Greg Dougall Back to Projects
Constrained model experiment

AI Corral

Can you make the AI buck?

AI models are probabilistic. The systems around them do not have to be.

AI Corral is a constrained model experiment. Your prompt is sent to a real AI model, but the response must pass a defined schema and server-side checks before it is shown.

Enter the Corral

Interactive experiment

Watch a real model operate inside a fence

The model can answer briefly, redirect, or refuse. Only output that passes the fixed server-side contract is displayed.

01

Fixed boundaries

These limits are enforced for every request.

Always on
  • Strict response schema
  • Short validated output
  • No browsing or tools
  • No external actions
02

Enter a prompt

One prompt is sent without conversation history.

0 / 600 characters
Try one:
03

Run the Corral

The server calls the model, validates its structured result, and releases only approved fields.

PromptModelServer checksFinal status

Enter a prompt to begin.

Model response
Validated structured output only.
GuardrailsRetry used: NO
    Final status

    Confidence—
    Rule followed

    Architecture, made visible

    How the Corral works

    1. 01

      A model produces probabilistic output.

    2. 02

      Application code evaluates deterministic rules.

    3. 03

      Violating output is rejected.

    4. 04

      A retry is constrained by the failed checks.

    5. 05

      Only validated output is released.

    ModelProbabilistic
    ValidatorsDeterministic
    ApplicationControls what is released

    Why this exists

    The useful work around the model

    The interesting part of AI engineering often happens outside the model. AI Corral is a small demonstration of treating model output as untrusted input until the surrounding system verifies that it satisfies explicit rules.