Geometric Hierarchies Organizing State Transitions:
Linearity Informed Neural Encoding

GHOSTLINE

Real-time behavioral monitoring for transformer inference.

Open Demo Contact

94.8% state classification accuracy

F1 0.94 fabrication detection

4 architecture families validated

Patent Pending 3 US provisionals filed

The instrument

See what the model is doing while it's still doing it

GhostLine spectrograph — Entropy, velocity, confidence, and attention signals as time-series across the full generation. Click to expand

What holds up

Validated results, honest boundaries

GhostLine hypothesis panel — Define signal rules, test against trajectory data, see accuracy and confusion matrix. Click to expand

Sharp

State separation across 4 architecture families
Fabrication detection (F1 0.94, held-out stress test)
Replayable signal inspection (58 signals, per-token)
Causal geometric intervention at 3B

Maturing

70B+ scale validation
Full parameter space coverage (temp, top-p)
Adversarial and multi-turn corpus

Applications

Where this goes

Compliance & Audit

EU AI Act Articles 9 and 12 require behavioral monitoring and logging for high-risk AI systems. GhostLine provides per-token behavioral state classification and audit trails — continuous monitoring that regulators can inspect directly.

Inference Optimization

Pre-generation prediction identifies prompts likely to hallucinate before any tokens are generated. Route, re-prompt, or gate chain-of-thought in ~15ms — before wasting compute on bad output.

Defense & Security

Geometric behavioral monitoring is independent of output content. An adversary can craft text that passes content filters but cannot easily control the geometry of the model's computation. Prompt injection screening at the activation level.

Model Development

Geometric profiling of training checkpoints reveals whether fine-tuning actually improves behavioral separation. Prompt corpus certification establishes geometrically bounded quality rules per model.

Under the hood

How it's wired together

Interactive architecture

Drag, zoom, and hover to explore how the instrument is structured.

Briefing pack Technical atlas Full lineage hub

Behind the project

Collin Civish

Research prototype

Solo inventor. No formal ML background. Built a real-time geometric monitoring system for transformer inference, validated it across 4 architecture families at 2 parameter scales, and filed 3 US provisional patents — in 48 days.

Seeking resources to take it from prototype to production.

Open Demo Contact

Intellectual property

P1 — Instrument	US 63/948,867	Dec 2025
P2 — Intelligence	US 63/975,787	Feb 2026
P3 — Interface	US 63/982,900	Feb 2026

P3: 27 claims, 22 figures, 90-page spec.

At a glance

6models

4architectures

2scales

48days to 3 filings

Geometric Hierarchies Organizing State Transitions: Linearity Informed Neural Encoding