Proprioceptor — On-Prem AI Audit, Fix & Certification Platform benchmark record · Feb 4, 2026 Expanding research convergence

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

Proprioceptor reads internal activity across heterogeneous AI systems — including transformers, Mamba, RWKV and MoE — to flag patterns associated with hallucination and deception before output, apply targeted corrections, and return signed, capability-checked evidence. In an on-prem deployment, model data remains inside your environment.

Request a Pilot → See it work ↓
ProprioceptorInside the control loop On-prem
Short video preview
Watch the short introduction
  1. FindInternal signals
  2. FixBounded correction
  3. ProveSigned evidence
Your model. Your environment.
55
Provisional Filings
40
Architectures Validated
85.8%
Reported Hallucination Reduction
1,376×
Peak Separation vs. Null

Reported figures reflect internal benchmark conditions and should be read with the associated methods, controls, nulls and test protocols.

What We Do

Find. Fix. Prove.

Many evaluations inspect only what a model says. Proprioceptor also measures internal-state patterns linked to behavior — closer to an MRI for an AI system than a surface-level filter.

01

Find

Trained behavioral probes read selected hidden states and flag patterns associated with hallucination, deception, manipulation and sycophancy before output. Internal held-out evaluations report AUC 0.96–0.999 under defined test conditions.

02

Fix

Trained correctors reduce targeted failure modes — internal benchmarks report an 85.8% reduction in confident-wrong output — while a dedicated capability probe checks for regressions. No full retraining from scratch.

03

Prove

You receive a cryptographically signed, capability-checked certificate documenting the measured outcome on your model's outputs — evidence suitable for technical review, customer diligence or governance workflows.

On-prem deployment — model data remains inside your environment One framework across validated heterogeneous architectures Zero-shot transfer in validated cross-model settings Runtime gate — policy-disallowed actions can be held before execution Honest by design — even nulls are signed

Built for finance, healthcare, defense & sovereignty, and frontier AI-safety labs.

See It Work

Watch a model get caught, corrected & certified

One simulated pass of the console: detection, correction, certification — and a runtime gate that can hold policy-disallowed actions before execution.

PROPRIOCEPTOR · LIVE AUDIT Simulated
How It Works

From hidden states to a signed verdict

Deploy on your hardware — or within a sealed enclave you control — with results returned as cryptographically signed evidence.

01
Your Model
Supported architectures, frozen weights
02
Hidden States
Mid-depth activations, on-prem
03
Random Projection
JL-512, seed-pinned
04
Probe Stack
Fiber + SAE behavioral probes
05
Signed Verdict
Dual-Ed25519 certificate

Then control it — in real time

Steer

Nudge activations toward the behavior you want, mid-generation.

Edit

Apply bounded suppression or amplification to selected behaviors without full retraining.

Fine-tune

Bake the corrections back into the weights when you're ready.

Gate

Hold policy-disallowed actions — a tool call can remain pending until its internal evidence checks pass.

Where We Fit

Internal-State Control Beyond Output-Only Evaluation

Outputs matter. Internal state adds earlier signal.

When a failure is detected only after an answer or action is produced, intervention comes late. Proprioceptor adds a pre-output internal-state layer: it measures behavior-linked signals, applies bounded targeted corrections, remeasures the result and produces signed evidence.

Method CategoryWhat It Does WellTypical Limitation
FOutput Filters & RulesFast, inexpensive surface checks and policy enforcement.Operate on prompts or outputs; they may miss latent-state failures and usually intervene after generation.
JLLM Judges & Constitutional EvaluationScalable semantic review, policy scoring and natural-language rationale.Add another model to the loop; judgments can vary with prompting and are generally post-output.
RRLHF & Preference OptimizationBroad behavior shaping through training-time preference signals.Can be costly to repeat at scale; weight changes may create trade-offs and provide limited event-level observability.
IActivation Monitoring & InterpretabilityDirect access to internal representations and powerful causal research tools.Often model-specific or research-oriented; sensing, correction, rollback and certification are rarely integrated.
PProprioceptive AISense + Control + VerifyPre-output monitoring, targeted intervention, post-action remeasurement and signed evidence across validated heterogeneous architectures.
Our Solution

Today's AI Is Flying Blind

Most deployed models lack an explicit operational-state interface.

