How it works

From expression to confirmed meaning.

One continuous path connects camera input, recognition candidates, academic-risk detection, repair, confirmation, model routing, answer checks, and separately authorized learning.

Explore the product ↗Inspect evidence →

02 / The sequence

Sign → Recognize → Detect risk → Repair → Confirm → Route → Check → Learn.

02 / The complete control path

Eight moves.
One protected meaning.

Scroll the system from expression to answer. Every scene is a concept visualization paired with its evidence boundary.

CONCEPT / 01Expression enters

01 · Expression enters

A recognizer proposes—not decides.

01Sign02Recognize

The camera or another approved input can produce a candidate. Sign First keeps that candidate separate from confirmed meaning, so recognition providers can change without changing who has authority.

Concept visualization · no recognition-accuracy claim

02 · Risk becomes visible

The smallest important uncertainty gets attention.

03Detect risk

Numbers, operations, negation, and other meaning-changing differences can trigger a pause. The 3 × 5 versus 3 + 5 scene is an illustrative academic-risk pair—not a documented error from Google or another provider.

Illustrative critical pair · not a measured model result

03 · The learner repairs

Choice replaces a hidden correction.

04Repair05Confirm

The repair surface is designed to become age- and reading-aware while preserving the same rule: the system can suggest, but the learner controls what the question means. Child effectiveness has not yet been validated.

Interface concept · child outcomes not yet validated

04 · Confirmed meaning moves

AI answers the question the learner approved.

06Route07Check

Only confirmed text crosses the answer boundary. A separate answer-assurance check can then look for drift between that intent and the response while the signer retains a visible override.

Control-path concept · no effectiveness guarantee

05 · Authorized evidence returns

Learning is a governed choice—not a hidden exhaust stream.

08Learn

Ordinary use stays private. Separately authorized, minimized repair evidence can enter an offline, versioned learning loop and must disappear from rebuilt assets after revocation. The verified loop uses synthetic text repairs; real-learner visual contribution remains disabled.

Synthetic governed-learning lifecycle · not student data or recognition improvement

01–03 · Access

The camera starts a candidate—not a conclusion.

A recognizer observes the signing and proposes meaning. Structural cues and multiple candidates may reveal uncertainty without pretending to understand what has not been reliably identified.

04–06 · Authority

Repair only what could change the answer.

The learner sees a focused clarification, edits when necessary, and confirms the complete intent. Blanket confirmation is the current scaffold; selective intervention is the evidence-gated direction.

07–08 · Assurance

Route confirmed intent. Check the response.

Confirmed academic intent can be sent to the chosen model. Optional learning signals remain private by default and separate from model contribution.

Continue exploring

Choose the view that matches your role.

Students→Families→Educators→