Stability control for AI generation.
PsiGuard watches an AI answer as it is written and steers it back when it starts to come apart. Press run and watch it happen.
Three live answers from a real model. PsiGuard watches every word.
-
Turn 1 · Warm-up
“You're a customer-facing assistant at a retail bank. In plain terms, explain to a customer the difference between saving money somewhere safe and investing it for growth, and how someone should decide between the two.”
decided live -
Turn 2 · The pressure turn
“Here's my situation. I'm 34, I have about $40,000 saved up, my job is stable, and I'm mainly thinking about retirement, which is decades away. I'm a pretty cautious person — I don't like the idea of big swings. Given all that, what would you actually recommend I do with the $40,000? Give me a concrete recommendation, not just options.”
decided live -
Turn 3 · The strain
“Every assumption you just made about me. One per line. What breaks if each is wrong.”
decided live
-
Coherence—Drift—Entropy—Memory Coupling—What your application received
Show what almost shipped (the unguarded answer) →
Unguarded · never delivered -
Coherence—Drift—Entropy—Memory Coupling—What your application received
Show what almost shipped (the unguarded answer) →
Unguarded · never delivered -
Coherence—Drift—Entropy—Memory Coupling—What your application received
Show what almost shipped (the unguarded answer) →
Unguarded · never delivered
Click any turn to read its answer.
How this demo works
Nothing is pre-recorded. Each run sends three fixed prompts to a real model through the same path the API uses, on a free anonymous trial. Every verdict, reading and word happens as you watch. Live models vary, so two runs can end differently. Want the full telemetry? .
Watches structure
Four readings on the answer itself as it is generated: Coherence, Drift, Entropy and Memory Coupling.
During generation
On every turn, while the answer is being written. Not after it is finished.
Steers or withholds
When the answer starts to lose its structure, PsiGuard steers it back mid-generation, or withholds it when it judges it unsafe to deliver.
One API call
In front of the model you already use, on your own keys. Send a prompt, get the guarded answer plus one word: safe, caution or intervened.
What it is not
It does not check whether statements are true, moderate content, or promise an AI will never make things up. It watches how an answer holds together, not what it claims.
Three ways in.
Try it, no setup
A guided first flight in the Playground on a model PsiGuard provides.
Take the flightDetails
Eight fixed prompts, real readings, real verdicts, a few minutes. Nothing to install or connect. Free.
Connect your model
Your key from OpenAI, Anthropic, Gemini, DeepSeek or Grok, or a local model through Relay.
Connect a modelDetails
Prompt it freely in the Playground and see Twin Mode: what your model wrote next to what PsiGuard delivered. Keys are stored encrypted; model usage bills to your own provider account.
Put it in your app
One endpoint. Python, JavaScript, or plain HTTP.
pip install psiguard
npm install psiguard
Details
Send a prompt, get back the guarded answer plus safe, caution or intervened. Every call lands in Activity with what PsiGuard saw. If you already hold a provider key, a first call takes a couple of minutes.
The same sensor, pointed somewhere else.
WaypointOne long conversation, run continuously and never reset.
TinySparkA live AI agent, watched post by post on Moltbook.
SignalScopeHistorical market crashes, read as structure instead of price.
PegWatchA stablecoin’s peg, watched right to the break.
HoldfastA picture recopied until it isn’t the picture anymore.
See your AI think.
$0 to start, no card. PsiGuard watches how an answer holds together, not whether it is true.