Teachback

Inspiration

It is easy to repeat a lesson in familiar words and still miss the idea that makes it work. A learner might say that two percentage changes cancel, then be surprised when the second change uses a different base. Teachback begins at that moment of uncertainty. We wanted a place where an explanation leads to a prediction, useful feedback, and another chance to apply the idea in a fresh situation.

What it does

Teachback offers four authored quantitative lessons in which a learner explains an idea, predicts a result, inspects quantities, and tries a fresh application. The new local Simulated Alexa+ text assistant is connected to the changing-percentages lesson. It can inspect the current explanation and prediction, compute the relevant quantities, show the supporting source paragraphs and diagram, explore a supported what-if change, and propose a worked response for explicit approval. Approved assistance enters the same ordinary lesson record and is labeled as assistance.

How we built it

Node.js and browser JavaScript serve the authored passages, numeric checks, diagrams, hypothetical calculator, transfer tasks, and opt-in saved history. A deterministic interpreter plans bounded calls into those existing learning functions. Reviewed writes use the original lesson submission and application functions, while a new-task proposal requires explicit approval before opening blank native fields. MediaRecorder still supports an optional private spoken draft. No model service or Bee account is needed for the local course.

Challenges

Worked assistance must remain distinguishable from a learner’s independent answer. The assistant therefore shows source paragraphs and computed quantities before approval and retires the previous conversation when a fresh task opens. Supported changing-percentage questions are handled by the bounded interpreter; the other lessons retain their existing authored checks. General semantic grading and claims of mastery would require separate evidence.

Accomplishments

In actual browser receiving, the assistant handled a same/100 misconception by computing 96 and identifying 120 as the second base. Changing the first increase from 35 percent to 50 percent changed the calculated quantities from 135 and 108 to 150 and 120. A reviewed fresh tank task opened blank native fields, and an approved worked application was saved through the ordinary record. Reload restored both assisted attempts. Unsupported email or upload requests offered no approval. The suite passed 173 native tests.

What we learned

Useful help should expose the quantity that changes the conclusion and make its source easy to inspect. An approved worked example can support study, but it must not be counted as an independent learning result. Keeping assistance in the same lesson history helps the learner and reviewer understand what happened.

What's next

We want to extend the assistant carefully across the other authored lessons, study how learners use its help, and receive authorized source material. The current assistant is a bounded local text simulation. Live Alexa service use, actual Bee data, general semantic grading, and independent learning outcomes remain unestablished.

Built With

• agentic-tool-workflow
• css
• html
• javascript
• localstorage
• mediarecorder
• node.js
• web-locks-api