When your body moves, you feel it — that sense is called proprioception. Most AI systems do not expose a dedicated, inspectable representation of their own operational condition, so problems often become visible only after output. Proprioceptor supplies an external sensing and control layer that monitors internal activity and model vital signs, so relevant risks can be identified earlier and measured after intervention.

Proprioceptive AI

We give AI systems an external operational-state layer that measures behavior-linked internal signals before they appear in an answer — analogous to proprioception, without implying human-like consciousness.

  • Watch: monitors selected internal activity in real time while leaving the base model weights unchanged for observation
  • Fix on request: applies a small, targeted correction to reduce a measured behavior without requiring a full retrain from scratch
  • Heterogeneous AI designs: one framework has been validated across transformers, Mamba, RWKV and mixture-of-experts systems
Independent Multi-Lab Convergence

Independent research points toward a compatible broader architecture

Across different teams, models and methods, a growing body of research explores actionable internal geometry, low-dimensional control, pre-output sensing and evidence-governed autonomy. Our convergence map identifies the specific connection each source has to the Proprioceptive AI program.

Trajectory Geometry26 Feb 2026

Semantic Tube Prediction

Huang · LeCun · Balestriero

Convergence: supports the view that hidden-state geometry can be functionally meaningful, directionally decomposable and actionable — not merely visualized.

Study details & scope

STP treats token sequences as geodesic trajectories, separates geodesic-parallel signal from perpendicular deviation, and shows that acting on this decomposition improves signal-to-noise ratio and data efficiency.

Boundary: a training objective over token trajectories; not a replication of Proprioceptor's behavioral channels, runtime controls or certification system.

Geometry-Aware Control6 May 2026

Manifold Steering

Wurgaft et al. · Goodfire and collaborators

Convergence: supports geometry-aware internal control rather than assuming that every useful intervention is a single Euclidean direction.

Study details & scope

Manifold Steering shows that interventions following learned activation manifolds generate more natural behavioral trajectories than Euclidean linear steering, including demonstrations across language tasks and a video world model.

Boundary: tightly controlled manifold experiments, not the architecture-portable operational-state interface or full read–act–verify control plane.

Internal Workspace6 Jul 2026

Anthropic J-space

A Global Workspace in Language Models

Convergence: supports the existence of consequential pre-output internal state that can be monitored and causally influenced.

Study details & scope

Anthropic identifies a silent internal workspace that can be read before output, deliberately modulated, causally edited and used for flexible multi-step reasoning. Its J-lens also surfaces private recognition of fabricated data, evaluation awareness, prompt injection and hidden goals.

Boundary: a Jacobian-lens workspace centered on verbalizable concepts; not Proprioceptor's model-independent instrumentation or governed correction product.

Governed Autonomy25 May 2026

Google ScientistOne

Chain-of-Evidence · CoE Integrity Audit

Convergence: supports the governance principle that autonomous discovery should remain separated from judgment and bound to reproducible evidence, protected evaluators and explicit failure records.

Study details & scope

ScientistOne performs end-to-end autonomous research while maintaining claim-level evidence chains and auditing score reproduction, specification compliance, reference existence and method–code alignment.

Boundary: validates the evidence-governance layer, not Proprioceptor's hidden-state sensing or inference-time behavioral control.

Research update · April–September 2026 Explore 11 additional convergence references Geometry & control · Monitoring & verification · Efficient computation

How to read the connections. The closest overlaps with our low-rank space methodologies concern compact behavioral representations, geometric sensing and targeted internal control. Other sources inform the surrounding monitoring, verification and efficiency architecture. The connection statements below are our technical interpretations of the authors’ reported work; they do not imply an identical low-rank method, endorsement or use of our IP. Updated ; dates identify the cited arXiv version or dated institutional publication.

Internal geometry & control

Six research references
Feedback control

Local Linearity of LLMs Enables Activation Steering via Model-Based Linear Optimal Control ↗

Skifstad, Yang and Chou use locally linear activation dynamics and layerwise Jacobians to construct feedback controllers that steer internal states toward semantic targets.

Connection: internal representations become operational control variables, linking state prediction to bounded intervention. The shared theme is model-based activation control; this does not establish the same low-rank decomposition or portability mechanism.

Sense & intervene

Hallucination as an Anomaly: Dynamic Intervention via Probabilistic Circuits ↗

Nielsen and colleagues model residual-stream density with probabilistic circuits and use an anomaly signal to trigger corrective decoding selectively during generation.

