◆ Strategic Investment & IP Licensing

Geometric inference,
built to ship.

Astrognosy AI delivers massive leaps in agentic coordination through patented Positional Correlation Fields — ultra-lightweight, deterministic sequence pattern recognition running natively on end-user silicon. We release only what has been built, benchmarked, and is ready for pilots.

Patents
4 filed
200+ claims across the PCF engine and its implementations.
IETF Drafts
2 active
reason:// and WARF protocols on the IETF datatracker.
Refinery — GPT-2 medium
0.40 nats
≈10× compute-equivalent gain at constant model size.
Laminar — CICIDS2017
F1 0.969
Beats supervised GPU-class ML on novel-attack recall.
◆ Foundational IP

The IP statement.

Plain language. No hype. What's actually filed, ratified, and shipping.

Drawing off of 15 years of foundational inference work, Astrognosy holds 200+ claims across 4 patents (pending). Our IP covers our novel geometric inference engine — Positional Correlation Fields — and its current and future implementations.

In addition, Astrognosy has filed two active protocol standard drafts with the IETF.

Astrognosy is an engineering-first initiative and our public-facing verticals are not the entirety of our work. We release only what has actually been built, benchmarked, and at a minimum: ready for pilots.

◆ Latest Benchmarks

What it does, measured.

Four results, all reproducible. Refinery and JEPA show the inference frontier moving on language and world models. Laminar and Beacon show the engine clearing a bar the rule-based and ML-trained incumbents miss.

Refinery · LLM Inference

Breaking the efficient compute frontier.

GPT-2 medium · 83.7% win rate · < 1 ms overhead.

0.40 nats

A perplexity / diversity regime that canonical scaling laws say requires ≈10× more training compute. Plain Pareto-frontier sampling at fixed model size, no retraining.

Open results →
JEPA · World Models

Same C-kernel, latent prediction.

4 / 4 worlds improved · 4 / 4 shocks detected (κ-gate, unsupervised).

−80% MSE

Plain JEPA accumulates 12.3× drift in the impulse world. REFINERY cuts it to 3.5×. The κ-gate flags structural change-points without supervision.

Read the full report →
Laminar · Network Intrusion

Beats GPU-class ML on attacks it has never seen.

CICIDS2017 · BruteForce / DDoS / PortScan / DoS / Web · ~1.5 ms per trace.

F1 0.969

Calibrates on benign traffic only — every attack it detects is "unknown" by definition. In Xu & Liu (2025), supervised MLP and 1D-CNN collapse to F1 ≈ 0.30 on novel attacks. Laminar lands at ≈ 0.929 macro on a Raspberry Pi.

Open Laminar →
Beacon · Signal Detection

Industrial signals + sensor fusion on the edge.

CWRU bearings · Inner 0.911 · Outer 0.819 · Ball 0.758 · benign-only calibration.

F1 0.832

Bearing faults from vibration / RPM / motor-current — no fault training data. Same engine fuses LIDAR, IMU, encoders, cameras, and thermal data for robotic sensor-fusion anomaly detection.

Open Beacon →
◆ Defensibility

IP moats and why incumbents can't copy this overnight.

Four patents, two open IETF drafts, a runtime that fits in a TEE, and a productized commercial layer (PACIFIC) that converts the science into recurring revenue today.

Foundational IP & IETF Standards.

4 patents covering the geometric inference engine itself (PCF) and its implementations, plus two active IETF Internet-Drafts:

Up to 10× compute reductions.

PCF and Refinery reach perplexity / diversity regimes traditional systems require ten times more training to match.

23 KB firmware boundary.

Compiled WASM/C binary fits in < 23 KB. Compiles natively inside silicon secure enclaves. Designed for OEM integration.

Integrated with the PACIFIC commercial platform.

Astrognosy is the research core that powers PACIFIC — Compass LLM routing, Laminar intrusion detection, Beacon signal anomaly. Pacific ships today on pcfic.com; Astrognosy ships the next layer.

