Skip to main content
Integrity & Anti-Cheat

Score-level integrity — not surveillance

LearnVyx is not a proctoring tool. We don't operate cameras, run facial recognition, lock browsers, or conduct room scans. Our integrity model works at the scoring layer: the Response Profile Integrity Score (RPIS) detects statistically improbable response patterns at the session level and applies a Bayesian penalty to the final ability estimate — making it harder to fabricate a score than to earn one.

See a Demonstration Read the Research
The Approach

Why hardware proctoring fails — and what works instead

Camera-based proctoring solves the wrong problem. It attempts to catch the act of cheating after the behavior has occurred — and in doing so, it treats every learner as a suspect and introduces significant false-positive rates for learners with certain disabilities, home environments, or internet conditions.

The behavioral signal approach works at the scoring model level. A learner who externally sources answers will exhibit statistically improbable latency profiles, answer clustering inconsistent with their established ability estimate, and revision patterns that deviate from the baseline distribution for their demographic cohort. You can't fake those signals.

LearnVyx's integrity engine doesn't flag individual items as "suspicious" — it computes a response-profile integrity score (RPIS) for the full session and applies a Bayesian penalty to the final ability estimate when RPIS falls below threshold.

Hardware Proctoring

  • High false-positive rate (12–18% in peer-reviewed literature)
  • Privacy invasive — requires facial recognition and room scans
  • Fails remote learners with inconsistent internet
  • Disproportionate impact on learners with disabilities

LearnVyx Behavioral Signals

  • Model-level integrity — works at the score, not the item
  • No hardware or browser extension required
  • Equitable across all test environments
  • Statistical confidence interval on integrity score
Signal Architecture

Four behavioral signal streams

Each signal stream is weighted independently, then combined using a calibrated Bayesian network to produce the Response Profile Integrity Score (RPIS).

Response Latency
Item-level time-to-answer is compared against the expected distribution for that item's difficulty and the learner's current ability estimate. Suspiciously fast correct answers on hard items are flagged; statistically normalized wrong answers on easy items are not.
Answer Confidence Patterns
If a learner self-rates high confidence on items that should be above their estimated ability level — and answers correctly — the confidence-accuracy correlation is used to detect either genuine mastery or response contamination.
Item Revision Patterns
How often a learner changes their initial answer, and which direction they change (correct-to-wrong vs wrong-to-correct), provides a secondary signal stream. External assistance tends to shift the revision direction distribution toward correct reversals on hard items.
Statistical Outlier Detection
Person-fit statistics (infit/outfit mean squares) are computed against the 3PL model's expected response pattern. Extreme person-fit values indicate a response pattern inconsistent with any ability level — a strong indicator of non-ability-based response behavior.
Abstract visualization of behavioral assessment signals — response latency, confidence patterns, and statistical outlier detection represented as overlapping waveforms
Get Started

Replace surveillance with measurement science

Book a 30-minute call to see how the RPIS model works on a live assessment — and how it performs against known-integrity cases in your subject matter domain.

Request a Demo