Every AI touchpoint, documented.
For each AI feature, we document exactly what happens, what data is sent, what a human does, and what the fallback is if the AI is wrong.
Core AI principles at Gregarious
How AI is NOT used here
These boundaries are intentional. They exist because of how the platform is designed, not because of technical limitation. A district or parent evaluating Gregarious should be able to verify these claims through the feature set.
No autonomous student grading
AI scoring is always a draft. A human teacher reviews and confirms every score before it is applied to a student record. AI cannot publish a final grade on its own under any workflow in the platform.
No free-text AI image generation for students
Students cannot prompt the AI to generate images. Studio visual projects use teacher-provided assets or the guided Pixel Art tool. The AI is not in the image pipeline at any point in the student experience.
No replacement for teacher judgment
AI outputs in teacher-facing tools (misconception analysis, class briefs, project score drafts) are advisory. They appear as suggestions in the "Needs You" feed. No AI output triggers an automatic action. The teacher decides.
No data sold to model providers
Student data is never sold, licensed, or shared with Anthropic or any other AI provider for model training. Anthropic's API terms prohibit training on submitted data. We additionally enforce Zero Data Retention (ZDR) headers on every API call.
No AI-written student work
The AI Coach enforces a Socratic stance. It asks questions; it does not write essays, stories, or responses for students. The system prompt explicitly prohibits answer disclosure, and this is enforced at the prompt-template level, not just by policy.
No unsupervised AI access for students
Every student AI interaction goes through lib/safety-wrapper.js before reaching the model. There is no open-ended chat interface for students. AI touchpoints are scoped, structured, and templated.
Models we use today
Current as of 2026-06-04. This list is updated when we change providers or model versions. We will not silently switch AI providers. Any material change to the AI stack will be documented here with an updated date.
Anthropic Claude
Used for: Socratic tutoring, item generation drafts for teachers, item rationale summaries, class brief generation, reflection prompts, concerning-content detection (Fall 2026).
We do not use AI providers other than Anthropic at this time. If we add a second provider (for example, for image analysis or audio processing), it will appear in this section with the same level of documentation.
Safety wrapper
Every student AI interaction is routed through lib/safety-wrapper.js before reaching the model and before the model output is returned to the student. The wrapper enforces the following, in order:
Socratic stance enforcement
The system prompt passed to every student-facing AI session instructs the model to ask questions and provide scaffolds, never to provide direct answers to academic tasks. The wrapper validates that the system prompt contains the required Socratic directives before the API call proceeds.
Crisis-keyword routing
The wrapper scans both the incoming student input and the outgoing AI response for crisis keywords (self-harm, abuse, suicidal ideation patterns). If a threshold is met on the input side, the API call is short-circuited and the hardcoded handoff string is returned instead. The model never sees high-crisis input. See /trust/safety for the exact handoff string.
No unmoderated image generation
The wrapper rejects any prompt that attempts to invoke image generation capabilities. Students cannot reach image generation tools through the AI Coach interface. The model configuration additionally does not include image generation endpoints.
Prompt injection sanitization
Student input passed into AI prompts is sanitized to prevent prompt injection attacks. The wrapper encloses student-controlled text in marked delimiters and the system prompt instructs the model to treat content within those delimiters as student input only, not as instructions.
Response validation
Before the model response reaches the student, the wrapper checks the output for patterns that would indicate the model has provided a direct answer, generated a student-ready essay, or disclosed content outside scope. Flagged responses are suppressed and a fallback message is shown to the student.
Bible §12 -- Kid Safety and AI (verbatim)
- --Every student AI interaction goes through safety-wrapper.js.
- --Studio uses a guided Idea Wizard, not free-text AI prompts.
- --No unmoderated AI image generation in the student experience.
- --Student tutor enforces Socratic method. Asks questions, does not write essays.
- --Self-harm or abuse disclosures: respond with care, redirect to trusted adult.
- --RLS on every Supabase table. Parent accounts link to child via parent_id.
Live features
The following AI features are active in the current version of Gregarious.
Close Reading (CR) Scoring
LiveAfter a student submits a written response to a close reading prompt, the AI scores it against a rubric defined by the teacher. The score is presented to the student and logged in the gradebook.
The rubric criteria, the prompt text, and the student's response. Student name is replaced with an anonymized session token before the call. No other student data is transmitted.
A numeric score (matching the rubric scale) and a short rationale (2-4 sentences) explaining what the student did well and what the response is missing.
The teacher reviews AI scores flagged as borderline (within 1 point of a grade boundary). Teachers can override any score at any time. Final grades are always teacher-confirmed.
The student can flag a score for teacher review from their dashboard. Teachers receive the original response, the AI score, the rationale, and a one-click override button. The AI score is never final on its own.
Project Scoring
LiveAt a teacher's request, the AI provides a draft score for a Studio project against a teacher-created rubric. This is a first-read assist, not an automatic grade.
The rubric, the project description, and any written components of the project. Student name anonymized. No audio, image, or video content is transmitted.
