposed into functionally distinct directions and manipulated to improve model behavior and data efficiency.

Anthropic's J-space research shows that a silent internal workspace can be read before output, causally edited, used for higher-order reasoning, and monitored for private assessments, fabrication and hidden goals.

Google's ScientistOne / Science One Framework independently converges on the governed-autoresearch layer: claim-level evidence chains, reproducible execution, method-code alignment and automated integrity audits.

Together these works provide strong external corroboration of distinct foundations of the Proprioceptive AI program. They are not direct replications of the complete Proprioceptor system or every reported result.

Deep Dive on our Architecture & Results

Click to watch Podcast ▸