Connection: an internal-state measurement drives a targeted correction loop. Its density-based detector and decoding intervention are specific implementations of that broader sensing-and-control pattern.

Low-dimensional control

Steering Robustness into World Action Models via Mechanistic Interpretability and Optimal Control ↗

Hong and colleagues identify low-dimensional separation between robustness-relevant activation patterns in some world action models and introduce a reduced-order feedback controller, WA-LQR.

Connection: one of the most direct links to compact behavioral subspaces used for control. Results vary by architecture, so the evidence supports a specific low-dimensional control principle rather than universal transfer.

Pre-output signalsRevised

Where Steering Signals Come From: Activation Source Selection in Activation Steering ↗

Ye and colleagues find that states immediately preceding a behavior can yield stronger steering signals than states selected merely because the behavior appears in text. Tail subtraction removes shared contextual components.

Connection: behaviorally useful internal signals can precede visible output, and separating irrelevant components improves their use for intervention.

Directional geometry

Feed-Forward Steering in Transformer Residual Dynamics ↗

Mudarisov, Burtsev and State model feed-forward networks as local steering fields. Their interventions distinguish the effects of tangential and radial components on residual dynamics and model quality.

Connection: decomposing internal geometry reveals functionally different directions for intervention. This is a geometric dynamics result, not a replication of our behavioral probe or certification system.

Geometric sensing

HalluTracer: Hallucination Detection via Depth-Averaging Truth Signals ↗

Guo and colleagues aggregate linearly accessible truthfulness signals across layers before the first answer token. Their geometric analysis explains how depth averaging reduces layer-specific noise.

Connection: internal representations can provide an early behavioral measurement. The work supports pre-output geometric sensing; its depth-aggregation detector does not itself demonstrate a correction loop.

Monitoring & verified improvement

Research and deployment evidence
Deployment evidence

OpenAI: activation monitoring during generation ↗

OpenAI describes classifiers that inspect internal activations, pause streaming when a concerning pattern appears, and route the content to a separate check before blocking or resuming. Its August update also describes token-level classifiers and escalation.

Connection: practical adoption of internal-state sensing combined with an external decision gate. These classifiers are trained per model; the documentation does not establish low-rank methods or our cross-architecture transfer approach.

Read the dated 18 August safeguards update ↗

Monitor evaluation

Measuring Activation Control in Large Language Models ↗

Kowalski and colleagues measure models’ ability to modulate activation direction and magnitude through instructions. They also find that this control can imperfectly evade several activation-based monitors.

Connection: consequential internal state can be deliberately modulated, making independent checks and monitor robustness essential parts of a control architecture. The study informs verification requirements rather than validating any monitor’s reliability.

Governed improvement

Anthropic: Automated researchers can reliably mitigate alignment failures ↗

Anthropic reports automated research loops that develop and test mitigations across ten alignment failures, with capability-preservation constraints, withheld evaluations and oversight of researcher actions.

Connection: autonomous improvement becomes a measured, evaluated process. This supports the governed-improvement layer of the program, rather than a particular low-rank sensing method.

Geometry & efficient computation

Adjacent architectural connections
Conditional computation

Language Models Can Control Their Own Attention ↗

Ho and colleagues introduce declarative attention: models express attention-scope choices that the inference engine uses to skip unnecessary key-value reads during long-context tasks.

Connection: model computation can be selectively controlled to improve efficiency. This is adjacent to the broader efficiency program; it does not demonstrate a low-rank behavioral monitor.

Constrained geometry

oHC: Orthogonal Hyper-Connections on SO(4) via Quaternions ↗

Guo and colleagues constrain mixing between residual streams using orthogonal geometry, with a quaternion parameterization for four streams that preserves norm and representational diversity.

Connection: explicit geometric structure can support stable computation. The SO(4) residual-mixing construction is an adjacent architectural result, not evidence of an identical algebra or low-rank control method.

Combined scientific meaning

Sense the state. Respect the geometry. Act on the state. Prove what happened.

SenseJ-space shows silent internal state can be read before output.
ActSTP, Manifold Steering and J-space show internal geometry can change outcomes.
VerifyScientistOne shows autonomy must remain bound to evidence and independent audit.

Together, these works provide meaningful independent, multi-axis corroboration for the research program behind Proprioceptive AI. No single paper reproduces the complete system; considered together, they strengthen the architectural case beyond any one line alone.

