← mindsweeper.sg View demo dashboard →

Mindsweeper: end-to-end flow

From multi-agent assessment to three-audience report
v0.5 · 2026-07-01 · Multi-agent architecture reference · 8 agents
Stage 1 · Agent app

Evidence collection

Two entry paths into the same engine. A student works through a portfolio step, the M³ self-report instruments and a short working-memory check, then a live session in one of two modes. Parents and teachers instead upload existing materials (portfolio, transcript, or both). Both paths produce structured evidence for the blackboard.

Landing
Role select
Student · Parent · Teacher
Student
Intake
grade · syllabus · confidence · goal
Confirm details
Portfolio upload
past work · optional
M³ self-reports
SRQ-A · MAI · CYRM-R
Working-memory check
digit span · word span · Capacity
Mode select
Socratic · Choose your question · Skip
Evidence bundle
self-reports + spans + chat + portfolio
Parent / Teacher
Intake
student details · context
Confirm details
Input choice
Both REC
PDF + transcript
Written only
PDF, no chat
Chat only
transcript only
Portfolio upload
Portfolio upload
Transcript upload
+ Transcript upload
(skip chat)
(no portfolio)
Evidence bundle
PDF + transcript
PDF only
gaps flagged
Transcript only
both paths converge · all evidence streams → assessment engine
Stage 2 · Assessment engine

Eight agents, one blackboard

Evidence from any mode is normalised and written to a shared student model. Portfolio evidence passes through a dedicated ingestion agent first. Five assessors read and write simultaneously across the three layers — Content, Control and Capacity; the orchestrator drives live probing and assembles the report; the planner turns the profile into a targeted intervention.

Live Socratic chat
real-time probing
Portfolio analysis
uploaded work, PDFs
Offline transcript
prior dialogue
live chat, self-reports & span tasks write directly to the blackboard · portfolio passes through the ingestion agent first
Portfolio Ingestion Agent 7 of 8
OCR + step extraction · crossings-out → resilience signals · schema: {question_ref, steps[], corrections[], ocr_confidence}
structured evidence ↓ · low-confidence OCR flagged; orchestrator down-weights accordingly
Shared student model: the blackboard
every agent reads & writes here · the system of record, not an agent
read evidence · write findings + confidence + open gaps
Knowledge & Skill 2
Content · mastery · errors · transfer · SAP Coherence Map
Mindset & Motivation 3
Control · SDT spectrum · SRQ-A · Dweck (2006)
Metacognition & Comms 4
Control · calibration · MAI · Flavell (1979)
Resilience & Affect 5
Control · cliff vs slope · CYRM-R · Yeager & Dweck (2012)
Cognitive 6
Capacity · working memory · digit & word span · load ceiling
self-report (survey) and observed (dialogue) reads are held side by side · the Orchestrator computes their convergence per dimension
↻ live mode only: reads open gaps → picks the next probe (Reasoning / Perturbation / Teaching / Calibration / Frustration / Capacity / Structural) → back to live chat
Intervention Planner 8 of 8
SDT-gated strategies · topological sort of prerequisite graph · LVN strategy library · ANet decomposition
Stage 3 · Unified report

One report, one data contract

The orchestrator merges each assessor's region of the blackboard into a single JSON report, written to localStorage as a bridge between the agent app and the teacher console.

One unified report
localStorage.mindsweeper:latestReport
Knowledge map (Content) + Mathematical character / M³ (Control) + Cognitive profile (Capacity) + Intervention plan + Narrative summary
Stage 4 · Teacher console

Report viewer: three audience views

Same underlying data, three voices. Each view shows / hides / rephrases sections based on the audience. Switched via toggle in the topbar.

Toggle view in topbar →
Student view
Warm · encouraging
Minimal jargon
Growth framing
2nd-person voice
Parent view
Concerned · supportive
Plain language
Action suggestions
Ways to help at home
Teacher view
Diagnostic · professional
Full CCSS detail
All sections visible
Intervention plans