Learning Twin
The substrate. Per-topic mastery + Bloom's level + Lexile + accommodation context, updated after every submission. Nothing else in the AI surface runs without it.
Audit-gatedPillar · AI personalisation
Personalisation by Bloom’s level — not by vibes.
Most “personalised” ed-tech sorts students into bins and calls it a day. CurioPilot builds a per-subject mastery model — the Learning Twin — and calibrates every activity, hint, and recommendation against it.
GDPRCOPPAFERPATraceLayer audit log
app.curiopilot.com / maya / mathLive
Learning Twin · Math · Grade 4
Maya R. · updated 2m ago
Equivalent FractionsApplying88%
Long DivisionUnderstanding64%
Decimals → FractionsApplying72%
Mixed NumbersAnalysing81%
Word ProblemsUnderstanding58%
Place ValueEvaluating94%
Active misconceptions
Treats fraction bar as subtractionProcedural · 3×
Ignores remainder in long divisionConceptual · 1×
Recommended next 3
1. Word problems · ZPD-pinnedWhy?
2. Long division with remaindersWhy?
3. Fraction → decimal practiceWhy?
The thesis
The bin model is broken.
Most “adaptive” ed-tech sorts a child into one of three buckets — easy, medium, hard — based on a single quiz score. Then it serves the next activity from that bucket. That’s not personalisation; that’s a difficulty knob with a smile.
Real personalisation needs a per-topic, per-subject, per-Bloom’s-level mastery model that updates after every interaction. It needs to track what kind of mistakea student just made, not just whether they got it wrong. It needs to know that “ignored remainder in long division” is procedural, fixable in two minutes; “doesn’t grasp fractions as parts of a whole” is conceptual, weeks of work.
That’s the Learning Twin. It’s not a marketing layer; it’s the substrate the rest of the AI features run on.
The loop
How the AI runs.
Five nodes, run on every activity. Every iteration writes to TraceLayer.
- 01
Student takes activity
Submission + hint requests captured.
- 02
AI extracts misconceptions
Bloom's classifier + misconception detector run on the submission.
- 03
Subject profile updates
Per-topic mastery + Lexile + accommodation context refresh.
- 04
Recommendations recalculate
Next 3 topics surface with a 'why' tooltip.
- 05
Next activity calibrates
Difficulty, format, and reading level pinned to ZPD.
Closes the loop → back to step 01 with a calibrated next activity.
The catalog
Everything that runs on the Twin.
Each feature is consent-gated and audit-logged. The shield means it’s enforced in code by TraceLayer.
Audit-gated
AI Quiz Generator
Pulls the student's current ZPD, Lexile, and active misconceptions, then drafts a 5–10 item set calibrated to the next-best topic. Teacher edits + approves before a student sees it.
Audit-gatedSpark — hint-giver
Hints, not answers. Subject-scoped, Bloom's-aware. Walks through reasoning step by step when a student is genuinely stuck — but won't write the essay or do the maths.
Audit-gatedAI Writing Coach
Inline feedback on argument structure, evidence, and clarity. Never rewrites the student's voice. Surfaces specific reverse-outline suggestions a teacher can adopt with one click.
Audit-gatedPersonalised Learning Paths
30/60/90-day topic plans drafted from the Twin, with a 'why this order?' rationale per step. Re-plans automatically when mastery jumps or stalls.
Audit-gatedSM-2 spaced repetition + spiral review
Forgetting-curve scheduling layered with weekly spiral review of past topics. Practice surfaces what's about to slip — not what was hardest yesterday.
Audit-gatedMay 2026 — new in the AI surface
Four capabilities that landed this sprint.
Photo-grading. Question variants. Multi-modal stems. Parent Coach Brief. Each runs inside TraceLayer; each respects the same consent and redaction gates.
Anti-cheat by paraphrase
Question variants library
Every quality-score ≥ 80 item spawns 3–5 AI paraphrases — same answer key, structurally validated. A class of 30 sees 30 variants of "What is 7 × 8?" Deterministic per-student assignment via variantIndexFor() means a page refresh shows the same variant a student started on. Generator runs as a 5-minute cron drainer so activity creation isn't blocked. Sibling de-dup is real.
Audit-gatedVisual stems
Multi-modal stem images + SVG diagrams
Two paths. Imagen 3 for STEM scenes that need a generated image — every image runs through fail-closed moderation before storage. Per-slot failure falls back to text-only; the activity never half-renders. For maths that doesn't need a model, seven pure-function generators (number-line, coordinate-plane, fraction-bar, function-graph, pie-chart, bar-chart, molecule) — screen-reader-friendly, mathematically correct, sanitiser strips every unsafe SVG vector.
