Written vision
Belay: A Vision for an Agentic Classroom Partner
Summary
Belay is an agentic classroom partner that watches student work between the moments a teacher can look, so the teacher spends her attention on the students who need it most. It reads every submission for more than the answers, drafts what the teacher would otherwise do by hand, surfaces only the exceptions, and follows up to check whether an intervention worked.
The name comes from climbing. A belayer holds the rope while the climber climbs. The belayer does not choose the route and does not pull the climber up. They pay attention, take in slack, and catch a fall before it becomes one. Belay plays that role for a teacher and a class: it supports, and the teacher decides.
This document describes the problem Belay addresses, the loop it runs, what it does on its own versus what it leaves to the teacher, the guardrails that keep it in its lane, and what the proof of concept shows. It is written for a school leader or product sponsor who has not seen the prototype. Every person, class, and number in it is invented.
The problem, from both seats
A secondary school mathematics teacher cannot look closely at every piece of work from every student, so most of it gets a glance and the students who need a closer look get the same glance as everyone else.
From the teacher's seat. Mdm Tan teaches Secondary 3 Mathematics in four classes, 108 students. A single homework set produces 108 submissions, some typed, many photographed from a notebook. Grading them takes an evening. Noticing patterns across them, such as the six students in one class who all missed the same idea, takes a second evening she does not have. Noticing patterns across weeks, such as a student whose work has been arriving later and getting shorter, is nearly impossible by hand. She knows the signals are in the stack. She has no time to find them.
From the student's seat. Jia Yi is capable and rushes. She finishes fast, skips the last two questions, and never hears about it until the grade comes back. Danish submits late, his handwriting is hard to read, and he is quietly falling behind. Nobody is ignoring Danish. He is simply not the loudest problem in a room of thirty, and by the time his grades make him visible, the gap is weeks old.
The two seats share one failure. The information that would help both of them exists in the work itself, but it arrives faster than any person can read it. Attention, not data, is the scarce resource. Belay exists to spend that attention well.
The vision
Belay is an agent, not a dashboard. The difference is that it acts between the moments the teacher looks.
A dashboard waits to be opened. Belay watches submissions as they arrive, reads each one, drafts feedback, compares the work to what came before, notices who is slipping and who is stuck on the same idea, and puts a short list in front of the teacher with a reason and a suggested next step. When the teacher chooses an intervention, Belay records it, watches the student's next few pieces of work, and reports back whether it helped.
The teacher remains the decision maker at every point that matters. Belay drafts grades; the teacher confirms them. Belay suggests a conversation with Danish; the teacher decides whether and how. Belay proposes a short practice set for the 3E3 cluster; the teacher assigns it or does something better. Nothing reaches a student or a gradebook without a teacher's say.
Students get a coach on the same rope. Belay reminds them before a deadline, flags a blank question before they press submit, explains the feedback they receive in plain language, and offers a short practice set when it notices a gap. For Jia Yi, that is a nudge to slow down before she loses points she knows how to earn. For Danish, that is a check-in before he falls far enough to be noticed the hard way.
In one line: Belay holds the rope so the teacher can teach and the student can climb.
A day in the life: Mdm Tan
Monday, October 12. Mdm Tan sets Laws of Indices Practice, the Chapter 4 assignment: twelve questions, due Thursday at 11:59 PM, with a late window to Friday at 8 AM. Belay asks what she wants it to pay attention to beyond the answers. She picks three things: every question attempted, work shown on questions 8 through 12, and submission on time. Belay turns that into a checklist it will apply to all 108 submissions.
Tuesday to Thursday. Students submit. Some type their answers. Thirty-five photograph notebook pages, 71 pages in all. Belay reads each one as it arrives, transcribes the handwriting, checks it against the checklist, drafts a provisional grade and two sentences of feedback, and compares the work to each student's recent history. On Wednesday at 7 PM it nudges the 23 students who have not started, and 19 of them submit on time. Mdm Tan does not open anything during these three days.
Friday morning. She opens Belay for fifteen minutes. She does not see 105 submissions. She sees three things that need her today, three non-submitters Belay has already nudged, and 97 students who are quiet because nothing is wrong.
- Class 3E3: six students (Wei Jie L., Zhi Hao O., Hafiz A., Kavya N., Shu Min K., Nur Aisyah B.) wrote 2^-3 = -8 and 10^-2 = -100 on questions 6, 7, 10, and 11. A relief teacher took the October 8 lesson. Belay suggests a ten-minute small-group reteach and offers a four-question practice set.
- Danish bin Rahman, 3E5: two photos at 12:47 AM Friday, 48 minutes late, and late on four of his last five. Legibility has fallen from 81 to 62 and accuracy from 88% to 70%. Belay flags him as quietly slipping and suggests a check-in.
- Tan Jia Yi, 3E1, low priority: typed Wednesday at 4:12 PM, nine minutes on task against a class median of 26, last question blank on three of her last four. Belay flagged two blanks before she submitted and she fixed one. Belay is already coaching her with a show-work reminder and suggests letting that run.
