AI at the front line. People hold the boundary.
AI drafts tasks, hints, practice, error scheduling and weekly summaries; teachers own disputed answers, content sign-off, progression decisions and safety escalation.
CIE / OPEN LIBRARY · AI STUDY COMPANION
Move learning forward to
the next verifiable action.
A controlled-autonomy learning pilot for Shenzhen secondary-school students, with China's National College Entrance Examination (Gaokao) as the long-term direction. AI organises the next task, asks for the basis and schedules practice; teachers review content, make key judgements and take over when needed.
CIE context
This is not a new CIE sub-brand. It is an Open Library research draft: Explore finds the gap, Build turns it into practice, Launch tests it with evidence, and Grow builds long-term capability.
01 / WHY
Start with a version built to improve learning, be audited and grow.
Improving examination performance is the pilot goal; any conclusion requires an equivalent baseline, post-test and comparison condition. Test the smallest useful slice before expanding.
One high-school grade, Chinese, mathematics and English.
Lower secondary and other subjects may reuse the product shell, but each needs its own curriculum, question bank, textbook version and measures.
Long blocks are built around rest.
Weekdays are 60–120 minutes. A Sunday total of 4–8 hours includes rest and paper practice; the working default is 4–6 blocks of 45–50 minutes.
North-star measure
Independent performance without hints + 7/21-day delayed retention.
Explicitly out of scope
No ghost-writing, exam-time answering, psychological or emotional diagnosis, or guaranteed score/admission conclusion.
Service boundary: the first slice is for ordinary study needs. ADHD, dyslexia and other specialist-support contexts require assessment and a separate service by qualified school, health or special-education teams. This draft is not an enrolment, paid-service or clinical-support route.
02 / HOW IT WORKS
Every session leaves a trail we can replay.
Try first. Work independently. Then review. Automation narrows when confidence or safety falls.
- 01
Learning profile Explore
Goal, year, subject, prerequisite gaps and an initial error label; a self-check is not a clinical diagnosis.
- 02
Plan proposal Build
AI proposes tasks, load and priority; student and teacher can see the basis, reschedule or leave.
- 03
Guided study Build
Keep the attempt, ask for the basis, then move from a directional hint to a relation and a micro-explanation.
- 04
Independent attempt Launch
Complete a parallel, variant or transfer task without hints. Paper and verbal explanation are evidence.
- 05
Spaced review Grow
Place the error in a 7/21-day queue, update mastery and produce the smallest next practice.
- 06
Human handover Human review
Uncertainty, three consecutive errors, a disputed answer or a safety signal pauses automation and routes to a person.
- 07
Shared reflection Grow
Student, teacher and a minimum-necessary guardian summary review the evidence and choose continue, adjust or exit.
StudentConfirm the goal · attempt first · pause, reschedule or leave
Teacher / adviserSign off content · review uncertainty · handle escalation and appeals
Parent / guardianSee the minimum-necessary summary · support load and rest, not every exchange by default
L0 · On demandStudent-initiated questions, hints and practice
L1 · RecommendAI proposes; the student confirms
L2 · Controlled proactiveAuthorised time-window reminders; always switchable off
L3 · Not availableDiagnosis, school contact, admission verdicts, ghost-writing or live exam answers
03 / CURRICULUM
A 12-week MVP turns the goal into a weekly loop.
Five weekday sessions of 60–120 minutes, plus an optional Sunday block split around rest. Each week produces independent work, an error record and the next review date.
Micro-lesson · baseline task
90–120 minTiered practice · feedback
About 90 minExplain aloud · error label
90–120 minShort paper · repair task
90–120 min7/21-day check · next plan
60–90 minWeekly paper · error sort · preview / STEM transfer
240–360 effective min, with restEntry and equivalent baseline
Goal, time, device, guardian consent and data boundary; create the first learning profile.
Output: consent record · baseline report · first-week planMap and close gaps
Concept cards, retrieval, tiered hints and the first 7-day delayed check.
Output: knowledge map · error card · review calendarTransfer and midpoint test
Context tasks, define-first and cross-chapter variants; calibrate load after an equivalent midpoint test.
Output: midpoint report · teacher-review task · adjusted planPaper calibration
High-frequency types, timed half/full papers and error-pattern review; uncertainty routes to a person.
Output: question pack · error report · second training taskMinimum intervention and exit
7/21-day retention, equivalent post-test and a choice for the next cycle; completion is not automatic renewal.
Output: retention · learning-gain report · continue / adjust / exitEight actions in a 60–120-minute sessionCheck-in → retrieval without hints → ask for the basis → tiered hint → independent parallel task → error attribution → 7/21-day schedule → one-sentence reflection
04 / METHODS & EVIDENCE
Borrow the methods. Do not copy the products.
US, Canadian and European products and learning-science research are inputs and risk signals only. Functionality, pricing, availability and evidence change; none is a CIE endorsement.
Socratic guidance + teacher-in-the-loop
Public patterns from Khanmigo, MATHia and Synthesis translate into “ask for the basis, tier the hint and show the process to a teacher”.
Khanmigo ↗Human escalation + curriculum alignment
CANtutor and Paper illustrate human support and school collaboration. Paper evidence cited publicly refers to its GROW human/mixed programme, not AI alone.
