What AI does in Gregarious, and what it does not.
Three features in the core product send text to an AI model: the sprint tutor, first-draft scores on short writes that the teacher confirms, and translating a family message when someone taps Translate. The decisions a school cares about most, which gaps to teach, what the class misread, and when a student may be unsafe, are made by written rules a teacher can read.
Core AI principles at Gregarious
AI in the core product
What is sent, what comes back, what a human does, and what happens if the AI is wrong.
Socratic sprint tutor
BuiltWhen a student misses a sprint question and asks for help, the tutor answers with a short, encouraging message and one guiding question about the passage. It does not state the answer. On a third try it walks through the reasoning step by step so the student is not left stuck. A conversation is capped at six exchanges. Each reply is labeled "AI tutor." Only signed-in students reach the AI; a visitor to the demo sees a written sample hint, labeled as a sample.
The question, up to 800 characters of the passage, the answer the student chose, the right answer (so the tutor can steer toward it), the standard, and the earlier exchanges on that question. No student name. Anything the student types is checked by the safety screen below before it is sent.
Two or three sentences: an acknowledgement and one guiding question.
No one reviews each reply as it happens. The tutor is limited by its instructions to questions and hints, and it never scores anything: the sprint score comes from the student's own answers. A student can tap "Something wrong?" under any reply and, if they like, say why: wrong, confusing, not appropriate, or something else. A reply marked not appropriate goes to the student's teachers and the school's admins the way a safety alert does, showing what the tutor said.
A confusing hint costs the student one question, not a grade. Each flag is counted for the school's admins with the question and which hint it was, so the instructions can be fixed; who flagged it and what the student wrote are not kept. You can also tell us at cgb003@gmail.com.
First-draft scores on short writes
BuiltWhen a student hands in a short written answer to a question the teacher posted, the AI drafts a score on three traits (answers the question, uses evidence, explains it), 0 to 2 each, and a one or two sentence note. The teacher sees the draft next to the answer and confirms or changes it.
The passage (from the question bank, or an excerpt the teacher pasted), the question and the student's answer. Once the teacher, or their team scoring together, has scored answers to the same question, also up to two of those answers with those scores, so the draft leans toward how that teacher scores; those come only from students whose family's opt-in still stands. No name, no code, no class. The answer passes the safety screen below first; one that matches is not sent or kept.
Three trait scores and a short note, for the teacher only, marked "AI draft: you decide."
The teacher confirms or changes every score. The student sees nothing until the teacher does, and only the score the teacher confirms is kept as the student's result.
The teacher changes it before anyone else sees it. A draft is a starting point, never the grade.
Translating a family message
BuiltWhen a teacher or parent taps Translate under a message they received, that one message is translated: into the language the parent chose, or into English for the teacher. The original stays on screen above the translation.
The text of that one message, with the student's, teacher's and parent's names swapped for placeholders. Other names someone typed are sent as written. Nothing else about the family or the student.
The translation, shown only to the person who tapped, labeled "Translated by AI. It may not be exact." It is not saved.
The reader decides what it means, with the original right there. A teacher sees which language a family reads, so they can write short, plain sentences.
A translation can miss tone or a detail. Tap "Something wrong?" under it to flag it: the school's admins see a count, never the message, the translation or the language. For anything that matters, such as a meeting, a grade or a worry about a child, use the school's interpreter services as well.
Decided by written rules, not AI
These are the parts of Gregarious most likely to be mistaken for AI. Each one follows rules that are the same for every student and that a teacher can see.
Finding learning gaps
Each standard a student has data on is called a big gap, a minor gap, an early signal or closed. Every sprint and every test a teacher logs counts. Attendance and behavior are never used.
Fixed rules, the same for every student. For example, a big gap is 3 or more data points, at least two in three below 70%, and under 60% overall. The teacher sees the rule under "How was this decided?" on the student's page.
Makes the call: agree, or say it is not a gap and why. A gap the teacher sets aside leaves the student's program and what is sent home, until a newer data point brings it back for another look.
What the class misread
For each class gap, the question the class missed most and the wrong answer most of them chose.
Counting. The wrong answers in the question bank are written as specific misreadings, so the most common one points at what to reteach.
Decides what to reteach and how. "Copy for your team meeting" turns the class picture into a note for colleagues, with no student named.
Choosing the next sprint question
Each sprint adjusts to the student: a right answer brings a harder question, a miss brings an easier one with the explanation.
Each question has a difficulty rating, and the next one is picked near the student's level. Students see "Adjusting for you" so the change is not hidden.
Assigns sprints by standard and sees, under each assignment, whether it worked: scores from before against scores since.
Safety screen
Every message a student types to the tutor, and every piece of Studio work they share, is checked for signs of self-harm or abuse before anything else happens.
Written patterns, not AI. When one matches, the student gets a caring reply that points them to a trusted adult, and the message is not sent to the AI.
The student's teachers and the school's admins get an alert that names the student, at most one per hour for the same place. An adult decides what to do.
The weekly family email
Parents who opt in get a short weekly note: sprints finished, EP earned, the biggest step, and one thing to try at home.
Fixed sentences filled in from the week's records. No AI writes it.
Teachers and families can also message each other directly in Gregarious.
Other AI in the code
The code also holds AI tools outside the core product, mostly for teachers: drafting lessons, questions and worked examples, and suggested feedback on other written work. No page in the core product uses them, and none of them is offered to schools today. Each one requires a signed-in user, and a student's text passes the same safety screen before it reaches the AI. If one is, it will be documented here first, with what is sent and what a teacher does, before it reaches a classroom.
Questions about AI use? Concerns about a specific feature?
cgb003@gmail.com