A draft score per rubric dimension and a brief justification for each dimension.
Teachers always review AI project scores before they are applied. Project scoring is never automatic. The teacher submits the final score.
Because teachers always review before applying, the AI score is never seen by students until a teacher has confirmed or overridden it. If a pattern of inaccurate scoring is identified, the rubric can be adjusted and re-scored.
Teacher Brief Generation
LiveThe teacher dashboard generates a plain-language brief summarizing their class's mastery trends, recent struggles, and students who may need attention. This appears in the "Needs You" feed.
Aggregated, anonymized class performance data: average mastery percentage by standard, number of attempts, and score trends. No individual student names are sent. The AI receives statistical summaries only.
A 3-5 sentence paragraph summarizing trends and suggesting focus areas. Formatted for quick reading, not data-dump style.
Teachers read the brief and decide whether to act on it. The brief does not trigger any automatic actions. It is advisory only.
If the brief misrepresents the data, the teacher will see the discrepancy when they look at the underlying data. Teachers can dismiss briefs and generate a new one. We log brief accuracy feedback to improve the prompts over time.
Worked Example Generation
LiveTeachers can request a model student response (worked example) for a given prompt and rubric. This is used to show students what a strong response looks like.
The prompt text and rubric only. No student data of any kind.
A model written response that meets the rubric criteria, formatted to match the expected student output.
Teachers review and edit the worked example before it is shown to students. It is never published automatically.
The teacher edits or discards the example before it reaches students. If a worked example is factually incorrect, the teacher corrects it in the editing step.
Reasoning Scaffold Samples
LiveWhen a student is stuck on a question, the AI provides a thinking scaffold: a structured set of questions to help the student reason through the problem themselves. The AI does not provide the answer.
The question text and the standard being assessed. Student name anonymized. No prior performance data is sent.
Three to five guiding questions designed to prompt the student's own reasoning, plus a brief framing sentence. No answer content.
No real-time human review. The scaffold is generated from a constrained prompt template that prohibits answer disclosure. Teachers can see when a student accessed a scaffold (logged as a help event).
If a scaffold inadvertently reveals an answer or misleads a student, the student or teacher can flag it. Flagged scaffolds are reviewed and the prompt template is corrected.
Reflection Prompts
LiveAfter completing a Sprint session, students are shown a reflection prompt tailored to what they practiced. The prompt encourages metacognitive thinking about their approach, not their score.
The standard(s) practiced in the session and whether the student showed growth (boolean: yes/no). No score values, no name.
A single open-ended reflection question appropriate to the standard and the student's session outcome.
Teachers can see reflection responses in the student view. No required action.
If a reflection prompt is off-topic or confusing, the student can skip it. Skipped and flagged prompts feed back into prompt quality review.
In development, Fall 2026
The following features are on the roadmap. They are documented here so schools and districts can evaluate them during procurement, not after deployment.
Parent Summaries
In development, Fall 2026A weekly email to linked parents summarizing their child's activity: standards practiced, Studio time, streak status, and any teacher notes. In development, Fall 2026.
Aggregated activity counts (sessions, standards attempted, EP earned) for the student. Student's display name. No assessment scores transmitted in this call.
A plain-language paragraph for a parent audience: warm, specific to the week, and free of jargon.
Parents receive the summary by email. Teachers can add a personal note before the weekly summary sends. Parents can adjust or disable summaries from their settings.
If a summary is inaccurate, parents can flag it or contact the teacher directly. Summaries are advisory; grades and official records live in the platform.
Concerning-Content Detection
In development, Fall 2026Student writing in Studio and Pulse check-ins is scanned for content that may indicate self-harm, abuse, or crisis. If detected, the teacher of record is notified in their dashboard. In development, Fall 2026.
The text of the student's submission. Student name is included in this call because the notification must identify the student to the teacher.
A flag (yes/no) and a confidence level. If flagged, a brief context note for the teacher.
All flagged content is reviewed by the teacher before any action. The AI flag is an alert, not a diagnosis. See /trust/safety for the full protocol.
False positives are expected and are treated as low-cost: a teacher gets a notification they review and dismiss. False negatives (a missed flag) are the greater risk. We document this honestly in /trust/safety and recommend teachers maintain human oversight of student writing.
Misconception Identification
In development, Fall 2026After a series of incorrect answers on a standard, the AI analyzes the pattern to identify a likely misconception and suggests a targeted intervention to the teacher. In development, Fall 2026.
The sequence of incorrect responses for the standard, anonymized. The standard descriptor. No student name.
A probable misconception (e.g., "confusing inference with summary") and one or two suggested intervention approaches.
The intervention suggestion appears in the teacher's dashboard as a recommendation. The teacher decides whether and how to act on it.
If the identified misconception is wrong, the teacher ignores the suggestion or provides a different intervention. The teacher's professional judgment always supersedes the AI recommendation.
Questions about AI use? Concerns about a specific feature?
cgb003@gmail.com