INSIDE GBB ONE

Multiple agents.
One thoughtful learning loop.

A single answer can tell us more than “right” or “wrong.” Our architecture is designed to turn that evidence into a better next problem.

THE ADAPTIVE LOOP

Separate responsibilities make the process easier to inspect, evaluate, and improve.

  1. 01Student activityAnswers & reasoning
  2. 02AssessmentIdentify patterns
  3. 03Skill modelUpdate learning evidence
  4. 04CurriculumChoose a target
  5. 05GenerationWrite original practice
  6. 06VerificationIndependently check
  7. 07Student practiceWork through the problem
  8. 08Grading & reasoningReturn evidence to the model
Updated skill model · A new learning cycle begins
SPECIALIZED BY DESIGN

Five roles. Shared purpose.

01

Assessment Agent

Understand the learner

Analyzes previous answers, tracks skills and subskills, and looks for recurring reasoning weaknesses. It updates a working estimate of mastery—not a fixed label.

EXAMPLE SIGNAL

Basic counting is secure; problems with multiple constraints need more attention.

02

Curriculum Agent

Choose the next step

Balances focused remediation with mixed review. It adjusts difficulty and prevents students from practicing only the topics they already know.

EXAMPLE DECISION

Select 3 targeted counting problems and 2 mixed-review problems.

03

Problem Generation Agent

Create original practice

Uses grade, target skill, recent mistakes, mastery estimates, and difficulty to compose a new competition-style problem. It does not copy or remix competition question banks.

EXAMPLE TARGET

Grade 5 · counting · complete case analysis · medium difficulty.

04

Verification Agent

Check before showing

Independently solves the proposed problem, validates the answer and any choices, and checks for ambiguity and appropriate difficulty. A failed check sends a problem back for revision.

RELEASE GATE

A valid solution, consistent answer choices, and clear wording are required.

05

Grading & Reasoning Agent

Learn from the answer

Evaluates answers and, when provided, reasoning. It distinguishes possible conceptual gaps from arithmetic slips, gives supportive explanations, and returns evidence to the skill model.

POSSIBLE MISTAKE CATEGORIES

Arithmetic slip · incomplete case analysis · missed constraint · misunderstood ratio.

BUILT WITH AGENTIC AI

Claude at the core.
Care at every step.

GBB One is being built with Claude as a core reasoning engine for problem generation, independent mathematical verification, mistake analysis, and adaptive tutoring.

Each practice cycle is designed to generate structured results: a target skill, an original problem, a checked solution, a reasoning assessment, and an updated learning plan.

A separate verification pass adds a quality check. It cannot guarantee correctness; evaluations and human oversight remain part of the design.

Where we are today

The public demo illustrates the learning loop with prewritten examples. The live Claude API pipeline, persistent skill model, and parent accounts are under development.

PLANNED TECHNICAL ARCHITECTURE
Student web appParent-managed profiles
Learning profile serviceSkill estimates & learning evidence
Agent orchestratorAssessment · Curriculum · Generation
Independent verificationMathematics · Wording · Difficulty
Claude APIStructured results & reasoning

Checked problems return to the student. Grading evidence updates the learning profile.

ORIGINAL BY DESIGN

Built around skills,
not copied question banks.

Problems are designed around mathematical concepts, reasoning patterns, and difficulty levels. We do not reproduce copyrighted competition questions or generate variants of historical AMC or Math Kangaroo problems.

Arithmetic reasoningFractions & ratiosGeometryCounting & combinatoricsNumber theoryAlgebraic reasoningProbabilityLogic & patterns