We have released v1.0.228. It comes in two halves. The first keeps required training going and proves it: recurring cycles, a post-course attestation, and certificates with a PDF, a verification page and an expiry date. The second is the AI learning platform: an API and MCP tools that ask questions at the learner's level, guide without giving the answer away, and say why an answer was wrong.
Highlights
Recurring training
Set "every 12 months" and the next cycle starts on its own after the due date. History is kept.
Attestation
Require "I have read and understood this". The course is not complete until the learner confirms it.
Certificate PDF, verification, expiry
A PDF with your name and logo, a URL anyone can check, and automatic renewal before expiry.
Concepts and a verified question bank
A map of what learners must understand, and questions whose answers stay right whatever the numbers.
Adaptive practice and staged hints
The next question follows the last answer; two misses lead back to the prerequisite. Five hint steps, none of them the answer.
Error analysis, mastery, safety
Why an answer was wrong, with evidence. Mastery per concept with review dates. Nightly checks and a kill switch.
1. Recurring training runs by itself
A learning path can now be set to recur every N months. When the due date passes, Manabu2 starts the next cycle.
- The previous cycle's completion is kept as history and stays visible in the compliance report
- The new cycle resets lesson completions and attestations, and the learner is told that a new cycle has started
- Paths assigned to departments recur the same way. When a learner has the same path both directly and through a department, the earlier due date applies
- The rollover runs daily at 08:30 Japan time
2. Attestation after a course
A course can require an attestation with a statement you write. After finishing every lesson, the learner must confirm "I have read and understood this" before the course counts as complete.
| What is recorded | Details |
|---|---|
| Time and wording | The statement as it read at the time. Changing the text later does not change past records |
| Channel | Whether it was given on the web, through the API or from an AI client |
| Origin | IP address and device, for audit |
The compliance report and the certificate detail page show whether and when the attestation was given. A new recurring cycle asks for it again.
3. Certificate PDF, verification page and expiry
PDF download
Learners can download a PDF of each certificate from My Certificates, typeset with a Japanese-capable font. The API can fetch it too.
Organization name and logo
Under HR > Branding, set the name and logo printed on certificates. Without them, the organization's display name is used.
Verification page
Every certificate gets a verification URL. A recruiter or client who is not signed in to Manabu2 can confirm it is genuine and still valid. Only the certificate itself is shown, never other learning records.
Expiry and renewal
A certificate can carry a validity period in months. 30 days before it expires the learner gets an email and the paths behind it reopen for renewal. Completing them renews the certificate automatically.
| State | Shown as |
|---|---|
| Valid | Within the period, or no expiry |
| Expiring | Within 30 days; the renewal paths are open |
| Expired | Past the date and not yet renewed. The verification page says so |
4. The AI learning platform, for learning apps and AI clients
For learning apps such as WakaRoute, the API and MCP now offer a way to build learning around whether a concept is understood. The design principle: the AI does not invent questions freely. People verify the questions; the AI guides.
4.1 Concepts: a map of what must be understood
Register concepts such as "adding negative numbers" or "absolute value", with prerequisites (to understand this, you need that) and common misconceptions (dropping the sign) together with how to address them. Each concept can carry the id of an external map.
4.2 A verified question bank
- A question is a template. Numbers become variables with constraints such as "a is not zero", so one question yields many variants, and the answer is a formula, so every variant is graded correctly
- Expected wrong answers map to misconception codes, so choosing one tells you which misconception it was
- Questions go draft → review → published, and someone other than the author publishes. Learners only ever see published questions
4.3 Adaptive practice sessions
- Start with a set of concepts and each next question follows the answers so far. Two misses in a row lead to a question on the prerequisite
- The same conditions give the same questions, so a learner who closes the app resumes on the same question. Every pick is recorded with its reason for a teacher to inspect
- Question count, time limit, retries and hints are configurable; hints off makes an exam
4.4 Hints that do not give the answer
| Step | What it does |
|---|---|
| 1. Intent | What is the question asking |
| 2. Small hint | When the last wrong answer matched a known misconception, its remedy |
| 3. Prerequisite check | "Before that, are you sure about ...?" |
| 4. Rephrase | The same question in other words, when the author wrote one |
| 5. A similar problem | Other numbers, worked through with the answer and explanation |
Seeing the answer is the learner's own choice, and a question revealed that way does not count as understood.