Audit-gatedPhoto-grading
Grade handwritten work from a phone shot
Snap a photo of a notebook. The platform OCRs the handwriting, runs the output through the same rubric the AI generator used to author the question, surfaces partial credit on multi-step problems, and lets the teacher confirm or override before the grade lands. Image moderation is fail-open on student-upload paths (so a child's submission isn't blocked by a timeout) and fail-closed on AI-generation paths. License-gated under photo-grading.
Audit-gatedParent Coach Brief
One short weekly read per child
Premium Family + B2B Standard+ parents receive a weekly per-child digest: what's growing, what's stuck, what to try this week. Mind Mirror bands only — no numbers. The Coach Brief is generated server-side from a fixed prompt template, pinned to opt out of the smart router because the writing voice and the no-numbers rule are tightly enforced. Free / Basic / Family see a monthly Curiosity Report instead.
Audit-gatedClass-level signal
Calibrate per cohort, without identifying any child.
Nightly aggregation computes each division’s standardDeviationVsGrade. When > +0.3, the cohort is stretch; when < −0.3, the cohort is easier.
The activity generator reads this signal as a soft bias — never as a hard rule. Privacy floor: signals nullify when studentCount < 3 so a small class can’t be uniquely identified through its difficulty profile. Sigma values never surface to prompts or UI — only the three-band signal.
Cohort signal · last 24h
5Astretchn=28
5Bon-graden=26
5Cflooredn=2
5Deasiern=24
floored · n < 3 · signal suppressed
detectCrisisLanguage
pattern matcher · < 1ms · no LLM
self-harm-stated0
suicidal-ideation0
abuse-disclosure0
bullying-victim0
severe-hopelessness0
eating-disorder0
substance-use0
~36 English + ~13 Arabic patterns. Negative patterns suppress pre-teen drama (“dying of embarrassment”) within 80-char proximity windows.
Child safety
A deterministic safety net on student-authored prose.
When a student writes — in the Writing Coach, a peer review, a reflection — detectCrisisLanguage scans the prose against seven categories at sub-millisecond latency.
No LLM on this path. Child-safety triage must be deterministic. Pattern tables are reviewed by a child-safety lead and cite Mental Health First Aid for Schools plus the SAMHSA youth-crisis taxonomy.
On a high-severity hit, the writing flow short-circuits with a CRISIS_DETECTED sentinel and an inbox alert fans out to school-admins — subject and preview never name the student or carry the excerpt. The queue UI is the authorised reveal surface. Counsellor escalation triggers the human intervention; the AI never tells the student “you’re in crisis”.
Adjacent surfaces: AI-prompt rejections land in safety_blocks for super-admin triage at /super-admin/safety/blocks. Image moderation runs at every image entry point.
Three copilots
Spark, the Parent Guide, and the Teacher Coach.
One architecture, three audiences. All three pass through the same consent gates. All three log to TraceLayer.
for students
Spark
Hint-giver, never answer-giver. Subject-scoped, Bloom's-aware. Will walk through reasoning step by step when a student is genuinely stuck — but won't write the essay or do the maths.
for parents
Parent Guide
Plain-English chat for parents. "What did Maya struggle with this week?" "Why is this activity recommended?" One-click actions: schedule a check-in, switch to a calmer topic, change the consent state.
for teachers
Teacher Coach
Friday digest with six action cards ranked by impact. Drafts reteach lessons, parent updates, conference briefs. Teacher edits + approves; nothing leaves the platform without a human signing off.
The honest list
What our AI doesn’t do.
Every “won’t” below is enforced in code, not just policy. Most of them refuse to compile if violated.
- Send student names, emails, or birthdays to AI providers — the redactor refuses to compile if you try.
- Train models on your students' data. (Enterprise tier contracts forbid this on the provider side too.)
- Replace teacher judgement on grading. Every AI score is overridable; the teacher's score wins.
- Personalise without consent. The DPA + COPPA gates run before any prompt is assembled.
- Make decisions about students that aren't reviewable. Every AI decision lives in TraceLayer, forever.
- Use third-party trackers in the AI surface — no behavioural ad pixels, no cross-site fingerprinting.
- Provide AI features without redaction. The PII boundary refuses to compile a route otherwise.
Each item above is enforced by TraceLayer + the build-time PII attestation. Read the full compliance posture at /compliance.
FAQ
Questions admins, parents, and DPOs ask.
After every submission. The Bloom's classifier and misconception detector run on the response, then per-topic mastery, Lexile, and accommodation context refresh in roughly 800ms. Every update writes a row to TraceLayer with the input hash, model version, and confidence score.
Demo
See the Learning Twin live.
30-minute walkthrough with you, your IT, and your DPO. We’ll build a Twin from a sample student dataset and show you how it updates in real time.