AI Transparency

Cognistase is not a consumer AI chatbot. It's not a general-purpose language model repurposed for children. It's a purpose-built clinical system where every output can be traced, calculations are done by dedicated software, and every educational claim cites a verifiable source.

Our AI approach

Cognistase uses dedicated clinical calculation software alongside specialized AI systems, each isolated and built for a single task. We don't use general-purpose AI agents that can make things up.

Clinical entity recognition

An AI that reads medical and educational documents and extracts relevant clinical information: test scores, diagnoses, observations. Think of it as a highly trained reader, not a conversationalist.

Document generation

An AI that combines clinical data, legal requirements, and educational guidelines into structured advocacy documents. It works like a legal brief writer who must cite every claim.

Task decomposition

An AI designed to help children break complex tasks into smaller steps. Built with a child-safety framework that prevents any harmful content.

Semantic matching

An AI that understands the meaning behind children's interests to find compatible peers. It works like a librarian who connects readers with similar tastes.

Clinical calculation

This is not AI at all. It's traditional, dedicated calculation software. Two plus two always equals four. No exceptions. No approximations.

How we choose our AI models

• Open-source: we only use models whose code and training methods are publicly available for review. No black boxes.
• Self-hosted: we download models and run them on our own servers. Your data never leaves our infrastructure.
• Built for the task: each model is selected or fine-tuned for its specific clinical or educational domain.
• Independently tested: we test each model against known clinical datasets to verify accuracy for our use case.
• Replaceable: no single model is irreplaceable. If a better open-source model comes along, we can switch without affecting your data.

Source attribution

Every AI-generated recommendation includes references to the sources it draws from. These are real citations to peer-reviewed papers, laws, and clinical guidelines, not fabricated references. If the AI can't find a verifiable source for a claim, it doesn't make the claim. This is enforced at the system level, not as a suggestion that can be overridden.

Confidence indicators

Each result includes a confidence level: high, medium, or low. Low-confidence outputs are clearly marked and explained. We'd rather tell you we're not sure than pretend we are. A high-confidence result means multiple verified sources support the conclusion. A low-confidence result means the evidence is limited, and the output should be treated as a starting point for further investigation.

How the reasoning works

The system explains which factors went into a particular analysis or recommendation. You can look at the reasoning chain, see which data points were considered, and understand why certain conclusions were drawn. This isn't a dumbed-down summary. It's the actual reasoning path the system followed.

What our AI does not do

• It does not diagnose your child. Only licensed professionals can diagnose.
• It does not make decisions about your child's education. Those decisions are yours and your school's.
• It does not do arithmetic. All scores use dedicated clinical software.
• It does not learn from your child's data. We don't retrain models on user data.
• It does not act on its own. All AI-generated documents are drafts for your review.
• It does not talk to other AI services. No API calls to third-party providers.