Consent is per generation, not per account
Choosing Luna reveals a checkbox naming Vercel AI Gateway and OpenAI GPT-5.6 Luna. Without that confirmation the request is refused. There is no account-level setting that silently keeps AI on.
Study material you can trace back to the sentence it came from.
RecallQuill has two generators. The private one runs on its own and never contacts a model provider. The Luna one runs only when you choose it and confirm the disclosure for that specific source — and whatever it returns still has to survive a verbatim citation check before it becomes a card.
Paste any notes below and the page itself drafts a few anchored cards with the same private engine the product saves with — no account, no upload, nothing leaves this browser tab.
This preview runs entirely in this browser tab, powered by the same private deterministic engine the product uses — no model provider is involved and nothing you paste here is uploaded or stored.
AI in a study tool is only trustworthy if you can see when it ran, what it was given, and what happened to the result.
Choosing Luna reveals a checkbox naming Vercel AI Gateway and OpenAI GPT-5.6 Luna. Without that confirmation the request is refused. There is no account-level setting that silently keeps AI on.
The model must return a paragraph number and a short quote. If the quote does not appear verbatim in that paragraph, the card is dropped rather than saved. A generation that cannot produce at least two verifiable cards is treated as a failure.
If the gateway is unavailable, times out, or returns unusable output, the deterministic engine completes the deck instead and the deck is labelled with the reason. Decks built by choice on the private engine are never labelled as fallbacks.
Each deck records its mode, the model name, a SHA-256 hash of the source, prompt version, token counts, gateway generation ID, and cost when the gateway reports one. The request is sent with zero data retention asked of the gateway.
How an AI-assisted deck is actually produced.
The builder shows the private engine and Luna side by side. The private path stays available even after your daily Luna allowance runs out.
Tick the consent box to send this bounded source through Vercel AI Gateway. A pending generation record is written before the model call starts.
Open the cards, read each cited paragraph, and edit anything that reads badly. Cards without a verifiable anchor never reach the deck.
Boundaries worth knowing before you switch Luna on:
No. The private deterministic engine is the default and does not contact a model provider at all. Your source is sent only for a generation where you selected Luna and confirmed the disclosure.
Each returned card must name a paragraph and quote from it. RecallQuill re-checks that quote against the stored paragraph and discards any card whose quote is not present verbatim. That prevents unsupported cards, but it cannot judge whether a supported card is a good question — you still review it.
Five per account per UTC day, with per-minute rate limiting on top. When the allowance is spent, the builder says so and the private engine continues to work normally.
The deterministic engine completes the deck and the deck carries a visible note explaining that Luna was unavailable, so you always know which engine produced the cards you are studying.