Proprioceptive AI — Meta Validation, Benchmarks & Investment

⬡ META FAIR · FEBRUARY 5, 2026 · 19K+ VIEWS

Convergent Discovery:
Meta FAIR Validates Our Thesis

FAIR at Meta published "Learning to Reason in 13 Parameters" — proving you can teach an 8B-parameter model to reason by training just 13 parameters. 26 bytes. We proved you can read behavioral failure from just 16 dimensions. Same geometric truth. Two independent groups. Opposite sides of the problem.

We filed the patents first. Every research group now publishing in this space — Meta, Apollo Research, university labs — is working within territory our 55 filed patents cover. Their papers validate our claims. Our patents protect the implementation.

META FAIR — WRITE SIDE

arXiv:2602.04118

13 parameters

To inject reasoning capability into an 8B model. RL succeeds where SFT fails — the cognitive update lives in a tiny subspace.

⟷ SAME TRUTH

PROPRIOCEPTIVE AI — READ SIDE

55 Patents Filed · 5 Architectures · Shipping

16 dimensions

To detect behavioral failure — sycophancy, hallucination, hedging, shallow reasoning — at every token, in real time. 1,376× separation.

PRODUCTION BENCHMARK

Not Just Detection — Intervention That Works

Industry-standard IFEval benchmark results with 3-condition attribution testing. We don't just find the problem — we fix it in real time with minimal intervention.

IFEval Improvement

+77.8%

Instruction-following benchmark

⚡

Surgical Precision

3.1%

Steering rate — only 37 interventions per 1,000 tokens. The model runs freely 96.9% of the time. We intervene only at the exact moments of behavioral failure.

🎯

Probe Attribution

86%

Of the improvement comes directly from our probes. Verified through 3-condition ablation: probes only, sampling only, temperature only.

🔬

3-Condition Attribution

Independent conditions tested: probes (86%), sampling strategy (11%), temperature adjustment (3%). Not just "it works" — exactly why it works.

Attribution Breakdown

Probes

86%

Sampling

11%

Temperature

CROSS-ARCHITECTURE PROOF

5 Architectures. Same Signal.

From transformers to state-space models, the behavioral subspace is universal. Published literature reports 2–5× separation. We measure 100–1,376×.

ALIBABA
Qwen-3B
3 billion params
TRANSFORMER
1,376×
Hedging detection

MISTRAL AI

Mistral-7B

7 billion params

TRANSFORMER

999×

Reasoning depth

TII UAE

Falcon-Mamba-7B

7 billion params

STATE-SPACE

999×

Cross-architecture

META

LLaMA-8B

8 billion params

TRANSFORMER

272×

Verbosity detection

NEW

ALIBABA

Qwen-7B

7 billion params

TRANSFORMER

999×

Multi-behavior

INDEPENDENT VALIDATION

Multiple Groups. Same Conclusion.

We're not the only ones finding this — we're just 100–500× ahead. Every major research group publishing in behavioral detection confirms the geometric truth our work is built on.

PROPRIOCEPTIVE AI · 2024–2026

Fiber Probe Behavioral Detection

16-dimensional probes detect 8 behavioral classes across 5 architectures. Per-token, real-time, with EMA spike detection and intervention.

1,376× peak separation

META FAIR · FEB 2026

Learning to Reason in 13 Parameters

TinyLoRA proves reasoning capability lives in ultra-low-dimensional subspace. RL writes to the same geometry our probes read from.

13 parameters sufficient

APOLLO RESEARCH · 2025

Deception Detection via Probes

Linear probes detect model deception from internal representations. High AUROC confirms behavioral signals are linearly separable.

0.96–0.999 AUROC

ICML 2025

Safety Features as Linear Subspaces

Safety-relevant features form linear subspaces within transformer activations — the same geometric structure our probes exploit.

2–5× separation

EMNLP 2025

Factuality Probes

Linear probes trained on hidden states distinguish factual from non-factual outputs — validating our per-token detection approach.

2–5× separation

They get 2–5× separation. We get 100–1,376×.

