Matched ½ and 2/4 using a fraction model. The drawing supports the equivalence, but does not show whether help was needed.
Teacher Copilot & Assessment
A child explains a fraction, changes their mind, draws a model and tries again. There is more to learn from that sequence than from the final answer alone. bina is building a teacher copilot that follows the lesson, examines the work it produces and brings the evidence back to the person teaching it. AI assembles a case for what a learner understands. A teacher decides whether that case holds, and what to teach next.
The context is in the lesson, not just the answer.
A transcript might say “they’re the same”. On its own, that tells us very little. Which question was the child answering? Were they comparing two fractions or repeating a classmate? Did the teacher supply the reasoning a moment earlier? A useful assessment needs the exchange around the answer, not merely a sentence extracted from it.
The copilot is being designed to gather that context during the lesson as transcript material becomes available. AI would follow the question, the learner’s explanation, the prompts and the revisions, building an evolving set of observations. A teacher could mark an important moment or add a quick voice or text note: “This was independent”, or “I showed the first step”. Those short additions can change the meaning of everything around them.
During-lesson context and a finished assessment are different stages. Transcript fragments can arrive late, words can be corrected and a speaker can be uncertain. The working interpretation stays provisional; a post-lesson pass reconciles the fuller transcript, the teacher’s notes and the work submitted. This is the intended workflow, with no fixed live-transcript latency promised.
From what happened to what it means.
The same lesson can leave several kinds of evidence. The interesting work is bringing them together without flattening the differences: what the child said, what they made, what the assessment asked and how much help they had.
Follow the lesson
Transcript excerpts, questions, prompts and marked moments build context as the lesson unfolds.
Read the work
Writing, drawings, worked solutions and assessment responses add evidence the conversation may not contain.
Test an interpretation
AI relates the evidence to a learning goal, keeping supporting passages, contradictions and gaps beside the suggestion.
Ask the teacher
Confirm, correct or send back for more evidence. The teacher’s decision is the assessment checkpoint.
Choose what comes next
A fresh challenge, a different explanation or another observation follows the reviewed interpretation.
Carry the reasoning forward
Reporting draws on reviewed assessments; the next lesson can test the questions that remain open.
Try the teacher’s side of the decision.
A fictional fraction lesson. The suggested assessment deliberately leaves room for judgment: was the explanation independent, or did a prompt make the difference?
Illustrative interaction · synthetic evidence · nothing is saved
1. Evidence to review
Teacher: “What happens if we split each half again?” Learner: “There are four pieces. Two still cover the same part.”
“Confident with the model. Check whether the idea transfers without a prompt.”
2. Copilot suggestion
Learning goal: explain equivalent fractions
Demonstrated independently.
Check the marked moment: the explanation followed a teacher prompt.
3. Teacher-reviewed outcome
Awaiting teacher review. No assessment has been confirmed.
Next step: decide what the evidence supports before choosing an activity.
What the copilot is made of.
These are the working parts of the assessment direction, rather than six separate products. Open each one to see what it contributes, and the question it has to answer before its output becomes useful.
Lesson contextFollow the exchange, including the help that made an answer possible.
AI analysis would group relevant transcript passages into an educational episode: the task, the question, a response and any follow-up. Timestamps let the teacher return to the exchange. Uncertain speaker attribution stays visible rather than attaching an achievement to the wrong learner. A pause or a quiet lesson is not a learning deficit; attendance and speaking time are context, not measures of mastery. The aim is to understand an explanation, not infer a child’s emotions from their behavior.
Student-created materialExamine the reasoning in the work, not only whether the answer matches.
A written paragraph, a drawing, a photographed worksheet and a worked solution each need a different reading. AI could identify a claim and its supporting evidence, a mathematical step or a revision between attempts. The task and any rubric travel with the work so an observation answers the question that was actually set. An unclear image or ambiguous handwriting should produce a request to check the source, not a confident assessment of something the system could not read.
Assessment contextBring formal results into the conversation without letting one score settle it.
Classroom assessments and standardized results, including STAR where relevant, offer another view. AI would help interpret their relationship to the learning goals alongside lesson evidence and student work. The level assessed, date, task conditions and assistance matter: a result from one setting does not automatically explain performance in another. If the sources disagree, that disagreement becomes a question for review. It may suggest a new task, a missing prerequisite or simply that more evidence is needed.
