REAL-TIME TEACHER-LED ADAPTIVE LEARNING
Every discipline LingoSide draws on, the synthesis that stress-tested it, and the verdicts that decide what gets built.
In Stanford’s randomised Tutor CoPilot trial, students of tutors who could call up AI-suggested replies were four percentage points more likely to pass the end-of-lesson mathematics check — nine points where tutors had the least experience. The tutor chose, edited or ignored each suggestion.
The tool was tutor-facing only. The tutor opened it; it read the lesson chat and offered three possible replies, which the tutor could send, edit, regenerate or leave unused. It never messaged the student and never decided what the student needed.
The trial studied text-chat mathematics tutoring, not LingoSide, and measured end-of-lesson checks rather than end-of-year tests — LingoSide’s own learning outcomes must be measured directly.
Research-informed by design
LingoSide draws from dozens of disciplines and research areas. They are not a technology checklist; together, they define how the system should pay attention, capture and structure evidence, and decide what to bring back.
Research on forgetting, retrieval practice, spacing, consolidation and metacognition guides how review decides when a lesson moment should return and what effort creates durable recall.
Second-language acquisition, psycholinguistics and language pedagogy shape how input, output, corrective feedback, vocabulary and meaningful interaction work together.
Speech recognition, listening, phonetics, phonemes, pronunciation and prosody inform what the speech analysis observes responsibly, and where the system stays silent.
Knowledge tracing, adaptive learning, error classification and learning analytics offer ways to update practice from observed performance rather than a fixed course sequence.
Conversation analysis, NLP, semantic search, retrieval and generative models help structure what happened, while human-computer interaction keeps that analysis from competing with the lesson.
Research is a design constraint, not a badge. It helps decide what belongs live, what belongs after the lesson, what needs validation and what should not be built yet.
Stress-tested, not cherry-picked
We gathered the science of how people really learn languages into one brief, then tried hard to prove ourselves wrong. Only the claims that survived get to shape LingoSide.
How to read this evidence
Each public claim names the evidence type, links to its source, states its boundary and carries a review date. These labels describe evidence maturity; they are not an ESSA rating or a certification of LingoSide.
The public synthesis narrows 261 sourced claims to 30 audited claims, 18 that held up across relevance checks and 4 findings that directly shape the product.
verifiable claims extracted, each tied to a published source
Verification results for the 30 key claims
Held up
Findings that shaped the product
Evidence maturity path
This path is a roadmap for accumulating evidence, not a claim that LingoSide has reached every stage.
research areas reviewed, from memory science to speech decoding
verifiable claims extracted, each tied to a published source
unique sources, with meta-analyses weighted first
feature ideas mapped from the evidence into a phased plan
Every key claim had to pass three simple tests: Is the research solid? Has it been repeated by other scientists? Does it apply to a real online lesson? Here is what happened to the 30 claims our product depends on:
Passed all three tests. We build on these.
One test raised doubts. We use them carefully.
Popular beliefs that failed the tests. They stay out of the product.
With captions in the language you’re learning, people simply understand much more of what they hear.
Meta-analysis of captioned video: listening comprehension g ≈ 0.99 vs uncaptioned viewing — Montero Perez et al. (2013).
Read the source
Source checked
→ That’s why live captions are the heart of the sidecar.
Pulling a phrase from memory strengthens it far more than reading it one more time, and the gap widens over weeks.
Meta-analysis of 188 experiments: g = +0.51 vs re-reading — Adesope et al. (2017).
Read the source
Source checked
→ LingoSide is building review that quizzes you gently instead of showing lists.
Short practice spread over days wins against one long session, especially weeks later, when it actually matters.
Meta-analysis: 48 experiments, 3,411 learners, medium-to-large effect — Kim & Webb (2022).
Read the source
Source checked
→ LingoSide is building practice for between lessons, never inside the live lesson.
Quick hints in your native language speed up beginners, and matter less and less as you advance.
Meta-analysis of native- vs target-language word hints: g = 0.33; the advantage shrinks with proficiency — Kim, Lee & Lee (2024).
Read the source
Source checked
→ LingoSide is building translation support that beginners can lean on and advanced learners need less.
The stronger the evidence, the bolder the feature. Solid findings become defaults and promising ones become options. Busted myths stay out, no matter how popular they are.
REAL-TIME TEACHER-LED ADAPTIVE LEARNING
Join the waitlist or tell us how your learners and teachers work today.