4.5 Error analysis
A wrong answer is classified as careless / misconception / prerequisite gap / misread / unknown, with evidence and a confidence. Rules go first (a match with an expected wrong answer, a flipped sign, off by one, swapped digits, answering with a number from the question); the AI is asked only when there is working to read or the rules cannot decide. The correct answer always comes from the bank, never from the AI, and when nothing reliable can be said the answer is "unknown".
4.6 Mastery and re-checks
- Per concept, observed facts (attempts, solved unaided, hints taken, answers revealed) are returned separately from the estimate
- The first unaided correct answer schedules a re-check tomorrow; each passed re-check stretches the interval: 3 days, a week, two weeks, a month, two months. A miss brings it back to tomorrow
- Stages (meaning → basics → connections → application → exam-stable) are defined per organization. The upper stages need passed re-checks over time
- A review mode practises only the concepts that are due
4.7 Learning activity for integrations
Session start, answers, hints, completion, lesson completions and mastery updates are available as a chronological feed. Pass the last position you read and only what follows is returned; each event has a stable id, so nothing is counted twice. The same events are pushed as webhooks. Deletions appear in the feed as well.
4.8 Safety and monitoring
- Banned terms: words unsuitable for minors never appear in questions, options, explanations or hints. Organizations can add their own
- Nightly evaluation: every published question is checked each night for broken variants, hints that leak the answer, and similar problems that are not actually different. No AI is used, so it costs nothing
- Kill switch: an administrator can stop AI use at once; rule-based features keep working
- Reports: learners can flag a question or hint as wrong or unsuitable; the organization reviews them
- Human review: when a teacher agrees or disagrees with an analysis, that verdict joins the evaluation set and becomes the standard for future runs
5. API and MCP additions
Developers will find the details in the OpenAPI document and the MCP tool list on the developer portal. The main additions:
| Addition | What it does |
|---|---|
/api/v1/organizations/{id}/concepts |
Concepts, prerequisites, misconceptions, stages |
/api/v1/organizations/{id}/question-banks |
Banks, variants, grading, review and publish |
/api/v1/me/learning-sessions |
Start, answer, hint, reveal, analyze, resume, end, report |
/api/v1/me/mastery, /api/v1/me/activities |
The learner's mastery and activity feed |
/api/v1/organizations/{id}/learning-sessions, /activities, /members/{userId}/mastery |
Audit, sync and mastery for the organization |
/api/v1/organizations/{id}/ai-ops |
Kill switch, banned terms, evaluation set and runs, usage, reports, review |
POST /api/v1/me/courses/{id}/attest, GET /api/v1/me/certificates/{id}/pdf |
Attestation and certificate PDF |
| MCP tools (concepts) | list_concepts, upsert_concept, link_lesson_concepts |
| MCP tools (question bank) | list_question_banks, create_question_bank, add_bank_question, render_bank_question (publishing is done by a person, so it is not an MCP tool) |
| MCP tools (practice) | start_practice_session, answer_practice_question, get_practice_hint, reveal_practice_answer, analyze_practice_answer, get_practice_session, get_my_mastery, report_practice_content. The AI client never sees the answer key |
MCP tool attest_course |
New: attestation |
| Webhook events | session.started / session.answered / session.hint / session.revealed / session.ended / mastery.updated / session.deleted / learner.data_deleted |
The paths API now returns labels and course counts in its list, and a course's progress includes the paths it belongs to.
6. Delivery and fixes
- Production deploys now run one at a time. Several updates in quick succession no longer overlap, so there is no gap in service while a new version starts
- Fix: attesting a course from an AI client (
attest_course) asked for an organization id and failed. Fixed - Fix: the developer portal's MCP tool list was missing the concept and question-bank groups. Fixed
What is next
The AI learning platform continues to be evaluated ahead of full use in WakaRoute. Next up: in-app notifications for due dates and renewals, and English content.