AI System Fact Sheet
How Trooth's AI features work, so you can judge how far to rely on them. Updated as the systems change.
What the AI does
Two features use AI: Kyrie, a conversational assistant, and questionnaire drafting, which proposes answers to security and vendor questionnaires. Both draft and explain; a person decides.
Foundation models
Both run open-weight models, not trained by Trooth, on Cloudflare Workers AI inside the product's network, under terms excluding inputs and outputs from training; Cloudflare is a listed sub-processor. These commitments hold if the model changes.
Feature by feature
Each model and its inputs, checked against running code on September 26, 2026. Cloudflare picks the location; no feature pins a region or falls back to another provider.
| Feature | Model | What the model sees | Stored by Trooth |
|---|---|---|---|
| Kyrie, the assistant on trooth.co | Workers AI: @cf/baai/bge-base-en-v1.5 to find passages, @cf/meta/llama-3.3-70b-instruct-fp8-fast to write the answer | Your question, curated knowledge-base passages and, when you are signed in, your company's posture summary | A count of requests for rate limiting; not the question |
| Questionnaire drafting | Workers AI: @cf/meta/llama-3.1-8b-instruct | The question being drafted and your own workspace's witnessed evidence | The draft you keep, in your workspace |
| trooth_ask, the Model Context Protocol (MCP) knowledge tool | None: a fixed, curated knowledge base and a capability registry, matched by keywords | Not applicable | That a question was asked; not its text |
Anthropic is on the sub-processor list for internal use by Trooth's founder. No customer-facing feature sends customer data to it.
What data the models see
A model sees only what its task needs: your prompt, witnessed evidence and posture data, and framework references. It does not receive credentials or integration secrets.
Trooth does not use your data to train AI models, and prompts are not contributed to provider training. The AI Use Policy states the full stance.
Limitations
- Output reflects the evidence available at the time and can be out of date.
- Models reason over language, not a ground-truth audit, and can miss nuance in your environment.
- Drafts start security and compliance work; they do not replace professional judgment.
Hallucination disclosure
Like all large language models, these can produce confident, inaccurate or fabricated output, so a person should review it before you rely on or submit it. Trooth labels AI content and keeps a person in control of real effects.