Learning that knows
the learner.
PRISM is an AI tutoring system running locally on Llama 3.2 3B — no cloud, no subscription, no data leaving the device. It tracks mastery per skill, detects emotion and attention in every interaction, logs every safety event, and generates structured learning reports for each session. Education that adapts instead of forcing adaptation.
Most learning systems are static.
PRISM isn't.
Most educational software assumes a baseline, delivers content linearly, and leaves the learner to struggle when the structure doesn't fit. PRISM inverts that model. It continuously adjusts pacing, explanation depth, topic focus, and interaction style based on what the learner actually demonstrates — not what a curriculum assumes they know.
Active Dialogue, Not Passive Delivery
PRISM engages in real-time conversation. When a learner says "fractions," PRISM responds with a question, a worked example, and a follow-up — adapting the explanation based on how the learner responds. Every session is a two-way exchange, not a broadcast.
Session Continuity
Learning sessions are not isolated events. PRISM maintains session state — active mode, current topic, lifecycle stage — across the full interaction. A learner who stopped mid-concept yesterday picks up from exactly that concept today. The tutor remembers. The learner doesn't have to re-explain.
A hard gate that cannot be overridden.
PRISM is designed for children. The safety architecture reflects that. Every query and every response passes through a two-phase safety gate before anything reaches a learner. The gate does not warn. It does not ask for confirmation. It blocks — and it logs every event in full detail.
PRISM knows what the learner
actually knows.
Every skill a learner encounters is tracked individually — not with a simple pass/fail flag, but with a continuous confidence score updated across every attempt. PRISM knows the difference between a learner who got lucky once and a learner who has demonstrated consistent understanding.
Per-Skill Confidence
Every skill has a continuous confidence score — a real number between 0 and 1 updated after every interaction involving that skill. The score reflects demonstrated ability, not just completion. A learner who guesses correctly does not get the same score as one who demonstrates understanding.
Attempt History
Attempt count and success count are stored separately per skill. PRISM can identify a learner who is attempting frequently but not succeeding — a signal to change approach, not increase pace. The data drives the teaching. Not the other way around.
PRISM reads the room. Every interaction.
Every interaction PRISM has with a learner is logged with two additional signals beyond the content of the exchange: detected emotion and detected attention. These signals inform how PRISM adapts in real time — slowing down, simplifying, encouraging, or refocusing based on what the learner is communicating beyond their words.
Emotion Detection
Every interaction is classified by detected emotional state. Frustration, confusion, engagement, confidence — each changes how PRISM responds. A frustrated learner gets a different explanation than an engaged one. The emotional signal is stored per interaction and available in session reports.
Attention Detection
PRISM detects attention state from the content and pattern of learner responses. Short, off-topic, or low-engagement responses trigger a mode shift — shorter explanations, more direct questions, structured re-engagement. Attention is treated as a teachable moment, not a failure.
Learning Reports
Every session generates a structured report: highlights (what went well), concerns (what needs attention), and recommendations (what to focus on next). Reports include skill confidence data, safety event summaries, and session lifecycle metadata. Designed for parents, teachers, and learners alike.
One system. Six ways to learn.
PRISM routes each session through one of six learning modes — selected based on the learner's current state, skill confidence, and session history. Modes shift dynamically within a session as conditions change.
Tutor Mode
Primary mode. Structured concept delivery, worked examples, guided questions. Adapts explanation depth based on comprehension signals.
Practice Mode
Skill reinforcement through repetition and variation. Difficulty scales with mastery confidence. Triggered when a skill falls below the confidence threshold.
Exploration Mode
Learner-directed discovery. PRISM follows the learner's curiosity while maintaining structured guardrails. Designed for high-engagement, high-confidence sessions.
Assessment Mode
Structured evaluation of skill mastery. Updates confidence scores. Generates report highlights. Activated at session milestones and on request.
Recovery Mode
Activated on frustration or attention signals. Simplified explanations, shorter exchanges, increased encouragement. Designed to re-engage without reinforcing negative experience.
Review Mode
Session-opening recap of prior material. Bridges the gap between sessions. Activates mastery data from previous interactions to calibrate where to begin.
Built for every learner.
Not just the median one.
Most educational software is designed for the average learner and adapted for everyone else. PRISM is designed with neurodiversity and accessibility as first-class requirements — not post-launch additions.
ADHD & Attention Variability
Adaptive pacing that responds to attention signals. Shorter explanations when attention is low. Structured re-engagement patterns. No penalty for refocusing — just a different path to the same destination.
Autism Spectrum
Structured reinforcement patterns. Consistent language and framing. Predictable interaction rhythms. Explicit comprehension checks rather than inferring understanding from conversational cues.
Down Syndrome & Cognitive Accessibility
Simplified abstraction layers. Concrete examples before abstract concepts. Repetition without friction. Mastery tracking that celebrates incremental progress rather than measuring against a fixed pace.
Visual & Hearing Impairments
Text-first interaction model that works with assistive technologies. Structured content that does not depend on visual layout for comprehension. Audio-compatible output designed for screen reader compatibility.
Runs on your hardware. Nowhere else.
PRISM runs locally on Llama 3.2 3B Instruct — a 1,926MB model that loads once and runs on the device. No query ever leaves the machine. No conversation is logged to a server. No child's learning data is stored anywhere but the device it was created on.
Response Times
Cold start on first query: ~6.9 seconds while the model loads. Subsequent responses: median 2,213ms, minimum 1,323ms. After the first turn, the session is fast. The model stays loaded for the duration.
The Tutor Persona
PRISM's tutor persona is defined in a 2,289-byte system prompt (~572 tokens) — covering learning philosophy, safety constraints, response style, mode behaviour, and the hard-gate rules. The persona is stable across every session on every device.
Privacy by Architecture
Local inference is not a privacy feature. It is a privacy guarantee. There is no mechanism by which PRISM sends data externally — not because the feature is disabled, but because the architecture does not include one.
PRISM is in active development.
Built on Haven's core. Running locally today. Designed for schools, homes, and every learner who deserves a tutor that adapts to them — not the other way around.