AI study help for the Digital SAT — a mock debrief and highlight-to-explain

Select a phrase — up to 400 characters, a phrase rather than a page — and one sentence comes back explaining it. It is mounted in lessons, in drills, and in the review screen after a sitting is submitted.

It explains; it does not solve. If the thing you highlighted is itself a question, it will tell you what the question is asking rather than what the answer is. That is not a disclaimer, it is in the instructions the model is given, and it is the difference between a study aid and an answer key.

After a mock, the same model drafts the wording of your results debrief from a summary of numbers our software has already calculated. It writes placeholders, not figures; your figures are filled in from your own results, and a draft that states a number of its own is rejected.

The model is DeepSeek V4 Flash, chosen because it is fast. The whole value of a gloss is that it arrives while you are still looking at the phrase; a slower, better sentence would be worse, because you would have moved on before it landed.

Every request goes through OpenRouter and carries a zero-data-retention requirement, so only hosts that have committed not to store what we send, or train on it, can receive it. DeepSeek's own service is excluded by name. If no qualifying host is available, the request is not sent anywhere else: you get our standard wording.

A request never carries your name, email address or account. The explanation request contains the phrase you chose; the debrief request contains numbers and skill names.

The selection is sent, an answer comes back, and both are dropped. There is no table recording what you highlighted, no event carrying the text, and nothing echoed back in an error message.

The budget counter counts calls, not content. A version of this feature that logged what students found confusing would be genuinely more useful to us to have, and it is not what the privacy policy says, so it is not what was built.

The exam runner does not import this feature. That is the mechanism — not a flag, not a permission check that a future change could get wrong.

Three separate published commitments require it. The honour code says no notes, no search, no help from anyone else, and specifically not looking things up mid-exam. The correctness invariant says the key is withheld until submission, and a model answering questions about a live item is a correctness oracle by inference even when no key field is in the payload. And the privacy policy says we never send the text you highlighted anywhere — so the one place that is allowed is where you asked for it, outside a ranked sitting, used once and discarded.

Twenty explanations a day and sixty a week, and six debriefs a day and twenty a week, per account. The budget is consumed before the request leaves, and if the counter cannot be reached the call is refused rather than let through.

That direction is deliberate and it is the opposite of how the rest of the platform degrades: a failure to check the budget fails closed here, because the cost of letting an unmetered feature through is unbounded and the cost of a student not getting one gloss is one gloss.

The cap is also what stops this becoming a paywall later. A runtime AI feature with no ceiling has exactly two futures — it gets restricted to paying users, or it gets removed — and both of those are worse than a number stated up front.