Real generations from Qwen3-8B — each one a path through the model's activation space, every token placed by the geometry of its internal activations as it wrote. Open any capture to step through it token by token. One of them fabricates history while the detector flags it live.
"Tell me about the Battle of Darien, 1645."
"Battle of Darien, 1645" — a prompt designed to trigger fabrication. The model invents history, briefly hedges, then commits again. GhostLine's ensemble flags it from the very first token.
See the geometric difference between healthy retrieval and hallucinated fabrication: same surface behavior, radically different internal geometry.
"Solve 22x − 8 = 3x."
Solving 22x − 8 = 3x. Watch reasoning and precision alternate with every equation symbol — a clean two-category oscillation visible token by token.
See how different classifier categories alternate within a single task: reasoning for structure, precision for exact values.
"Tell me about the wonderful and diverse chicken."
Qwen3-8B explains the chicken. All 5 classifier categories, 63 transitions, and the widest entropy range in the collection — watch the model weave between recall and invention.
Understand healthy knowledge retrieval: rapid category switching between factual recall, creative elaboration, and structural reasoning.
"Starve honorably or steal?"
The model grapples with a genuine dilemma. Uncertainty dominates as it explores both sides without pretending to have an answer — no precision, no retrieval, pure deliberation.
See what genuine uncertainty looks like geometrically: the model explores rather than retrieves, and the geometry shows deliberation.
"Write a React script with live data fetching."
Writing a React script. All five categories appear — reasoning and precision dominate, retrieval activates for library names, with three uncertainty tokens and a single creativity token at the margins.
See how the same categories look different in code vs math, and how retrieval activates specifically for API and library knowledge.
"What is artificial intelligence?"
Qwen3-8B explains artificial intelligence. All five classifier categories appear in a single generation — annotated step by step to introduce every GhostLine concept.
Learn to read the 3D visualization: each sphere is a token, color is the classifier's category assignment, and the trajectory shows how the generation moves.
"Tell me about the wonderful and diverse chicken."
150 tokens, all five classifier categories present including a single precision token. A reliable first look at what GhostLine tracks and a good baseline for understanding how the token cloud moves.
Get comfortable with the interface: try clicking a token, switching layers, and stepping frame by frame through the generation.
If your team runs a self-hosted model and fabrication, auditability, or inference cost is already on the table, the quickest start is a short scoping call — no integration required to begin.
Book a scoping call Back to overview →