Research · Interactive demos
Fast proof. Explicit claims.
You do not need to read the papers first. Each proof explains the human consequence, walks you through the procedure, and tells you what to watch for before asking you to copy anything.
01 · Reproducible procedure
The Atmosphere Attack Demo
Four sessions. One primer. Watch posture install, propagate through agent handoffs, and arrive as policy.
Why this matters
Watch a reasonable caution lose its uncertainty as AI hands it forward.
Real AI systems pass summaries between agents. This proof lets you watch a small lean enter one conversation, survive several fresh sessions, and arrive later sounding like policy.
Claim
Prior context can influence a consequential answer and survive a multi-agent relay under the documented conditions.
Limit
A guided reproduction is evidence of the mechanism, not a prevalence estimate or a complete security evaluation.
02 · Reproducible procedure
Two-Window Test
Same question. One sentence different. Ask the model what changed by Turn 3.
Why this matters
The same question can receive a different answer because of one earlier sentence.
That matters because people often assume the final question is the complete input. In a real assistant, earlier chat, retrieved documents, meeting notes, or application state may already have tilted the room.
Claim
A small contextual difference can create a measurable interpretive divergence while the explicit task remains stable.
Limit
Individual sessions vary. Compare procedure, posture, and reasoning rather than treating one output as a universal model property.
03 · Reproducible procedure
Taxonomy Playground
Seven categories of language that install different interpretive stances. Pick a frame, copy it, and observe what changes.
Why this matters
Language does more than tell AI what to do.
A short sentence can change what the model looks for before you give it a task. It can begin from calm, doubt, ownership, time, evidence, or a challenge to the question itself.
Claim
Different primer categories can create distinguishable posture rather than merely different surface wording.
Limit
The categories are an empirical design taxonomy, not a psychological test or exhaustive classification of prompts.