--- title: "Education AI, built and evidenced | YPAI" url: https://ypai.ai/solutions/education/ description: "Practice companions that correct on the learner's own words, consent-gated captions, assessment data and operations AI. Metrics carry their denominators." source: "src/copy/routes (route /solutions/education/)" --- # The correction lands on the word she wrote. The number leaves with its denominator. > Practice companions that correct on the learner's own words, consent-gated captions, assessment data and operations AI. Metrics carry their denominators. Practice companions, live captions, assessment data and programme operations AI, built into learning programmes and evidenced the way an education review asks for it. ## A tutor that talks is easy. Three things make it evidence Whether it corrected the right word. Whether a person saw the skill. Whether the number that reaches your board shows how much stands behind it. ### Did it correct the right word? anchored to characters 6 to 12 of her sentence - A correction that describes the mistake in a block underneath is a lecture. One that is anchored to the exact characters the learner wrote can be drawn on her own word, and can be checked: if the offsets no longer point at the original text, it is a model problem and it never reaches a learner metric. ### Did a person ever see the skill? Observed · past tense, in conversation with a person one question to the volunteer · about the skill, never the person - A practice system on its own only records the learner talking to a machine. One optional question to the volunteer after a session asks whether one observable thing happened. Not a score and not a judgement of the person. The third answer is "not enough context", which is never counted as a no. ### Can the number show its evidence? Reused the corrected form · 2 / 3 numerator and denominator · the rate is withheld below five - Every metric leaves with a numerator and a denominator. A rate is refused below a floor rather than printed as a percentage that moves twenty points on the next learner, and test traffic is excluded before the first reading. ## Start from the programme you run Four engagements · one programme at a time Practice, captions, assessment data and operations, each bought on its own and configured for one programme and one cohort. Four things are settled before the first learner is enrolled. - The programme, the language and the cohort - What the model may say, and where a person rules - Which controls apply, from FERPA to WCAG 2.2 AA ### Conversational practice systems - A practice partner that replies in character, corrects on the learner's own word, and knows when to stop. - A companion built on tool-use: one reply that asks one question, at most one correction anchored to the learner's characters with a one-line explanation, at most two vocabulary items in context, a recap when she has said enough, and voice in and out. Scenarios and pause states are designed, not improvised: on frustration or a jailbreak attempt the companion stops and hands the learner to a person, and it answers course questions only from approved material, with the source cited, or refuses. - **Reply**: In character · one question at a time - **Correction**: flag. Anchored to characters 6 to 12 - **Recap**: Offered after the third turn Drawn on her word · logged with its offsets ### Live captions and session transcripts - Captions in a tutoring call that both people agreed to, and a transcript that exists only if someone chose to keep it. - Live transcription in video sessions, started only when both participants allow it, with speaker attribution, captions in the session language, and a transcript saved only when a person decides to keep it after the call. Recordings are consented separately. - **Consent**: confirm. Both people, before the first caption - **Speakers**: Attributed, in the session language - **Transcript**: Kept only on a person's choice Captions on · nothing saved without a choice ### Assessment and evaluation data - Rubrics, rater agreement and held-out sets for the models your programme will trust. - Evaluation data for education models: rubric-scored responses, two-rater agreement measured on the delivered material, held-out test sets, native-speaker evaluation in the languages the model will meet, and red-team probing for the things an education model must not say to a learner. - **Raters**: confirm. Two per item · agreement measured - **Held out**: Kept out of training - **Languages**: Including Danish, Norwegian, Swedish, Finnish Agreement on the delivered set, not assumed ### Programme operations AI - The admin side of a learning programme, with a person holding every consequential decision. - Intake triage and cohort assignment with a person confirming, course provisioning into the LMS