Transparency by Design

How ARIE Works Under the Hood

Every number, label, and recommendation in ARIE is explainable. No black boxes. This page documents exactly how scores are computed, trends detected, and decisions made.

On This Page
1. Observation → Score Pipeline2. Five Readiness Dimensions3. Readiness Score Formula4. NLP Observation Mapping5. Regression Detection6. Trajectory Engine (ESTE)7. Vocational Matching8. Confidence Scoring9. AI Growth Plan Layer
Overview

From Observation to Action

ARIE follows a deterministic pipeline. A teacher writes a plain-text observation. The system processes it through discrete engines, each with documented logic, to produce scores, alerts, and recommendations.

01
Observation
Free-text input
→
02
NLP Engine
Extracts dimensions
→
03
Vector Computation
Normalizes scores
→
04
Temporal Engine
Detects trends
→
05
Output Layer
Scores + Actions
Dimensions

Five Readiness Dimensions

Every observation maps to five measurable dimensions. Each dimension has a score from 0 to 100, computed from specific subcomponents with defined normalization rules.

Task Performance30%
Completion accuracy (35%), Multi-step adherence (25%), Error frequency (20%), Task endurance (20%)
Behavioral Stability25%
Escalation frequency (30%), Recovery speed (25%), Emotional volatility (25%), Peer interaction (20%)
Cognitive Adaptability20%
Task-switch latency (30%), Instruction retention (30%), Learning velocity (20%), Repetition requirement (20%)
Supervision Independence15%
Independence level (0-4 scale), normalized to 0-100
Consistency10%
Attendance (50%), On-time (30%), Routine stability (20%)
Scoring

Readiness Score Formula

The overall readiness score is a weighted sum of the five dimensions. This is a deterministic calculation with no AI involvement.

Formula
Readiness = (Task Performance × 0.30)
         + (Supervision Independence × 0.15)
         + (Behavioral Stability × 0.25)
         + (Cognitive Adaptability × 0.20)
         + (Consistency × 0.10)
Example: A student with Task=65, Supervision=50, Behavior=60, Cognitive=55, Consistency=70 → Readiness = 19.5 + 7.5 + 15.0 + 11.0 + 7.0 = 60.0
NLP Engine

Observation to Structured Data

Teachers write free-text observations. ARIE's NLP engine maps these to the five readiness dimensions using keyword extraction and pattern matching. Supports English and Hindi input.

Input
“Needed two reminders to complete packaging task, lost focus after 15 minutes, improved after encouragement.”
Output
task_performance: 55
supervision_independence: 40
behavioral_stability: 60
cognitive_adaptability: 45
consistency: 50

Each observation creates a new “snapshot” — the dimension scores are merged with historical scores using an exponential moving average to prevent single observations from causing large swings.

Safety Net

Regression Detection

ARIE continuously monitors for regression — sustained drops in readiness scores that may indicate a child needs additional support. Two severity levels are tracked.

High Risk
  • Readiness drops >12% over 2 consecutive weeks
  • OR Behavioral Stability drops >15% in any single week
Medium Risk
  • Readiness drops >6% in a single week
  • OR Supervision Independence drops >15%
  • OR Cumulative drop >10% over 4-week sliding window
ESTE

Early Support Trajectory Engine

ESTE analyzes recent readiness score history to compute trajectory direction, stability, and whether an early support window should be activated. It uses the recent 4-week window for slope calculation and the full history for volatility analysis.

Direction
Linear regression slope over last 4 weeks:
Improving: slope > 0.5
Needs Attention: slope < -0.5
Stable: between -0.5 and 0.5
Stability
Measures volatility (residual std dev) from the full-history regression line:
High: < 3.0
Moderate: 3.0 – 6.0
Unstable: > 6.0
Early Support Window
Activates when:
• Slope < 0 + not high stability
• 2+ consecutive weekly declines
• Medium regression risk detected
Confidence is based on data volume: ≥8 weeks = High, ≥5 weeks = Medium, <5 weeks = Low
Pathways

Vocational Matching Algorithm

ARIE matches students to suitable vocational pathways by computing cosine similarity between the student's readiness vector and predefined job profile vectors, with constraint-based penalty adjustments.

Matching Process
  1. Compute cosine similarity between student vector and each job vector
  2. Check dimension deficits against job requirements
  3. Apply penalty for deficits exceeding 15-point tolerance
  4. Disqualify if any single deficit exceeds 30 points
  5. Rank by effective similarity score
Job Profiles
Each job has a target vector across all 5 dimensions:

Packaging Assistant — High consistency, task focus
Data Entry Clerk — High cognitive, consistency
Cleaning & Maintenance — High consistency, moderate task
Sorting & Inventory — High task performance
Trust

Confidence Scoring

ARIE computes how confident you should be in its assessments using three independent factors, combined into a composite score.

Temporal Maturity
How many weeks of data exist. Full maturity at 8 weeks. Fewer weeks = lower temporal confidence.
Score Volatility
Standard deviation of readiness scores across snapshots. High volatility (>8.0 std dev) reduces confidence.
Observation Frequency
Fewer than 0.5 observations per week reduces confidence. Regular data input strengthens trust in scores.
Composite ≥ 0.75 → High Confidence  |  ≥ 0.45 → Medium  |  Below → Low
AI Layer

AI-Powered Growth Plans

This is the only component that uses generative AI (Gemini). The growth plan takes all deterministic outputs as input and generates 3 actionable recommendations with weekly focus areas.

AI Receives
  • Current readiness vector (5 dimensions)
  • Trend slopes per dimension
  • Regression risk level
  • Vocational match results
  • Confidence level
  • Recent observation texts
AI Produces
  • 3 specific growth recommendations
  • Each with a focus dimension and weekly action
  • Framed in supportive, actionable language
  • Never labels, never predicts outcomes
Key distinction: The AI does not compute scores, detect regression, or make classification decisions. It only generates human-readable recommendations based on deterministic inputs. If the AI is unavailable, all other features continue to function normally.

See It in Action

Every score on the dashboard is computed using the logic documented above.

Open Dashboard →