Standards-linked hypothesesSay which learning goal the evidence supports, and show why.
The standards map gives the interpretation somewhere precise to land. A suggestion would identify the relevant goal and the evidence for it, with any supporting conditions or unresolved contradiction. Mentioning fractions is not evidence of understanding equivalent fractions; answering with a model is not necessarily evidence of transferring the idea without one. The design calls for tentative hypotheses with source references, rather than turning every matching keyword into a completed standard.
Teacher reviewA short, specific decision while the lesson is still familiar.
The teacher sees the proposed assessment beside the material behind it. They can accept it, change the interpretation, correct the learner or goal involved, or ask for another observation. A voice note should be a natural way to contribute, not another long form to complete. Review effort is a hard design constraint: a copilot that generates more checking than it saves has missed its purpose. We need to learn which suggestions are useful enough to deserve a teacher’s attention.
Teaching and reportingUse the reviewed judgment in the next lesson and in the account of progress.
A confirmed assessment could suggest a suitable next activity, surface a prerequisite to revisit or help prepare a progress report. The teacher chooses the response. The same reasoning should remain available when a family asks how a conclusion was reached. A later lesson can challenge an earlier interpretation; progress is a developing account of what a learner can do, not a collection of permanent labels. Placement changes remain a separate school decision with their own review.
Different tools answer different assessment questions.
A score can be exactly what a teacher needs for a bounded task. A lesson summary can be useful for remembering what happened. The copilot’s additional job is to connect evidence to a judgment that somebody can inspect. Each approach comes with a different cost.
| Question | Task scoring | Lesson summary | Teacher copilot |
|---|---|---|---|
| What does it establish? | How a response compares with the task’s answer or rubric. | What was discussed and which moments stood out. | A proposed learning-goal assessment across relevant sources. |
| What context matters? | The task, response and marking criteria. | The conversation and lesson structure. | Those sources together, including prompts, work and teacher observations. |
| Where is the uncertainty? | In the rubric, scoring method and conditions of the attempt. | In transcription, attribution and what the summary leaves out. | Beside the hypothesis, with gaps and contradictory evidence for review. |
| What happens next? | The result informs the teacher’s broader judgment. | The teacher reconstructs its meaning for each learner. | The teacher confirms or corrects a suggestion and chooses a response. |
| What is the trade-off? | A clear result with a bounded view of learning. | Less material to read, but interpretation still falls to the teacher. | Richer interpretation, with more integration, model evaluation and review to get right. |
The checkpoints matter as much as the model.
Before an observation can support a judgment, the relevant passage or work must be readable, attached to the right learner and understood in relation to its task. Before a suggestion can be confirmed, the teacher needs to see its reasoning. Before it shapes a report, the reviewed assessment must be distinguishable from the machine’s original proposal.
Those boundaries make the system testable. We can ask whether an observation matches its source, whether the suggested goal is relevant, whether a teacher had to correct the interpretation and whether the review was worth the time. A model’s fluent explanation is not enough. Useful evaluation needs examples with teacher judgments, disagreement cases and a way to compare successive versions against the same material.
The data lake supplies the underlying records and their provenance. The copilot owns the interpretation and the review experience. Keeping those responsibilities separate lets the assessment approach improve without making the school’s evidence depend on one model’s account of it.
A better memory for teaching.
The long-term value is not a faster mark book. It is being able to pick up a learner’s story where the last lesson left it: the explanation that became clear, the prompt still needed, the idea ready to be tested in a new setting. Teachers already work with this knowledge. The opportunity is to make more of it available when they need it, without asking them to reconstruct every lesson from memory.
That is also a demanding product test. Does the copilot help a teacher notice something useful, reach a sound judgment with less preparation, or choose a better next question? Those are the gains worth demonstrating before claiming a change in learning outcomes. The technology earns its place through the teaching it helps make possible.
In development. The diagrams and interactive example describe the assessment direction. They use fictional material. During-lesson context gathering, multisource analysis and teacher-reviewed assessment are being developed as one workflow; this page is not a demonstration of a deployed live service.