- Grades: 105 drafts, 98 confident, 7 held for her eye. Forty-one blanks across 17 students, already noted in the feedback.
She approves the reteach and assigns the practice set. She marks Danish for a check-in and leaves Jia Yi to Belay's coaching. She looks at the seven held grades, changes two, and confirms the rest. Fifteen minutes, against the two hours grading used to take.
Monday and Tuesday. She catches Danish at lunch on Monday. His phone camera is cracked, which explains the photos, and he has been watching a younger sibling after a parent's shift, which explains the hour. She tells him to type his answers from now on. On Tuesday she spends ten minutes with the six 3E3 students on negative indices. She tells Belay the outcome of both in one line each.
The following week. On Laws of Indices II, due October 22, Belay reports that five of the six 3E3 students got the negative indices items right, up from none. Nur Aisyah B. is still stuck, and Belay suggests a one-on-one on Monday. Danish submitted on time, Thursday at 8:40 PM, nine of twelve correct, and Belay says it will keep watching before calling the trend reversed. Jia Yi answered every question and spent 22 minutes on it. Mdm Tan did not have to remember to check any of this.
A day in the life: Danish
Thursday, October 15, 11:20 PM. Danish gets his younger sibling to bed after his parent's shift and opens Laws of Indices Practice on his phone. Belay has sent one reminder that afternoon: due at 11:59 PM, twelve questions, show work on 8 through 12. No lecture, no red badge.
Friday, 12:47 AM. He works through it in his notebook and photographs the two pages with a cracked camera. It is 48 minutes past the deadline and inside the late window Mdm Tan set, and Belay says so plainly, without a penalty it has no authority to set. Belay cannot make out his work on questions 8 and 11 and asks him to confirm what he wrote, rather than guessing or marking them wrong.
Friday afternoon. His feedback arrives. Belay has transcribed his handwriting next to the photo so he can see what the teacher saw. The grade Mdm Tan confirmed is there, with two sentences on what went well and one on where he lost the thread on question 11. He taps the sentence and Belay explains negative indices in a different way from the textbook, then offers three practice problems. He does two.
Monday, lunch. Mdm Tan asks him to come by. She knows four of his last five submissions were late and that his handwriting has been getting harder to read, and she asks what is going on, not why he is lazy. He tells her about the camera and the sibling. She says he can type his answers from now on. Danish did not know Belay had noticed the pattern. He only knows his teacher did.
The next assignment. Belay reminds him a day earlier, because he asked it to. He types his answers and submits Thursday at 8:40 PM, on time, nine of twelve correct. Nobody announces this. Belay notes it, and when Mdm Tan asks how Danish is doing, the answer is there.
The loop: what Belay does on its own, what the teacher decides
Belay runs a six-stage loop for every assignment. At each stage it acts on its own up to a clear line, and the teacher owns everything past that line.
Stage 5 is the one moment the teacher has to be present. Everything before it is Belay working ahead of her; everything after it is Belay carrying out and checking what she chose.
| Stage | Belay does on its own | The teacher decides |
|---|---|---|
| 1. Set the assignment | Suggests what to check for beyond answers, turns it into a checklist, estimates time to complete | What the assignment is, what matters, the deadline |
| 2. Students submit | Accepts typed or photographed work, transcribes handwriting, reminds before the deadline, flags a blank question before submit | Whether late work is accepted and on what terms |
| 3. Analyze | Checks timeliness, completeness, legibility, work shown, and trends against each student's history and the class | Nothing yet; this stage is Belay's alone |
| 4. Draft grades and feedback | Writes a provisional grade and short feedback per submission, marks the ones it is unsure about | Every grade; which feedback goes to students, in what words |
| 5. Surface exceptions | Shows only the students and clusters who need attention, with the reason and a suggested next step | Who actually needs attention and what the right step is |
| 6. Follow through | Records the intervention, watches the next few submissions, reports whether it worked | Whether to intervene, how, and whether the outcome is good enough |
Stage 6 feeds stage 1. What Belay learned about Danish, Jia Yi, and the 3E3 cluster shapes what it suggests checking for on the next assignment, and the loop runs again.
Looking beyond the answers
Most of what a teacher learns from a stack of homework is not in whether the answers are right. Belay reads for the rest of it.
| Dimension | What Belay reads | Why it matters |
|---|---|---|
| Timeliness | When the work arrived relative to the deadline, and the trend across assignments | One late submission is noise. Four of the last five, each with photos harder to read, is Danish, and it is a conversation about his evenings, not about math. |
| Completeness | Whether every question was attempted, and where in the assignment the blanks fall | The last question blank on three of the last four assignments is Jia Yi rushing, not Jia Yi struggling. The fix is different. |
| Legibility and neatness | How readable the handwriting is, and whether it has changed | Faint or deteriorating handwriting can mean a rushed student, a tired one, or a cracked phone camera. It is a prompt to look, not a judgment. |
| Work shown | Whether the steps are there on multi-step problems, not only the final line | A right answer with no work and a wrong answer with good work call for opposite responses. |
| Correctness | Which questions are right, wrong, or partly right, and the kind of error | This is the traditional grade. Belay drafts it and treats it as one signal among several. |
| Trends and clusters | How this submission compares with the student's own history, and which students share the same error | A student slipping against their own baseline, and six students missing one idea the same way, are the two things a teacher most wants to know and is least able to see. |
The point of the extra dimensions is not more data for its own sake. Each one is a thing Mdm Tan would notice if she had an hour per student, phrased so she can act on it in a minute.
