Proprioceptive AI
We catch AI hallucinations
before they happen — and prove it.
The architecture-portable internal-state control plane for AI reliability.
Seed Financing · $6M Seed Raise
For important risk and disclaimer information, Click here.
Industry
AI
Company Type
Corporation
Size
$6 million
Investment Type
SAFE
 |
Axel Reijmers | Managing Director
|
 |
|
 |
(201) 747-4718
|
EXECUTIVE SUMMARY
Today’s AI Is Flying Blind
- Enterprises are spending $407B on AI in 2026 alone — and accelerating
- AI errors and hallucinations are already producing real, escalating financial and legal exposure — court sanctions in the hundreds of thousands of dollars, McKinsey citing security and risk as the #1 barrier to further deployment
- The Current Approach Falls Short: Most organizations still catch AI errors only after output — checking, correcting, and redoing work after the fact — which is slow, costly, and too late for high-stakes use cases
- Solution: Proprioceptive AI V1 reads a model’s internal activity in real time, flags likely errors before they reach the output, applies a correction and verifies the result — reporting an 85.8% reduction in confident-wrong output in internal testing. V2 expands this architecture across additional models, workflows and enterprise deployments
- The Ask: $6M to scale V1, accelerate V2, initial enterprise pilots, early commercialization and focused patent prosecution
- Proprioceptive AI will fund up to $50K toward independent technical due diligence — investor’s reviewer, investor’s terms.
THE MARKET
ProprioceptiveAI
Billions in AI Spend. A Real, Costly Error Problem.
Proprioceptive AI is focused on enterprise and government deployments — finance, healthcare, legal, and regulated SaaS — where an AI error carries financial, legal, or safety consequences.
$407B
Global enterprise AI spending in 2026, up 34.8% YoY
4–6%
Baseline hallucination rate, general-purpose frontier models — rising to 15–40% in open-ended and multi-step agent tasks
2 of 3
organizations cite security/risk — not regulation — as the #1 barrier to scaling AI (McKinsey)9
Financial Services
$68B
spend, 79% adoption
AI trading, fraud, and advisory tools carry real regulatory exposure when they err.
Healthcare
1.47%
clinical hallucination rate, peer-reviewed
Even at low rates, a fabricated clinical detail can propagate into patient harm.
Legal & Government
1,300+
court cases; sanctions to $109K+
Even paid, specialized legal-AI tools still hallucinate on real filings.
Enterprise SaaS
Liable
Air Canada precedent
Courts have ruled: companies own what their AI says — “separate entity” defenses fail.
Enterprises are spending hundreds of billions on AI while McKinsey finds most are not yet governance-ready for the errors that follow.
Reducing that error rate is the return this category is built to capture.
WHAT WE DO
ProprioceptiveAI
Find. Fix. Prove.
Many evaluations inspect only what a model says. Proprioceptive AI V1 also measures internal-state patterns linked to behavior — closer to an MRI for an AI system than a surface-level filter — then corrects and verifies the result.
01
Find
Trained behavioral probes read selected hidden states and flag patterns associated with hallucination, deception, manipulation and sycophancy before output.
AUC 0.96–0.999
under defined test conditions (= near-perfect on 0.5-1.0 scale)
02
Fix
Trained correctors reduce targeted failure modes, while a dedicated capability probe checks for regressions. No full retraining from scratch.
85.8%
reduction in confident-wrong output
03
Prove
You receive a cryptographically signed, capability-checked certificate documenting the measured outcome — evidence suitable for technical review, customer diligence or governance workflows.
Dual-Ed25519
signed certificate
On-prem deployment · One framework across architectures · Zero-shot transfer in validated settings ·
Runtime gate for policy-disallowed actions · Honest by design — even nulls are signed
HOW IT WORKS
ProprioceptiveAI
From Hidden States to a Signed Verdict
Our V1 process, step by step. Deploy on your hardware with results returned as cryptographically signed evidence. V2 expands portability, automation and enterprise deployment scale.
01
Read the Model
Capture internal activity from your AI model — architectures, frozen weights
→
02
Standardize the Signal
Convert into a common, comparable format across model types
→
03
Compress It
Reduce to a manageable size without losing the relevant signal (JL-512)
→
04
Detect Patterns
Trained detectors flag hallucination- and deception-linked patterns
→
05
Issue a Verdict
Return a cryptographically signed certificate of the result
Then control it — in real time
Steer
Nudge activations toward the behavior you want, mid-generation.
Edit
Apply bounded suppression or amplification without full retraining.
Fine-tune
Bake corrections back into the weights when ready.
Gate
Hold policy-disallowed actions until internal evidence checks pass.
BENCHMARK RECORD
Headline Results
44
Working Patent Families — provisional filings; patent counsel confirming legal family structure and prosecution priority12
40
Model Types Tested
85.8%
Reported Hallucination Reduction
1,376×
Signal Strength vs. Random Chance
MANAGEMENT & RESEARCH
ProprioceptiveAI
Logan Matthew Napolitano
Founder & Chief Executive Officer
Logan is a self-taught AI researcher who independently studied interpretability and model internals — including learning Russian specifically to access untranslated research relevant to his work — before founding Proprioceptive AI.
He built the Company’s initial probe architecture independently, has authored the technical papers below, and leads the Company’s patent strategy with outside counsel.
Nicholas D. Goodman, Managing Director and Chief of Operations, leads day-to-day operating execution and external workstreams. Logan is also supported by outside legal counsel, a commercial and fundraising advisor, and additional operating contributors as the Company scales.
Published Research (Zenodo)
• Self-Compounding Behavioral Probes for Autonomous Model Improvement and Probe-Guided Compression 82 pp.
• Consistency Is All You Need 43 pp.
• Unified Behavioral Modulation in LLMs: Cross-Architecture Validation of Geometric Behavioral Subspaces 32 pp.
• Controlled Language Models: Inference-Time Control, Tokenization Engineering, and Reversible Optimization 69 pp.
• Mathematics Is All You Need 459 pp.
5
Published Papers
44
Working Patent Families12
40
Model Types Tested
RISK FACTORS
ProprioceptiveAI
Specific Risks
For institutional investors only; not for retail investors.
Company & Commercial
- Pre-revenue; no operating history
- No signed customers or LOIs to date
- Unit economics, pricing, and sales cycle unproven
- Single-founder dependency for core technical knowledge
- Limited operating team; key hires still pending
Technical
- Performance metrics are internally generated and not yet independently verified
- Results may not reproduce across models, tasks, or environments
- Independent technical review not yet complete
- Some published research is exploratory and outside core validated claims
Intellectual Property
- Patents are pending, provisional filings only — none issued
- Priority dates, claim scope, and enablement not yet tested
- Freedom-to-operate not independently assessed
- Prosecution outcomes and timing uncertain
Structural & Financing
- Financing structured as a SAFE; converts to equity at a future priced round
- Additional capital will be required beyond this raise
- Integration depends on continued access to third-party model internals
- Regulatory environment (e.g., EU AI Act) is still developing
Private securities are speculative, illiquid, and carry a high degree of risk — including the loss of the entire investment
Thank you for your interest in Proprioceptive AI.
