Reading Science + Adaptive AI

The Science Behind Finley

How AI and decades of learning science work together to teach your child to read with real understanding, not lucky guesses.

Built on the Science of Reading

The Science of Reading is not a curriculum or a teaching method. It is a body of converging evidence from cognitive science, linguistics, neuroscience, and education research that describes how the human brain learns to decode written language. This research spans decades and thousands of peer-reviewed studies.

Finley is built directly on this evidence base. Our reading progression follows Dr. Elfrieda Ehri's four phases of word reading development, the most widely cited model for how children move from pre-readers to fluent, automatic readers:

Phase 1: Pre-Alphabetic (Levels 1-3, Ages 4-5)

Children recognize words as visual shapes. They might identify "McDonald's" by the golden arches, not by decoding the letters. Finley's early levels focus on letter recognition, phonemic awareness, and connecting letters to their most common sounds.

Phase 2: Partial Alphabetic (Levels 4-6, Ages 5-6)

Children begin using the first and last letters of a word to guess at pronunciation. "Ball" might be recognized by the "b" and "l." Finley introduces systematic phonics instruction, CVC (consonant-vowel-consonant) blending, and high-frequency sight words.

Phase 3: Full Alphabetic (Levels 7-9, Ages 6-8)

Children can map every letter to its sound and decode unfamiliar words independently. Finley builds decoding fluency, introduces more complex phonics patterns (digraphs, blends, vowel teams), and increases text complexity.

Phase 4: Consolidated Alphabetic (Levels 10-15, Ages 8-12)

Larger letter patterns, morphemes, and word families are stored as units. Reading becomes automatic and fluent. Finley shifts focus toward reading comprehension, vocabulary expansion, and exposure to varied text types.

Each child enters Finley's 15-level reading ladder based on an AI-powered assessment, not a parent's estimate. The system continuously reassesses as the child progresses, ensuring they are always working at their optimal challenge level.

Story lesson screen

How the Adaptive AI Works

Most reading apps claim to be "adaptive." In practice, they get slightly harder when your child gets answers right and slightly easier when they get answers wrong. That is not adaptive learning. That is a difficulty slider.

Finley uses a fundamentally different approach built on three interconnected systems:

Bayesian Knowledge Tracing (BKT)

BKT is a probabilistic model developed in educational research to estimate the likelihood that a student has truly mastered a specific skill, not just gotten a question right. For each skill (letter sounds, CVC blending, sight words, digraphs, etc.), BKT maintains four probability estimates:

  • P(know): the probability the child has actually learned the skill
  • P(guess): the probability the child got a correct answer by guessing
  • P(slip): the probability the child knew the answer but made a careless error
  • P(learn): the probability the child transitions from not knowing to knowing after each practice opportunity

This means Finley can distinguish between a lucky guess and genuine mastery. A child who gets "cat" right once might have guessed. A child who consistently blends CVC words across multiple contexts has demonstrably mastered the skill. Finley does not advance until mastery is real.

Spaced Repetition (Leitner System)

Finley uses a modified Leitner system to schedule when previously learned skills and words resurface. Skills that the child finds difficult appear more frequently. Skills that are well-mastered appear at longer intervals. This is the same principle behind effective flashcard systems, but applied automatically and continuously within the reading experience.

The result: nothing slips through the cracks. A word mastered in Week 1 is revisited in Week 2, then Week 4, then Week 8, ensuring long-term retention.

AI-Generated Content Engine

Every story, word puzzle, and phonics activity in Finley is freshly generated by AI for each session. This is not a content library. When your child opens Finley, the system evaluates their current mastery profile, identifies skill gaps, and generates content that targets exactly what they need to practice, incorporating their interests to keep engagement high.

The AI content engine sends anonymized parameters to generate appropriate content: the child's current reading level, target skills, and interest tags. No child names, identifiers, or personal information are sent to the AI generation service.

Phonics activity screen

Parent Coaching: The Research-Backed Multiplier

One of the most consistent findings in reading research is the power of parent involvement. The Best Evidence Encyclopedia (Johns Hopkins University), drawing from over 100 rigorous studies, found that one-on-one human interaction is the single strongest driver of reading gains in young children.

After every Finley session, parents receive an AI-generated insight report that includes:

  • What skills the child practiced and how they performed
  • Which words were mastered and which need more practice
  • A specific coaching prompt: one question to ask or one activity to do with the child (takes about 60 seconds)

A parent asking "What was your favorite part of the story?" or "Can you find three words that start with 'sh' in this room?" is acting like a personal reading tutor. No teaching degree required.

Coaching prompts are optional. Your child benefits from Finley whether or not you use them. But the research is clear: families who engage with the prompts see the strongest gains.

Parent dashboard screen

What Makes Finley Different

Feature Typical Reading App Finley
Adaptivity Gets harder/easier based on correct/incorrect answers Bayesian Knowledge Tracing at the individual skill level
Content Fixed content library AI-generated, personalized every session
Mastery Detection Score-based (80% = pass) Probabilistic (distinguishes guessing from mastery)
Retention No review system Spaced repetition resurfaces skills at optimal intervals
Parent Role Progress bar or star count AI-generated coaching prompts grounded in research
Reading Science Marketing claim Ehri's phases structure the entire 15-level curriculum
Data Privacy Varies widely COPPA-compliant, no ads, no data sales, parent-controlled

Privacy by Design

Every engineering decision in Finley starts with the question: "Would I be comfortable with my own child's data being handled this way?"

  • No child names or personal identifiers are sent to AI services
  • No advertising SDKs, behavioral tracking, or third-party analytics that profile children
  • The child-facing experience contains no external links, no social features, and no way to contact other users
  • The parent dashboard is behind a PIN lock; children see only the reading experience
  • All data is encrypted in transit and at rest

For complete details, see our Privacy Policy and Data Retention Policy.

Technical Architecture (For the Curious)

  • Frontend: Vite + React, optimized for low-bandwidth and offline-capable reading sessions
  • Authentication: Firebase Auth with email verification for parent accounts
  • Database: Cloud Firestore with field-level security rules separating parent and child data
  • AI Content Generation: Anthropic API with structured prompts; no PII transmitted
  • Payments: Stripe for subscription management (PCI DSS Level 1)
  • Hosting: Cloudflare Pages for the marketing site; Cloudflare Workers for the application
  • Adaptive Engine: Custom BKT implementation running client-side for real-time responsiveness