Products & Platforms Seven sections · One research core

One research core.
Seven ways forward.

From the published research behind Proprioceptor to enterprise agents, security audit, interpretability tooling, open-source resources, hands-on partnerships and the hosted steering setup kit. Choose a section to explore.

Explore the platforms, read the research, or start a conversation. Product capabilities and availability vary by platform.

THE ORIGINAL RECORD

Research you can inspect.

Published papers and an archived portfolio snapshot, straight from the source.

PORTFOLIO SNAPSHOT / 27 AUG 2026Original screenshot panel showing Total Records: 303Inspect original image ↗
303 portfolio recordsHistorical record count, not an issued-patent count.
Actual first page of From Internal Prediction to Measured Control by Logan Matthew NapolitanoRead original PDF ↗
From Internal Prediction to Measured ControlPublished PDF · 25 September 2026
Actual first page of The Proprioceptive Transformer version 2.5 by Logan Matthew NapolitanoRead original PDF ↗
The Proprioceptive TransformerResearch candidate v2.5 · 29 July 2026

Decorative architecture reference: NIST Advanced Measurement Laboratory · Gail Porter/NIST. Shown at 10% opacity; not a Proprioceptive AI facility.

01 · Research

The Research Moat

Every platform on this page rests on the same foundation: original research in internal-state sensing, bounded control and verification. It is published, archived and backed by an active patent program.

44
Patent Families with Counsel
40
Architectures Validated
1,376×
Peak Separation vs. Null
23
Public Research Deposits
Read the publicationsproprioceptiveai.com/publications

Patent families are being processed by counsel (as of 6 September 2026); applications are not issued patents. Public deposits span February to May 2026. Reported figures reflect internal benchmark conditions and should be read with the associated methods, controls and nulls.

CISO CONSOLEDEVELOPMENT CAPTURE ↗Actual Proprioceptive AI CISO console, showing the backend-ready indicator, CISO navigation and observed signed Repair and Preserve outcomes beside fail-closed containment

Actual development interface. Open the image to inspect the console at full resolution.

View the control-proof workflow ↗
02 · Security & Audit · Working demo

CISO & Auditing Platform

Give security and audit teams a clear view of model decisions: inspect the evidence, authorize bounded changes and verify the result before release.

  • Pre-output risk signals

    Internal-state flags for hallucination, deception, manipulation and sycophancy, raised before output reaches a user or a tool.

  • Evidence-gated outcomes

    Every case ends as REPAIR, REFUTED or ABSTAINED. When the evidence is insufficient, the system makes no edits and records that it abstained.

  • Signed, replayable audit trail

    Outcomes are released in signed, monotonic order, and the audit trail is signed with dual-key Ed25519 and designed for third-party replay.

  • Runs on-prem

    Deployed as a controlled-access on-prem appliance, so model data stays inside your environment.

HUGGING FACE / LOGANRESEARCHResearch you can explore.MODELS / DATASETS / MODEL CARDS
03 · Open Source · Hugging Face

Open the work.
Explore the models.

Find Logan Matthew Napolitano’s public model repositories and datasets on Hugging Face. Start with a model card, inspect the accompanying files and follow the published research for context.

Availability, licensing and usage terms are specific to each repository. Check its model card before use.

04 · Enterprise agents · In development

Cygnus Corporate Agent Harness

A harness for putting corporate AI agents to work under control. Cygnus is being built on Proprioceptor's runtime gate: before an agent's action takes effect, it can be held while internal evidence checks run, then approved, denied or routed to a person under your policy.

  • Control before external effect

    Proposed actions and tool calls can be held until internal evidence checks pass. Your policy decides whether each one is approved, denied or routed to a person.

  • Internal-state checks

    Readers on the model's internal activations score hallucination, deception, manipulation and sycophancy while it is still generating, before an answer or action is released.

  • Gated corrections

    Corrections are applied automatically only after causal, preservation, regression and rollback checks pass. Otherwise a person approves them.

  • Signed record, on-prem

    Decisions are written to an audit trail signed with dual-key Ed25519, and the system is designed to run on-prem so model data stays inside your environment.

05 · Interpretability

Hidden State & Interpretability Platform(s) for Mechanistic Control

Go beyond observing a model's outputs. Read its internal state, apply bounded, reversible control in its own coordinates, and verify every change, with one framework that carries across model families. It includes Proprioceptor, our on-prem sensing, control and verification system.

01

Read

Trained behavioral probes read selected hidden states and estimate behavior-linked patterns before output. Internal held-out evaluations report AUC 0.96–0.999 under defined test conditions.