Guardrails
An agent that acts on its own needs clear limits, and Belay's are part of the product, not an afterthought.
Grading is support, never the headline. Belay drafts grades because it saves the teacher an evening, and it marks the ones it is unsure about. It never publishes a grade. The product is built around the exception list and the follow-through, and grading sits behind them on purpose. If a demo of Belay leads with auto-grading, the demo is wrong.
The teacher's judgment comes first. Every suggestion is a suggestion. The teacher can dismiss a flag, change a grade, ignore a suggested step, or tell Belay it read a situation wrong, and Belay learns from that for this class. When Belay and the teacher disagree, the teacher is right by definition.
Student dignity. Belay describes work, not children. It says a submission was 48 minutes late, not that a student is lazy. Flags go to the teacher, never to the class or to other students. A student sees their own feedback, their own reminders, and their own practice, and nothing about anyone else. Handwriting legibility is a prompt for the teacher to look closer, never a mark against the student.
Privacy. Student work stays within the school's systems and is used to help that student and that class. It is not used to train general models and is not shared outside the school. The teacher can see everything Belay holds about a student, and so can a parent on request.
Transparency about what the agent did. Every action Belay takes on its own is logged in plain language: what it read, what it noticed, what it drafted, what it sent to a student, and why. A flag always comes with its reason. A teacher should never be surprised to learn what Belay did, and a student should never wonder whether it was a person or Belay who sent a reminder.
What the proof of concept demonstrates
The proof of concept has three parts built from one storyline, so a viewer can meet Mdm Tan, Jia Yi, Danish, and the 3E3 cluster in any of them and recognize them in the others.
- A clickable prototype. The teacher's side of the loop, from setting the indices assignment to confirming grades, acting on the three exceptions, and reading the follow-through report a week later. The student's side as Danish sees it: the reminder, the blank-question check, the feedback with transcription, and the practice set.
- A dynamic slide story. The same week told as a narrative for a room, with the exception list and the follow-through as the two moments the story turns on.
- This written vision. The argument, the loop, the guardrails, and the path forward, for someone who will read before they click.
Together they are meant to show five things:
- Belay reads photographed handwritten work as readily as typed work, so it fits how students actually do math.
- Belay reads for timeliness, completeness, legibility, work shown, and trends, not only correctness, and each of those produces something a teacher can act on.
- A class of 108 students reduces to three items that need the teacher today, each with a reason and a suggested step, and grading takes fifteen minutes instead of two hours.
- An intervention is recorded, followed, and reported on without the teacher having to remember to check.
- Grading is present and useful, and it is visibly not the point.
What the proof of concept does not demonstrate: integration with a real gradebook or learning platform, accuracy of handwriting transcription on real student work at scale, or how the exception thresholds should be tuned for a given class. Those belong to the pilot.
From proof of concept to classroom pilot
The shortest honest path from this proof of concept to a classroom is one teacher, one subject, one semester.
- Choose the teacher. One secondary school Mathematics teacher with several classes, who already collects homework and is willing to be a design partner. Mdm Tan is the model. The loop needs volume to show its value, so a teacher with one small class is the wrong first pilot.
- Connect the submissions. Students submit through whatever the school already uses, plus a phone camera for handwritten pages. Belay reads from there. No new device, no new login for students beyond what exists.
- Run the loop read-only for two weeks. Belay analyzes, drafts, and flags, and the teacher sees all of it, but nothing reaches a student and no grade is confirmed through Belay. This calibrates the exception thresholds against the teacher's judgment and checks transcription accuracy on real handwriting.
- Turn on the teacher's side. The teacher confirms grades, acts on exceptions, and records follow-through in Belay. Measure the time spent per assignment against the two weeks before, and how often the teacher overrides a draft or dismisses a flag.
- Turn on the student's side. Reminders, the blank-question check, feedback explanations, and practice sets, with parent notice beforehand. Measure on-time rate, completeness, and whether students use the explanations.
- Decide at the semester's end. The questions that matter: did the teacher get time back, did any student get noticed sooner than they otherwise would have, did any follow-through demonstrably work, and did the teacher trust it. A pilot that returns the teacher's evenings but misses a student like Danish has failed.
What is needed to start: a school willing to host, a data agreement covering student work and the privacy commitments above, and a product team to build the integration with the school's existing systems. The analysis, drafting, and follow-through shown in the proof of concept are the parts that already exist in principle; the integration and the calibration are the work.