IIThe method

Five signals,
computed openly.

NoteEach measurement has a documented cognitive-science root, a transparent formula, and a published failure mode. Scepticism is welcomed; we publish what we'd need a sceptic to know.

Signal · /0101

Clarity.

Sessions-to-mastery on a new concept.

How it is computed

Bayesian Knowledge Tracing. Each child holds a latent probability of mastery per concept; every attempt updates the posterior. The signal we report is the number of practice attempts before the posterior crosses 0.85, smoothed across similar concepts.

Rooted in

Corbett & Anderson (1995) — Knowledge Tracing in the Cognitive Tutor. Still the most-cited baseline because it survives translation across subjects.

Where it fails

Children who guess well on multiple-choice. We mitigate with item-response-theory weighting; we do not pretend it is solved.

Sample reading · Class 9 student

Placeholder · real pilot data soon

Sample reading replaces this placeholder when pilots conclude. Each reading will link to a per-concept breakdown.

Signal · /0202

Retention.

Recall probability over time without revision.

How it is computed

A modified forgetting-curve fit per concept per child. We fit a stretched exponential to each learner's recall data and project the curve forward. The signal is the projected recall at the next assessment date.

Rooted in

Ebbinghaus (1885) and the modern half-life regression family. Closer to what the spaced-repetition community calls 'stability'.

Where it fails

Concepts that are reinforced outside school (a child uses fractions while baking with a parent). We track external practice when reported, but blind spots remain.

Sample reading · Class 9 student

Placeholder · real pilot data soon

Sample reading replaces this placeholder when pilots conclude. Each reading will link to a per-concept breakdown.

Signal · /0303

Application.

Transfer accuracy on unseen problem shapes.

How it is computed

We tag every question with a Bloom's-Taxonomy level and a problem family. Application is the gap between accuracy on previously-seen patterns and accuracy on novel patterns within the same concept. A child with high clarity but low application has memorised, not understood.

Rooted in

Bloom (1956), Anderson & Krathwohl (2001) revised taxonomy. The signal that most often correlates with adult competence.

Where it fails

Subjects where 'novel' is hard to define (humanities, art). We currently scope this signal to math and science.

Sample reading · Class 9 student

Placeholder · real pilot data soon

Sample reading replaces this placeholder when pilots conclude. Each reading will link to a per-concept breakdown.

Signal · /0404

Exploration.

Attempt-rate on questions outside the student's confident range.

How it is computed

A child is offered a stretch question they have not been formally taught. We measure whether they attempt it and, separately, whether they think out loud. High explorers learn faster across the board, even when their immediate accuracy lags.

Rooted in

Dweck's growth-mindset research, intersected with classroom productive-failure (Kapur, 2008).

Where it fails

Students whose courage is context-dependent (will attempt at home, freezes in class). The signal is averaged across contexts; we surface the variance, not just the mean.

Sample reading · Class 9 student

Placeholder · real pilot data soon

Sample reading replaces this placeholder when pilots conclude. Each reading will link to a per-concept breakdown.

Signal · /0505

Calibration.

Confidence vs. actual performance, per concept.

How it is computed

Before answering, the student rates their confidence on a 1–5 scale. After grading, we compute the calibration error — Brier-style — per concept and roll it up. Low calibration error is the single best predictor of exam-day collapse in our data.

Rooted in

Metacognitive research from Flavell (1979) through Dunning & Kruger (1999). The signal that most respects the child's interiority.

Where it fails

Cultural variance — some children are taught to under-report confidence. We control for baseline per-child rather than per-cohort.

Sample reading · Class 9 student

Placeholder · real pilot data soon

Sample reading replaces this placeholder when pilots conclude. Each reading will link to a per-concept breakdown.

VIComposition
How they compose

The signals are interdependent.

A child cannot trade Retention for Application without one of them falling. This is intentional. It is also what stops a school from gaming the new measurement the way it sometimes games marks.

In our internal use, we summarise a child's state as a five-axis radar. The shape of that radar — not any one axis — is what teachers, parents, and the child themselves work with. A symmetric pentagon means consistent strength; a skewed one is a roadmap for the next four weeks.

What we don't measure
  • · Behaviour, attendance, or social dynamics.
  • · Emotional state in real time.
  • · Anything outside the classroom and the homework session.
  • · Anything that would identify a child outside their own school.

We measure understanding. Not children.

Measurement is an act of attention. Done well, it is the same thing as care.
Daniel L. Schwartz · 2016

☞ Every signal links to its source paper when pilots publish.

Run a pilot ↗