Same underlying science. Same direction. We're further ahead because we've been building the infrastructure — fiber projections, per-token labeling, EMA detection, cross-architecture transfer — since before anyone else was looking.

COMPANY TRACTION

From Research to Revenue

Sales force deployed. Engineering team hired. Enterprise partnerships active. Probe weights live on HuggingFace. This isn't a research project — it's a company.

🗺️

US Territory Directors
Deployed Nationwide

⚙️

Engineering Team
Hired for Production

🚀

Batch #2
Launching Now

🤝

Live

Enterprise Partnerships
& License Agreements

🤗

Probe Weights
Live on HuggingFace

TARGET VERTICALS

🏥 Healthcare

💰 Finance

⚖️ Legal

🏛️ Government

🛡️ Defense

MANDATORY DEMAND

Regulation Is Creating Non-Optional Buyers

These aren't future possibilities. Regulatory frameworks are creating mandatory demand for behavioral monitoring — exactly what we've built.

🇪🇺 EU AI ACT

European AI Act

Mandates behavioral monitoring and risk assessment for high-risk AI systems. Organizations deploying AI in healthcare, finance, legal, and government must demonstrate continuous behavioral oversight.

EFFECTIVE NOW

🏥 FDA GUIDELINES

AI Medical Device Framework

FDA drafting requirements for AI-powered medical devices including real-time behavioral monitoring, hallucination detection, and output validation — core capabilities of our probe system.

DRAFTING REQUIREMENTS

📊 SEC DISCLOSURE

AI Disclosure Requirements

SEC implementing AI disclosure requirements for financial services. Firms using AI for trading, risk assessment, or customer interaction must demonstrate behavioral oversight and failure detection.

IMPLEMENTING NOW

Every regulated industry deploying AI needs what we've built. The question isn't if — it's who gets there first.

FREQUENTLY ASKED

Questions

For investors, partners, and anyone interested in the technology.

Meta FAIR's "Learning to Reason in 13 Parameters" (Feb 2026) proved you can write reasoning capability into a model using just 13 trained parameters. We independently proved you can read behavioral failure using just 16 dimensions. This is convergent discovery — two groups arriving at the same geometric truth from opposite sides. Meta proves the cognitive geometry exists on the write side. Our probes prove it's detectable on the read side. The combined implication: the cognitive structure of LLMs lives in absurdly low-dimensional subspaces, and our technology is uniquely positioned to monitor and correct it in real time.

We're currently raising on Wefunder. Anyone can invest — accredited or not. Visit our campaign page to review our materials and join the raise. For questions, call (323) 393-0811 or email logan@proprioceptiveai.com.

Five verticals where AI behavioral failure carries the highest risk: Healthcare (AI diagnostics, clinical decision support), Finance (trading, risk assessment, customer-facing AI), Legal (contract analysis, legal research AI), Government (public-facing AI, policy analysis), and Defense (autonomous systems, intelligence analysis). Each faces regulatory mandates creating mandatory demand.

Most AI safety works at the training level (RLHF, constitutional AI) or output level (content filters). We work at the inference level — reading hidden states in real time while the model generates each token. We detect behavioral failure before it reaches the user. Our 1,376× separation is 100–500× ahead of published results (2–5×), and our approach works across 5 architectures including state-space models. Multiple groups — Meta FAIR, Apollo Research, ICML/EMNLP researchers — have validated the underlying science.

Separation ratio measures how distinctly our probes can tell behavioral classes apart in the model's hidden state geometry. A ratio of 1,376× means the distance between clusters (e.g. "hedging" vs "direct response") is 1,376 times larger than the variation within each cluster. Academic published work achieves 2–5×. Higher separation means more reliable detection with fewer false positives — critical for enterprise deployment.

Meta Just Proved
Our Thesis.

Convergent Discovery:
Meta FAIR Validates Our Thesis

Attribution Breakdown

Join the Raise

Meta Just Proved Our Thesis.

Convergent Discovery: Meta FAIR Validates Our Thesis

Attribution Breakdown

Join the Raise

Meta Just Proved
Our Thesis.

Convergent Discovery:
Meta FAIR Validates Our Thesis