Paper ↗Retrieval, spacing + continuous assessment
CENTURY, Seneca, Sparx, StudySmarter and Nolej show structured practice and content production; CIE must re-test each method against Guangdong materials.
CENTURY ↗Put mechanisms into the interaction
Retrieval, spacing, mastery, metacognition and transfer are hypotheses to test. Report performance without AI as well as with it.
EEF review ↗Rule for source-page counts
Subject, chapter, knowledge-point and question counts shown in the supplied pages are page-displayed or planned values (snapshot: 30 August 2026), not CIE inventory, completion or outcome measures. A formal question bank needs textbook/version, source, copyright, answer and dual-review gates.
Year / subject / textbook version / skill / type and difficulty / stem, answer and explanation / source rights / reviewer / release version.
Attempt, time, hint level, error label, repair action, parallel or variant check, 7/21-day retention and human handover.
Unverified questions stay out of the formal bank. Drafts, released items, policy information and personal records are separated, retractable and rollback-ready.
05 / PROOF
Measure independent performance before talking about scores.
No online minutes or message count substitutes for learning results. Any effect claim must state the baseline, equivalent post-test, comparison condition, confidence interval and adverse measures.
What we will report
- Independent performance and 7/21-day delayed retention
- Equivalent pre/post gain, effect size and comparison condition
- Answer accuracy, source coverage and human-handover rate
- Fatigue, sleep, family conflict and dependency signals
What we will not claim
- Source-page counts as CIE inventory
- Cases, usage or satisfaction as causal score gain
- Guaranteed marks, admission, competition or progression results
- Competitor features as partnership, endorsement or ranking
Research entry points: retrieval and spacing review · US Department of Education AI report · UNESCO guidance
06 / SAFETY
For a minor-facing programme, safety is a process.
This research draft carries no real student reports. Accounts, uploads, proactive messages and voice features open only after privacy, safeguarding, supplier and incident processes are reviewed.
Collect less, retain briefly
Keep only year, subject, attempts, errors, plan and necessary dialogue. Voice, video, camera, screen and precise location stay off by default.
Age-appropriate and verifiable
Show purpose, scope and retention. Student and guardian can pause, access, correct, export, delete or appeal.
Trigger the human protocol
Repeated errors, ghost-writing/cheating, bullying, self-harm or privacy exposure pause automation and follow the human plan.
Answers have a source
Formal questions bind to textbook, source, year, rights and reviewer; no-source or uncertain answers are not delivered as fact.
Think first, then see the answer
Tiered hints, independent attempts and parallel checks; no ghost-writing, live exam answers or proctoring evasion.
Retractable, rollback-ready, appealable
Record model, knowledge source, hint level, human edits, publisher, deletion request and version rollback.
CollectOnly fields needed for the current task
UsePlan, feedback, review and human support
RetainShort retention, access tiers, de-identification and audit
ExitPause, access, correct, export, delete and appeal
Gate before opening
Mainland-China counsel must review the personal-information impact assessment, minor consent, supplier no-training/no-retention terms and incident rehearsal. This page is product triage, not legal advice.
07 / PILOT
Production comes after the gates, not before them.
Clean content, compliance and safeguarding first; then invite a small cohort. Places, fees, opening date and textbook version are not yet decided and are not promised here.
Prepare
Interviews, three-subject content clean-up, consent text, rights register and risk red team.
Usability test
Up to 20 people; check entry, study loop, errors, weekly report and teacher handover.
Controlled pilot
A 60–120-person candidate window with pre-registered measures and weekly human review.
Careful expansion
Only after safety hard stops and learning evidence pass: more years, modalities or school dashboards.
Before entry
Confirm year, textbook version, goal, time, device, guardian consent and visibility; complete the equivalent baseline.
During service
Student-led service. The proposed pilot response target is review of uncertain content within 24 hours and immediate pause/human handover for safety signals; the approved protocol controls.
Exit
Compare equivalent post-test and 7/21-day retention, issue a learning-gain report, then continue, adjust or leave. No automatic renewal.
Is this an answer machine?
No. The student attempts first, then receives a tiered hint and a parallel check; teachers review key answers.
Can you guarantee a score increase?
No. Improving examination performance is the pilot goal; any conclusion needs equivalent baseline, post-test and comparison conditions.
Will guardians see every exchange?
Not by default. Share the minimum-necessary summary, load and risk signal needed for the next action.
How do I join, pause or leave?
Participation conditions will be announced after approval. Once enrolled, a student can pause, reschedule, access, correct, export, delete or leave under the agreed process.
When will it open and what will it cost?
After content, data and safeguarding gates pass. Places, pilot support and formal pricing remain subject to approval.
08 / DOWNLOAD
Take the complete proposal back to the team.
The 16-page A4 brochure covers the student profile, learning loop, weekly rhythm, 12-week curriculum, international methods, question bank, evidence, safeguarding, pilot and FAQ.
Download the complete brochure (PDF) PDF · A4 · 16 pages · research preview · service not openThe public download is photography-free and uses CIE token geometry. The supplied student photograph remains in the local full-photo internal review edition; public use requires documented guardian, photographer, venue and third-party-mark permissions.
香港科创资助与专项文件准备
按项目准备商业计划书、创始人介绍、研发证据、预算、股权与申请材料；政策、期限与证据状态分别核实。
打开本项目资助申请与资料清单 →