Belay

Architecture vision

How Belay holds the rope

Belay is built as a set of small agents that work while submissions arrive, an orchestrator that decides which of their proposals can go ahead on their own and which must wait for the teacher, and a per-student memory that lets every new piece of work be read against what came before. Nothing in the design waits for a teacher to open a dashboard. Everything that changes a grade or a student's day waits for a teacher to say yes.

Calvin · Every person, class and number on this page is invented, drawn from the Mdm Tan storyline.

Four decisions that shape the system

Event-driven

Each submission, deadline and missed deadline is an event that wakes the right agent. Belay acts between the moments the teacher looks.

Agents propose, the orchestrator gates

Agents never write to a gradebook or decide an intervention. They hand proposals to one approval gate the teacher controls.

Memory per student, owned by the school

A history ledger per student is what turns one late photo into "late on four of the last five". It stays in the school's systems.

Grading sits behind attention

The draft grader is one agent of six. Its output feeds a review queue, never the front page of the teacher console.

The system, layer by layer

Two surfaces sit on top: the teacher console Mdm Tan opens once the work is in, and the student app Jia Yi and Danish use on their phones. Under them, the orchestrator routes work to six agents, and every agent reads and writes the same memory. Integrations bring in rosters, timetables and submissions from the systems the school already runs. Privacy rules wrap every layer below the surfaces.

PRIVACY AND GOVERNANCE · PDPA CONSENT · DATA MINIMISATION · SCHOOL-HOSTED SURFACES Teacher console Mdm Tan · web · opened once the work is in Attention view Follow-through Grade review Set assignment Student app and coach Jia Yi, Danish and 106 others · phone first Type or photo Blank check Explain feedback Practice set approvals, decisions flags with reasons, drafts work, questions nudges, feedback Belay orchestrator Turns events into agent work, and holds every proposal until the right person says yes Event stream Scheduler: deadlines, nudges Policy engine Approval gate Action log teacher owns the gate dispatches work returns proposals Intake reader Photo to transcript Legibility score Asks to confirm 71 pages read Signals analyst Late, blank, legible Work shown, trend vs own history every submission Draft grader Provisional grade Two-line feedback Holds unsure ones 98 sure · 7 held Attention agent Ranks exceptions Groups shared errors Reason, next step 108 → 3 items Follow-through Records the step Watches next work Reports outcome 5 of 6 fixed Student coach Deadline nudges Blank-question check Practice sets 19 of 23 on time Model layer · Reasoning model: handwriting, judgement, clusters · Fast model: high-volume checks, nudges · chosen per agent read and write history Student history ledger Signals, flags, steps, outcomes one timeline per student Assignment and rubric store Questions, checks, deadlines late window, answer keys Work vault Photos, transcripts, typed work photos purged on school schedule Action log Every agent action, in plain words append-only, teacher can read all rosters, timetable, submissions in confirmed grades, notices out SLS or school LMS assignments, submissions Student info system rosters, classes, contacts Timetable and relief who taught which lesson School sign-in roles, class access Messaging app push, parent notices
Proposals travel up through one approval gate in the orchestrator. Agents never talk to the surfaces directly, which is what makes the action log complete. The dashed draft grader is deliberately one box among six.

The loop, and where the teacher decides

One assignment, Laws of Indices Practice, followed through all six stages. The middle lane is what Belay does with nobody watching. The orange marks are the four points where the system stops until Mdm Tan decides.

Stage
1. Setteacher starts
2. Submitup to the deadline
3. Analyseas work arrives
4. Draft gradesas work arrives
5. Surfaceafter the deadline
6. Follow throughon the next assignment
Students
Sees 12 questions, show work on Q8 to Q12
Type answers or photograph notebook pages: 35 students, 71 pages
Nothing to do
Nothing yet; no grade is visible
Kavya N. and five classmates get a 4-question practice set
Danish types his answers; Jia Yi answers all 12
Belay on its own
Turns three chosen checks into a checklist for 108 students; schedules nudges
Transcribes photos; flags blanks before submit; nudges 23 before the deadline, 19 then on time
Reads timeliness, completeness, legibility, work shown and trend against each history
105 drafts with feedback; 98 confident, 7 held for review
Reduces 108 to 3 items with reasons; handles 3 non-submitters; 97 quiet
Records each step, reads Laws of Indices II, reports 5 of 6 fixed
Teacher decides
What to set, what matters, the deadline gate
Sets the late window once
Nothing; this stage is Belay's alone
Changes 2 of the 7 held grades, confirms the rest gate
Approves the 3E3 reteach, marks Danish for a check-in, leaves Jia Yi to coaching gate
Judges the outcome; books a 1:1 for Nur Aisyah B. gate
Runs without the teachergateStops until the teacher decides
Stage 6 feeds stage 1: what Belay learned about the 3E3 cluster shapes the checks it suggests for the next assignment.

