Skip to main content

Trust and your data

What OryxLearn keeps, what the AI decides and what plain code checks, what an evidence page can and cannot tell an employer, and how to delete everything.

What we store

  • Who you are: a random guest id in a signed cookie. No email, no password, and no name unless you type one to publish.
  • Your maps: the posting, PDF text or topic you gave, and the skills built from it.
  • Your attempts: your explanation, the follow-up question and your answer, the score, the words that met each checkpoint, and how you wrote it: word count, active typing time, whether anything was pasted.
  • Your pre-test results: if you took a course's optional pre-test, which topics you answered, whether each was known, missed or skipped, and when. The answers you typed are checked and not kept. The questions themselves are shared by every learner of the course.
  • What you chose to share: evidence pages (with the name you typed). Anyone with an evidence page's link can read it until you withdraw it.
  • Study and practice: flashcards and quiz questions made for your maps.
  • Study circles: the name you gave the circle and which copy of the map is yours. Other members see your name, which skills you proved and when, never what you wrote.
  • Streak: the days you proved something, your sparks this week and when you last relit each topic. Only you see it.
  • People and review: your public profile on /people (name, city, contact link) and your spaced-repetition schedule and review days.
  • Only in this browser: unsent drafts, the theme, the text size and the Accessible learning mode setting.

It is kept in the app's cloud storage. Postings, explanations and answers are sent to an AI model (Claude Haiku, from Anthropic) to build maps and lessons and to grade. Courses and lessons made from a topic are saved once and shared, so two learners who ask for the same course get the same one. They hold no names, accounts or source text.

What the AI does, and what code checks

The AI model

  • Picks the skills in a posting, their order, and 3 or 4 checkpoints for each.
  • Says which checkpoints your explanation covers, quoting your words.
  • Points out factual errors.
  • Writes flashcards and quiz questions.

Plain code, no AI

  • Checks each quote from the posting is really there. If not, it says so.
  • Checks each quote from your explanation is really in what you wrote. If not, that checkpoint does not count.
  • Gives no credit for pasting a checkpoint's own wording back.
  • Works out the score: checkpoints met out of 100, minus 20 per factual error. 50 passes.
  • Decides what unlocks, and measures how you wrote, in your browser.
  • Fingerprints and seals each evidence page, so a reader can check it was not edited.

What an evidence page shows, and what it does not

It shows that this person's own words covered each checkpoint, with the quotes checked by code. It shows they answered a live question about what they wrote, how they wrote it, and that the record has not changed since it was published.

It does not show that they did it without help. A chatbot given the checkpoints can get through: in our saved runs, 15 of 16 chatbot answers written with the checkpoints in hand reached the pass mark, and the same chatbot passed the follow-up 13 of 13 times. Pasting is shown as a signal, never as an accusation, and it does not block a pass.

It also does not show who the person is, since the name is whatever they typed, or that an expert set the bar: the checkpoints are written by AI from the posting.

How well the grading works, measured

70 of 71
test explanations got the verdict their label expectedLikely true rate: 92 to 100%
0 of 48
explanations that should fail were passedLikely true rate: 0 to 7%
1 of 23
explanations that should pass were failedLikely true rate: 1 to 21%
8 of 8
made-up quotes planted in a grade were caught by code, with no AI involved

Read these with their limits

  • An AI (Claude) wrote all 71 test explanations and the verdict each should get. No teacher or expert has checked them yet.
  • 71 cases is a small test, and it has not been tried with real learners.
  • The chatbot figures come from saved runs with the older pass mark of 70, read against today's rules, not rerun.
  • Measured with gpt-oss-120b on November 7, 2026, before grading moved to Claude Haiku. The full report is docs/eval-results.md in our repository.

Accessible learning mode

Turns on Lexend, a typeface designed for easier reading, with larger text, more space between lines and letters, and no motion. It also adds “Read aloud” buttons to each skill and each result. Those use your browser's own voice: nothing is sent to us. The setting is saved in this browser only.

Everything works with a keyboard. On the map, Tab moves between skills, Enter opens one, and Escape closes it.

Delete my data

Deletes your maps, attempts, scores, proven skills, and pre-test results from our storage, takes down any evidence page you published, takes you out of every study circle (and closes the ones you started), and starts this browser as a new guest. If someone joined a study circle you started, their copy of the map is theirs and stays.

Not covered yet: the flashcards and quiz questions made for your maps. They hold no words you wrote, and they can no longer be linked to you once your session is replaced.