you already run, certificates and badges issued from the record, campaign copy drafted only behind an approval marker and never from rosters, insight narratives with a banned-word filter, and voice notes transcribed into the record. - **Triage**: flag. Proposed by the model · approved by staff - **LMS**: Course created in your system - **Campaigns**: confirm. Drafted only behind an approval marker Every consequential step has a person's name on it ## Six kinds of programme, each with its own proof Language · assessment · higher education · corporate · accessibility · K-12 A language app, an assessment vendor and a university do not ask the same question of the evidence. The engagement is written for the question your programme asks. From edu-AI start-ups to proctoring vendors, each engagement names the kind of programme it is written for. - Practice companions and tutoring tools, evidenced by anchored corrections, observed skills and numbers with denominators. - Rater agreement on the delivered set - Rubric data, two-rater agreement, held-out sets and red-team probing for what a model must not say to a learner. - Built into the system you run - Retrieval over course material, provisioning into the LMS, certificates from the record, and the documentation an institutional review asks for. - Skills evidence a manager can read - Workplace practice, session captions with consent, and impact metrics that show their denominators to the people who fund the programme. - WCAG 2.2 AA and Annex III controls - Assistive reading and captioning built to WCAG 2.2 AA, and the high-risk documentation an exam-integrity or proctoring deployment carries under the EU AI Act. - Publisher AI, district software and adaptive platforms built under Article 8, COPPA and FERPA, with age-appropriate design and a person at every consequential step. ## Write the sentence yourself Six rules · two languages · one person You are the learner. A toy companion of six rules replies, draws the correction on your own word, and a person is asked one question. Every number keeps its denominator. A toy companion with six rules, so the mechanics are visible. Not a language model. Nothing on this page is a real delivery or a real result. ### Practice · At work - I dag har gå jeg på møte med kollega - Jeg har et jobb på sykehuset - I går jeg snakket med sjefen om pause - I goed to work early today - I found an job at the clinic Want a recap of what we practised? ### In your conversation, did she use the past tense? Met the learner in a session this week The question opens after a correction has landed. Answer recorded · observation, never proof of cause Captions in a session only start when both people have allowed them. The transcript is saved only if someone chooses to keep it afterwards. Transcript kept · by a person's choice Tutor: Fortell meg om dagen din på jobb. The companion answers a question only from the passages the programme approved, cites the one it used, and refuses when nothing matches. A refusal hands the learner to a person. - Why is it "har gått" and not "har gå"? - What is the word order after a time word? - What is the capital of Peru? That is not in the course material. I will not guess. Your tutor can take this one. Handed to the tutor · logged - After "har" the verb takes the perfect form: har gått, har spist, har snakket. - When a sentence starts with a time word, the verb comes second and the subject third: I går snakket jeg. - Jobb is masculine: en jobb, jobben. Møte is neuter: et møte, møtet. Rates are withheld below five turns. ### The correction is drawn, not described. - characters 6 to 12 of her sentence - The companion returns offsets into the learner's own message. If the offsets no longer point at the original text, the correction is a model problem and never counts. ### The person is asked about the skill. - observed · past tense · a volunteer - One optional question after a session, about one observable thing. Not a score, and "not enough context" is not a no. ### The number keeps its denominator. - 1 / 3 · rate withheld - Every metric is a numerator over a denominator. A rate is refused below the floor, so three learners and one retained form is reported as 1 over 3, never as 33 percent. ## Every consequential step arrives with a name Demand · attendance · recap · certificate Demand becomes a class, absences become follow-up, a session becomes a recap and a certificate. The model proposes each step. You or a teacher confirms it, and the log keeps the name. A toy desk with fixed demand and a four-learner roster, so the mechanics are visible. No email is sent, no course is created. Nothing on this page is a real delivery or a real result. ### Programme desk · Autumn cohort Every consequential step carries a person's name Level 40 · slot 30 · waiting up to 30 Crisis tag