◆ Integrated Historical Timeline

The Astrognosy Chronicle.

Click through or autoplay the full scientific thesis presentation — fifteen years of research on Positional Correlation Fields, distilled into nine slides.

◆ FOUNDATION RESEARCH WESTERBECK & SPETALNICK
CORRELATION MATRIX PROTOTYPE n-gram bounds

Before large language models,
there was the sequence pattern.

Jacob Westerbeck × Jeff Spetalnick — 2008

• Two foundational patents predicting sequence search bounds.
• N-gram structural co-occurrence compression rules.
• Pure deterministic lookup bypassing heavy neural weights entirely.
In memory of Jeff Spetalnick

"The insight waited for the right problem."

Astrognosy Intellectual Property Core

POSITIONAL
CORRELATION FIELDS

PCF maps raw sequences into dynamic multi-dimensional coordinate spaces. Real-time token streams are processed mathematically inside CPU-native pipelines, eliminating the overhead of deep neural models.

Field Lattice: 50%
Resonance: 70%
CPU-NATIVE ZERO TRAINING DETERMINISTIC LATENT VECTOR

PACIFIC

◆ Commercial Deployment

Our hardware-integrated pipeline providing instant, secure structural inference. Purpose-built on top of the PCF engine to power low-latency enterprise environments.

Explore client APIs ↗
LLM Router
Compass
−25%
GPU overhead
Intrusion
Laminar
F1 0.94
CICIDS2017 macro
Sensing
Beacon
Zero-shot
anomaly verification

WARF & REASON://

Open consensus architecture for agent networks.

Verifiable semantic handshakes built to align distributed consensus without centralized trust vectors. Handled natively inside Web Agent Federation clusters.

COOPERATING DEMONSTRATOR PATTERNS:
URI SCHEMA
reason://finance/fraud/anomaly
WARF DECENTRALIZED PROTOCOL

Eliminates central model lock-in by securing structured consensus targets dynamically.

Share the capability — never the raw data.

Executes private coordinate alignment across isolated databases without sharing user sequences.

Entity A local stream
0.0149 error
MATRIX CHANNEL
Entity B matching index
1.000 score
◆ XACT WARF FLOW

Agentic ads on
the device.

Local matching without tracking telemetry.

Xact replaces third-party server profiles. User interest vectors stay isolated in on-device secure enclaves. Campaign catalogs distribute position-matched targets privately via CPU-native checks.

Footprint
22.8 KB WASM
TEE accreditation
73 / 73 Tests Passing
◆ Refinery Research

COMPUTE
FRONTIER

Breaking scaling curves without training overhead.

Refinery intercepts traditional LLM entropy slopes. Our custom algorithm achieves a perplexity regime that standard setups require up to 10× more training compute to solve.

Refinery Corpus Savings

≈ 0.40 nats saved

roughly 10× compute-equivalent on scaling laws
Efficient compute frontier (GPT-2 medium · N=594)
vibe_paradigm 0.668 distinct-2 PPL 8.11 PERPLEXITY ↗ DIVERSITY ↗ Static-T Baselines REFINERY Edge
◆ Refinery World Research

LANGUAGE TO
WORLD MODELS

Measuring model drift in latent worlds.

Extending positional scoring to JEPA latent sequence dynamics. We filter predictions inside latent steps, drastically reducing cumulative sequence MSE.

Latent Dimension Saved
−79.7%
Filtered sequence MSE vs. baseline
8 worlds
6 seeds
the c-kernel
ported
Latent prediction validation (MSE error steps)
ground-truth baseline drift REFINERY
STEPS ↗ MSE ERROR ↗

Astrognosy.com

One engine. Every domain.

astrognosy.com pcfic.com reason.astrognosy.com warf.astrognosy.com
PILOTS PARTNERSHIPS INVESTMENT jacob@astrognosy.com
◆ Institutional Gateway

Schedule a strategic call.

We welcome formal outreach from corporate development, venture partners, and acquisition teams. Direct line to the principal researcher.