02

Steer

Targeted, low-dimensional interventions are applied at decode time while the base model's weights stay frozen. No full retraining from scratch.

03

Verify

Each intervention is remeasured and checked against protected criteria and a capability probe. Verified changes are retained; the rest are rolled back.

The control loop

01Capture
02Normalize
03Sense
04Predict
05Act
06Remeasure
07Verify
08Retain / Roll Back
Transformers Mamba state-space models RWKV Mixture-of-experts 40 validated architectures 410M–405B parameters

Deployment requires supported access to hidden states during inference. Reported figures reflect internal benchmark conditions.

Architectural photograph of a contemporary glass-walled conference room with executive chairs and a city view
CONSULTING & PARTNERSHIPSResearch.
Pilots.
Partnerships.
Ideas, evidence and enterprise.
Architectural photography · Aleksandrs Karevs / Unsplash · Representative setting
06 · Partnerships

Consulting & Partnership Platform

Proprioceptive AI is currently in discussions and working across multiple pilots and partnerships. We work directly with teams on scoped evaluations, research collaboration and advisory engagements, with each relationship shaped around its goals and stage of development.

Scoped pilots

We are working across multiple pilot engagements and discussing new evaluations, using shared open models or a team’s own model inside its environment.

See pilots

Research partnerships

We are in ongoing discussions and working on partnerships around interpretability, internal-state monitoring and mechanistic control research.

Discuss a partnership

Advisory & consulting

Hands-on help evaluating, monitoring and governing the AI systems you already run.

Book a conversation

Free 30-minute conversation with Logan and the team. No preparation needed.

07 · Hosted Steering · Client Setup

Connect to the
steering service.

Run our connection client on your computer to use the hosted models and request experimental steering. Inference runs on Proprioceptive AI’s servers, so you do not need model weights or a GPU.

Python & Node.js clients · Windows / macOS / Linux · 13 November 2026

INCLUDED IN THE DOWNLOAD

Your connection to
the hosted preview.

  • Python and Node.js clients
  • Five documented model routes
  • Steering-enabled examples
  • API reference, receipts & file checksums

Prompts are sent to our hosted service. This kit connects to the listed hosted models; it does not install steering into models on your computer.

  1. 01

    Sign up & download

    Register with your email, download and extract the kit. You can also register through the signup API. Open the proprioceptive-ai-connect folder containing README.md, python and node.

  2. 02

    Choose a client

    Use a current supported release of Python or Node.js. The kit requires Python 3.10+ or Node.js 20+, with no extra packages.

  3. 03

    Run from that folder

    Open Terminal or PowerShell in the extracted folder. Copy the command below, run it and read the returned answer and steering result.

Build your setup command

Experimental steering requested with --loop
RUN INSIDE proprioceptive-ai-connect

Choose your options, then copy the command. Replace the example question with your own prompt.

Read what the service actually did.

The client prints JSON. Read answer for the response, inspect x_loop for the processing result and keep receipt if you need help.

Steering is experimental.

Requesting steering does not guarantee a change or a better answer. The service may keep the baseline answer, abstain or report a dependency failure. A successful connection alone does not establish an improvement.

Sign up through the API

Email is the only required field. Name and company are optional. No password, existing API key or email-verification step is required.

POST https://proprioceptiveai.com/steering-signup.php
Content-Type: application/json

curl 'https://proprioceptiveai.com/steering-signup.php' \
  -H 'Content-Type: application/json' \
  --data '{"email":"you@example.com","name":"Your name"}'

A successful signup returns registered: true, your user_reference and a download_url valid for one hour. The link works without keeping browser cookies. This is registration for the setup kit; it does not introduce an inference API key.

Website and API signup details are saved privately and emailed to logan@proprioceptiveai.com. Use the complete guide for request fields and response details.

Full API signup instructions ↗
Connection details & troubleshooting

Hosted endpoint: POST https://api.proprioceptiveai.com/v1/vault/chat. The documented public preview accepted requests without an API key when this kit was prepared. Availability and response quality may vary.

One prompt per request: the route forwards only the last message. It does not retain conversation context. The documented route does not support streaming, tools or images.

If it times out: completion is unknown. The clients wait up to 200 seconds and do not retry automatically. Preserve the error or receipt before making another deliberate request.

Building a browser app? call the service through your own server. Direct browser requests currently lack CORS support. The included README and API reference describe the request and response fields.

Verify the ZIP: SHA256 f94a9eef41a6b7df7becddefa5533c105507211e3a64220db1cd689bf9588c97

Open the complete connection guide ↗

Build Safer AI Systems

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