pinned · shown, never scored Class created · course provisioned in your LMS · approved by you No teacher can take the largest demand yet. ### Amal ### Bao ### Ceren ### Dana "We missed you" · templated · sends only if the cohort opted in Personal check-in · a person writes it Phone or WhatsApp · a person calls Template suppressed · crisis tag pinned · a person decides Wording allowed · the tone rules hold Refused · the rules never allow The rules refuse any threat to a learner's place and any streaks, badges, points or leaderboards. A learner is paused, not dropped, and gets fewer emails, not more. ### Amal ### Bao ### Ceren ### Elif Draft · a teacher edits and publishes Published · by the teacher · learners see this text only Rule: at least 75 percent attendance, over at least eight sessions Certificate issued · signed by you · in the record Nothing yet. Propose a class to begin. ### The proposal shows its reasons. - Demand, teacher capacity and the size rule are written next to the proposal, and the candidate score is three visible parts. A crisis tag is shown and never scored. ### The tone rules are written down and enforced in the system. - Follow-up is tiered and the wording is checked against rules the programme wrote down. A pinned crisis tag suppresses the template and routes the learner to a named person. The model detects; it never intervenes. ### A person signs every consequence. - Class creation, check-in, recap and certificate each wait for a named confirmation, and the desk log records who did what. ## The ledger your funder and your review will read One row per turn · totals with denominators Every turn with its correction and its anchor, whether the corrected form came back, and what a person observed. Written by what you did above. Example ledger, built from what you ran above No turns yet. Write a sentence in the room above. Valid · offsets point at her text ## One programme, one language, one cohort A pilot runs the same engagement on one programme, one language and one cohort. The scenarios, the controls and the ledger's fields are fixed before the first learner is enrolled, one cohort runs with a person at every consequential step, and the pilot ends with a ledger and a decision. You judge it against your programme and your review. - What the companion may say, and where a person rules - The controls that apply, from consent to WCAG 2.2 AA - The floor below which no rate is reported - What the evidence ledger must contain - The situations the companion practises, the corrections it may make, and the pause states, written and agreed. - Consent, age and data controls mapped to the programme, with the person named at each step. - One cohort runs, with captions and observations gated on people, and test traffic excluded. - Every turn, correction, anchor, observation and metric, with denominators, in the delivery. - Continue, change or stop, on your own criteria, with the ledger as the record. The pilot ends with a ledger your review can read row by row. ## The controls a review of learner data will ask for GDPR Article 8 · COPPA · FERPA · Annex III A Norwegian company under GDPR. Learner data stays in Europe by default, is handled for the learner's age under Article 8, and carries the documentation an Annex III education deployment needs. US programmes run under FERPA and COPPA. - **Jurisdiction**: Norway · GDPR-native - **Children**: GDPR Article 8 handling · COPPA where it applies - **Records**: FERPA controls for US educational records - **EU AI Act**: Annex III education documentation - **Accessibility**: WCAG 2.2 AA at code level · no overlay widgets - **Transparency**: AI-drafted text labelled to learners · Article 50 - **Storage**: EEA by default - **Consent**: Captions and recordings consented separately - **Contracts**: Standard DPA terms · SCCs available - **Erasure**: 30-day end-of-contract SLA - [The AI Implementation service line](https://ypai.ai/ai-implementation/) - [AI Data & Evaluation](https://ypai.ai/data-solutions/) - [Human and model evaluation](https://ypai.ai/data-solutions/evaluation/) ## Tell us the programme A feasibility read · a bounded first cohort We scope against the programme as it runs today, agree what the model may say and where a person rules, and prove the method on a bounded first cohort. Scope · Scenarios · Controls · First cohort · Ledger A bounded first cohort carries its own acceptance criteria and its own ledger. You judge it against your programme. Briefs are treated as confidential. We are used to programmes that cannot be named and learner data that cannot leave the EEA. Tell us the programme, the learners and what the model will do for them. We reply with a feasibility read. We come back with a feasibility read on the programme, the learners and the controls. The brief is treated as confidential.