Stealth assessment: how AI can know if a child is truly reading without making them feel tested

Post 15 of 17 · The Somastars Phygital Thesis · The Solution

A teacher in a Standard 6 class in Nairobi manages 42 students. In a 40-minute reading period, she can meaningfully observe perhaps four or five children. She knows the fast readers and the slow ones. She does not know — in any precise sense — why the slow ones are slow: whether they are Foundation readers masquerading as First Floor readers, whether their vocabulary is thin or their inference is weak, or whether they are simply not reading at all.

This is the problem that AI-assisted assessment was designed for. Not to replace the teacher's judgment — which remains irreplaceable for relational and motivational tasks — but to provide a diagnostic precision that one human monitoring 42 children simultaneously cannot physically achieve.

The Zone of Proximal Development as a control target

Lev Vygotsky's Zone of Proximal Development (ZPD) describes the space between what a child can do independently and what they can do with support. Effective instruction keeps the child in that zone: challenged, but not overwhelmed; progressing, but not coasting.

SomaStars' adaptive engine targets the ZPD through a discrete-time control algorithm. The formal expression: D(t+1) equals D(t) plus alpha multiplied by (P(t) minus T), where D is the current difficulty level, P is the child's actual performance as a fraction of correct answers, T is the target mastery threshold (0.8), and alpha is the adaptivity coefficient. If a child answers more than 80 percent of questions correctly, the difficulty increases. Below 80 percent, it decreases. The child is always working at the productive edge of their capability.

What 'stealth' means in practice

The assessment is described as stealth because it collects data without interrupting the reading experience. Two signals are captured for every question: semantic accuracy (is the answer correct?) and temporal accuracy (how long did the child take to answer?). A correct answer given in 2 seconds at Level 8 is different from a correct answer given in 45 seconds. The first suggests fluent comprehension. The second suggests laborious decoding of the question itself — a Foundation signal disguised as performance at a higher level.

The engine also tracks self-correction frequency: how often a child changes their answer before submitting. High self-correction rates at Levels 5 through 8 indicate active inferential processing — the child is weighing options, which is exactly the cognitive behaviour the First Floor of the Reading House demands. Low self-correction at the same levels, combined with low accuracy, indicates guessing. These profiles are different problems requiring different responses.

"A correct answer given in 2 seconds at Level 8 is different from one given in 45 seconds. The engine knows the difference."

The human-in-the-loop safeguard

Every AI-generated question in SomaStars passes through a moderator dashboard before reaching a child. The dashboard shows the question, the correct answer, and the pedagogical level the AI assigned. If the AI labels a question Level 13 but the answer is explicitly stated in the text — a DOK 1 response to a claimed DOK 4 question — the moderator rejects it. The target is 98 percent accuracy in AI question-level assignment. Every question that passes review carries a source anchor: the specific panel or paragraph that generated it.

The takeaway

Stealth assessment does not mean hidden surveillance. It means learning measurement that is indistinguishable from learning engagement: a quiz that feels like a game, a score that feels like an achievement, and a diagnostic that the teacher sees on Monday morning as a precise map of where every child's Reading House needs work. That is what 42 students and one teacher actually need.