Boundary: these works do not collectively replicate Proprioceptor's exact decompositions, cross-architecture transfer results, trained correctors, runtime authorization, signed certification stack or complete governed-improvement system.

Chronology discipline: comparisons use original paper dates and original dated records. Scientific convergence, patent priority and infringement are separate questions; exact patent entitlement depends on what each as-filed application actually supports.

Read STP ↗ Read Manifold Steering ↗ Read Anthropic J-space ↗ Read ScientistOne ↗
Emerging AI Architecture

An early position in the next AI control plane

Recent research increasingly points beyond post-output evaluation toward systems that expose, predict, govern and verify internal operation. Proprioceptive AI's strongest position is not ownership of every hidden-state method — it is an early, integrated architecture for the control layer around increasingly capable models.

Family I · Parent Architecture

Dual-Path Neural Computation

Task production + operational-state pathway

A primary pathway produces text, predictions, plans or actions. A functionally distinguishable proprioceptive pathway represents the system's own operational condition while computation is still underway. Sensors read it; validated controllers act; a verifier accepts, adapts or reverses the intervention.

Designed to support: both retrofit products attached to frozen models and native future architectures with a trained operational dynamics core.

Family II · Interoperability Layer

Architecture-Universal Operational-State Interface

Target-relative coordinates + portable operators

Each target model exposes its own stable operational coordinates from unlabeled internal state. Representational ambiguity is resolved, a portable read or control operator is applied, and the requested action is lifted back into the target's native coordinates—without requiring identical hidden dimensions, tokenizers or architectures.

Technical differentiation: target-relative normalization, ambiguity stabilization, portable operators, target-native action and post-action verification.

Family III · Governed Autonomy

Predictive Control & Verified Self-Improvement

Forecast → bounded action → independent judgment

The system commits a predicted state transition before acting, compiles a bounded candidate intervention, executes it outside the canonical system, evaluates it against protected criteria and non-equivalent controls, and promotes or rolls back only from externally grounded evidence.

Designed to block: circular self-evaluation, unsafe direct mutation, relabeled trial-and-error and unverified “improvement.”

01Capture
02Normalize
03Sense
04Predict
05Act
06Remeasure
07Verify
08Retain / Roll Back

The strongest accurate priority statement

Proprioceptive AI appears to occupy a potentially valuable early position across an architecture-portable internal-state control plane: pre-output sensing, low-dimensional operational state, inference-time adjustment, streaming control, cross-model normalization, reversible target-native action, protected evaluation and evidence-governed improvement.

This is not a blanket claim to all AI architecture, all activation steering, every latent manifold, Anthropic's exact J-lens, or every autonomous research agent. The defensible value lies in the concrete integrated combinations and their continuation families.

Candidate chronology

27 Jan 2026Candidate umbrella date for a low-dimensional behavioral or operational manifold used for internal monitoring and control.
29 Jan–3 FebHidden-state exploration, inference-time adjustment, per-token sensing, streaming control, fiber methods and state-space extensions.
4 Feb 2026Candidate date for label-free per-model normalization and cross-architecture probe or operator transfer.
Feb–Jul 2026Independent work from multiple laboratories converges on geometry, internal workspaces, causal intervention and evidence-governed autonomy.

Patent notice: dates above are candidate priority positions from the internal portfolio record, not a legal conclusion. Effective dates and enforceable scope require limitation-by-limitation support mapping, continuity verification, prior-art analysis and successful prosecution.

IP Architecture

Three featured crown-jewel families

A parent machine architecture, a portable interface across heterogeneous models, and a predictive evidence-governed improvement loop—separated to preserve distinct prosecution and continuation paths.

Counsel has begun processing 44 patent families. Portfolio update ·
55
Reported Provisional Filings
Hundreds of Claims
Drafted Claim Concepts
Dual-Path Neural Computation with a Proprioceptive Operational-State Pathway
The broad parent architecture: a task-producing pathway plus a functionally distinguishable pathway that senses, predicts, governs and verifies the system while computation is ongoing.
Architecture-Universal Operational-State Interface
Target-relative coordinate construction, ambiguity stabilization, portable read/write operators, target-native lifting and heterogeneous operational-state packets.
Proprioceptive Predictive Control and Verified Self-Improvement
Prediction before action, immutable counterfactual records, bounded candidate compilation, isolated execution, protected evaluation, promotion, rollback and success/failure memory.
Candidate record · Jan 27–Feb 4, 2026
Effective claim dates depend on as-filed support and prosecution
Team
Logan Matthew Napolitano

Logan Matthew Napolitano

Founder & CEO

Father. Husband.