One submission, end to end

Danish's late photos show how the layers cooperate. No single reading of his work says much. The ledger is what turns it into a pattern worth a moment of Mdm Tan's attention.

Danish R. 3E5 · 2 photos after deadline cracked camera Intake reader 2 pages read legibility 81 → 62 Q8, Q11 unclear Signals analyst submitted late in late window accuracy 70% History ledger late 4 of last 5 accuracy 88% → 70% over 5 assignments Attention agent quietly slipping next step: check-in Teacher console Mdm Tan marks a check-in talks with Danish photos text signals trend gate flag Student coach asks Danish to confirm Q8 and Q11 unclear confirm? Action log "Flagged Danish R.: late 4 of 5, legibility 81 → 62" logs
Belay asks Danish about two unclear answers instead of marking them wrong, and describes his work rather than him. Only the teacher learns the trend, and she decides what to do with it.

The 3E3 cluster takes the same route with one extra hop. The attention agent groups six submissions that all wrote 2-3 = -8 and 10-2 = -100 on Q6, Q7, Q10 and Q11, then checks the timetable feed and finds that a relief teacher took the lesson that introduced negative indices. The reason arrives with the flag, so the suggested step is a reteach, not six separate remedial notes.

The six agents

Each agent has a narrow job, a short list of things it may do on its own, and a clear hand-off. The model behind each is a choice per agent, so a cheap fast model can run the volume work and a stronger one can do the judging.

Intake reader

Reasoning model · image input

Reads
Typed answers, notebook photos, retakes
On its own
Transcribes handwriting next to the photo, scores legibility, asks for a retake or a confirmation when unsure
Hands off
A transcript per question, with confidence, to the signals analyst

35 students, 71 pages, 4 retakes requested. It never guesses an unclear answer.

Signals analyst

Fast model · structured output

Reads
The transcript, the assignment's checklist, the student's ledger
On its own
Measures timeliness, completeness, work shown, legibility and time on task, and compares each with the student's own baseline
Hands off
A signal record per submission to the ledger

Jia Yi: about a third of the class median time on task, Q12 blank.

Draft grader

Reasoning model · supporting role

Reads
Transcript, answer key, rubric for shown work
On its own
Writes a provisional grade and two lines of feedback; marks the ones it is unsure about
Hands off
Every grade to the review queue. It cannot publish.

105 drafts, 98 confident, 7 held. The teacher reviews 7 closely, not 105.

Attention agent

Reasoning model · clustering and ranking

Reads
All signal records for the assignment, every ledger, the timetable
On its own
Groups shared errors, ranks students against their own trend, writes a reason and a suggested step for each item
Hands off
A short list to the attention view, through the gate

108 students became 3 items: the 3E3 cluster, Danish, and Jia Yi at low priority.

Follow-through

Fast model checks · reasoning model reports

Reads
The step the teacher chose, the next few submissions
On its own
Records the intervention and its outcome note, watches later work, writes a short before-and-after
Hands off
An outcome report and, if needed, a next suggestion

3E3 went from 0% to 83% on negative indices. Nur Aisyah B. still stuck; 1:1 suggested.

Student coach

Fast model nudges · reasoning model explains

Reads
The student's own deadlines, draft, confirmed feedback
On its own
Reminds before a deadline, flags blanks before submit, explains feedback another way, offers practice
Hands off
Nothing about grades or other students. It speaks only from approved message types.

Jia Yi fixed one of two flagged blanks before submitting: 11 of 12.

Guardrails, and where each one lives

A guardrail that exists only in a policy document does nothing. Each of Belay's is enforced by a specific component.

GuardrailEnforced byWhat it means in practice
The teacher approvesApproval gate in the orchestratorGrades, interventions and anything new to a student wait in the gate. Agents can propose but have no write access to the gradebook.
Agents stay in their lanePolicy engineThe coach can send reminders, blank checks, explanations and practice sets the teacher allowed, and nothing else. Every other message type is blocked.
Nothing happens unseenAction log, append-onlyEvery agent action is written in plain words with its reason. Students see when a message came from Belay rather than from Mdm Tan.
Grading stays supportTeacher console layoutThe attention view opens first. Grade review is a queue behind it, and low-confidence drafts are held, never published.
Data minimisation and PDPAWork vault and governance rulesCollect for a stated purpose, keep transcripts and derived signals, purge photos on the school's schedule, host in Singapore, never train general models on student work. Parents can see what is held about their child.
Access by roleSchool sign-inMdm Tan sees 3E1, 3E2, 3E3 and 3E5. She does not see 3E4. A student sees only their own work.
Student dignityPrompts and the Attention agentFlags describe work, not children: "submitted after the deadline", not "lazy". Flags go only to the teacher. Legibility is a prompt to look closer, never a mark.
The teacher is rightHistory ledgerA dismissed flag or a changed grade is recorded and tunes the thresholds for that class.