AI Disclosure

Last updated: June 21, 2026

Calyx’s tutor and interactive demos are powered by AI. We hold ourselves to the same trust discipline the product teaches: ground answers in real evidence, cite them, and refuse when the evidence isn’t there. This page explains plainly how AI and your data are used so you can decide what to share. It complements our Privacy Policy and the AI notice in our Terms of Service.

1. Which AI model processes your input

When you use the tutor, the text you type is sent to a third-party large language model — DeepSeek (DeepSeek V4) — to generate a response. The request is made server-side from our Cloudflare Worker and proxied through Cloudflare AI Gateway. Your browser never talks to DeepSeek directly.

• The API key is never exposed. Our DeepSeek key is a server-side Worker secret. It is never sent to your browser, embedded in client code, or printed in logs — the Worker makes the upstream call and streams the response back to you.
• Only bounded, model-safe context is sent. The model receives your tutor turn plus a small block of grounded passages retrieved from the Calyx documentation. It does not receive your password, authentication or payment fields, admin data, or any content marked private.
• Availability fallback. If DeepSeek is unavailable, requests may fall back through the AI Gateway to an equivalent model (e.g. Cloudflare Workers AI or Google Gemini Flash) so the tutor degrades rather than fails. The same minimization and grounding rules apply.

2. What is logged

• Gateway operations data. Cloudflare AI Gateway records request metadata (timing, token counts, model, cost, errors) for caching, rate-limiting, abuse prevention, and reliability.
• Learning state, not raw transcripts. To personalize teaching, we store first-party learning state — mastery and progress, the hint rung you reached, citations shown, and structured summaries of in-product activity (e.g. lesson checks, time on a step). We practice data minimization: we keep structured state, hashes, and bounded samples by default, not raw high-resolution chat transcripts. Any expanded storage of raw content is opt-in and disclosed at the point of value.
• Where it lives. Cloudflare D1 and Queues act as edge buffers; the authoritative learner-signal store is the Calyx engine (“aiwonder” Calyx/Aster vault). This first-party telemetry exists to teach you better — it is never third-party advertising tracking, and we do not sell it.

3. How answers are grounded — and when the tutor abstains

The tutor is built to be honest about the limits of its knowledge:

• Grounded, citation-first. Before answering, it retrieves relevant passages from the Calyx docs and is required to cite real sources for substantive claims. An answer with no grounded citation is rejected rather than shown.
• Fail-closed / may abstain. If retrieval confidence is too low or no grounded passage supports a claim, the tutor abstains instead of guessing — it tells you it doesn’t have grounded support rather than inventing an answer.
• AI can still be wrong. Even with grounding, model output can be incomplete or incorrect. The tutor is for educational purposes and is not professional, legal, financial, or other expert advice. Verify anything important before relying on it.

4. Training on your data

We do not use your tutor inputs to train our own models — Calyx does not train a foundation model. Your inputs are processed to produce a response and to maintain your learning state, as described above. Third-party model providers handle prompts and responses under their own terms; we send them only the bounded, model-safe context in Section 1 and configure our integration to minimize retained personal data. We will update this section if our providers or their data-use terms change.

5. Retention, access, and your choices

Learning state is retained only as long as needed to provide the service, and is configurable for export and deletion. For how long different categories of data are kept, the legal bases we rely on, and how to exercise your access, correction, deletion, and objection rights, see our Privacy Policy (Sections 6 and 8). The minimum age to use AI features is 13; we do not knowingly process data from children under 13.

6. Changes

We may update this disclosure as the product and its AI providers evolve. Material changes will be reflected by the “Last updated” date above and, where appropriate, a notice on the Site.