An independent research program that grew into an architecture-portable control layer for sensing, correcting and verifying AI behavior across heterogeneous model families.

Nicholas D. Goodman
Managing Director
Staff Researchers, coders & other staff Hiring

Our staff is currently in the low double digits. Researchers, coders, and other staff are being evaluated and hired as the team grows.

Additional consultants and founding members contribute to the company and are not individually listed here.

Careers & opportunities →
Michael Napolitano
Legal Counsel
Stealth Private Placement Bank Exploring potential seven-figure opportunities Under contract

Following several months of discussions, Proprioceptive AI recently entered into a contractual engagement with a multibillion-dollar private placement bank to explore potential seven-figure opportunities.

Engagement inquiries →

Legal inquiries: legal@proprioceptiveai.com · Contact Legal →

Model-Assisted Critiques

Read the Full Critiques from Google AI Mode, Grok, ChatGPT & Claude

We supplied research materials and asked each model to critique the work. These are model-generated analyses, not endorsements or independent validation; the full conversations are linked so readers can inspect the context.

Claude
Anthropic
Read the full conversation
Claude's complete model-generated critique, hosted by Anthropic.
Google AI Mode
Google Search
Read the full conversation
Google AI Mode's shared conversation about Proprioceptive AI, hosted by Google.
Grok
xAI
Read the full conversation
Grok's complete model-generated critique, hosted on X.
ChatGPT
OpenAI
Read the full conversation
ChatGPT's complete model-generated critique, hosted by OpenAI.
FAQ

Questions, answered

The Technology

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 learned monitors (probes) that read selected hidden states and estimate behavior-linked patterns — hedging, repetition, sycophancy and shallow reasoning — before output. The deployed system gains an external operational-state monitor, not consciousness or human-like self-awareness.

Cognitive probes are small learned monitors attached to selected hidden states in supported models. They estimate patterns associated with behaviors such as hedging, hallucination, shallow reasoning or repetition before output.

RLHF (Reinforcement Learning from Human Feedback) modifies model weights and can require substantial labeling and retraining. It may also introduce capability trade-offs. Proprioceptor can leave the base model frozen during monitoring, read selected hidden states and apply targeted interventions at decode time.

Probe scores can also supplement or partially automate reward-model signals rather than replacing all human judgment. We refer to this as Probe-Guided Reward Modeling; related implementations are described in the company's provisional filings.

In supported deployments, detection runs before output, so a response can be held before display and an agent's tool call can remain pending until internal evidence checks pass. The gate can approve, deny or route the proposed action according to deployment policy.

The reported separation range compares measured behavioral separation with a defined null or chance baseline under the documented internal protocol. It indicates a strong signal in those experiments, but it is not a universal multiplier of model quality and should be read with the associated controls and test conditions.

Architecture & Priority

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.

Product & Company

The framework is architecture-portable in validated settings across 40 architectures from 410M to 405B parameters, including transformers, Mamba state-space models, mixture-of-experts and attention-free designs such as RWKV. Deployment requires supported access to hidden states during inference. In internal experiments, training a probe for a new behavior can take about 20 minutes on a consumer GPU, depending on the model and dataset.

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

As of 6 September 2026, counsel has begun processing 44 patent families. The internal portfolio also reports 55 provisional filings and hundreds of drafted claim concepts, with 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 filings and cross-architecture internal results. First commercial deployments are targeted for Q3–Q4 2026 in clinical AI.

Research & Media

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.

The convergence map also covers Manifold Steering and 11 additional research and deployment references from April through September 2026, including low-dimensional feedback control, HalluTracer, activation monitoring and governed alignment research.

The closest connections concern low-dimensional activation control, geometric state sensing and intervention before output. Other references concern monitoring, governed improvement and efficient computation. These are our technical comparisons, not claims that every source uses our low-rank methods or replicates the complete Proprioceptor system.

Deep Dive on our Architecture & Results

Full Video
Click to watch Podcast ▸

Build Safer AI Systems

Start with a scoped pilot — on a shared open model, or on your model inside your environment.

Request a Pilot → (479) 343-8795 Research Archive →