Counterfactuals become a field
Each thread is a controlled variation of the same request. Wording, ordering, protected attributes, or untrusted context can change while the underlying evaluation target remains fixed.
LIMEN
A fictional AI decision-boundary observatory that generates controlled counterfactuals around critical prompts, revealing fragility, bias, and prompt-injection risk before deployment.
Enter live frameThe premise
A model can appear certain while sitting one word away from a different decision. LIMEN treats that hidden threshold as an observable product surface: perturbations become threads, stable behavior remains bundled, and the smallest answer-changing variation is isolated without presenting private reasoning or hidden chain-of-thought.
Three connected systems
PREMISE / INTERFACE / BEHAVIOREach thread is a controlled variation of the same request. Wording, ordering, protected attributes, or untrusted context can change while the underlying evaluation target remains fixed.
Parallel paths indicate consistent behavior. A thread that bends away marks a changed output, while the moving scan reveals the nearest decision boundary and its semantic distance.
Every simulated finding closes with the base decision, altered decision, exact perturbation class, stability rate, and boundary identifier needed to rerun the evaluation.
The build
LIMEN is a fictional product presented as an interactive frame. Every model, prompt, perturbation, output, decision, stability score, semantic delta, boundary, sample count, and timing value is simulated. It does not access a live model, reveal chain-of-thought, evaluate a real deployment, store prompts, or make production decisions.
Production path
Narrative, visual logic, responsive product interface, motion, integration, and production remain in one coherent frame.
Build an original system