Mechanism

How it works

Three sources feed one decision about what's happening on the page. They don't carry equal weight, and the ordering is the product, not an implementation detail.

Authority 1 · ground truth

Reader responses

Answers to retrieval questions (the only ground truth in the system). An answer outranks anything the sensors say. Right resolves to on pace and stops the system pressing; declining a question asserts nothing about comprehension.

Authority 2 · primary path

Browser telemetry

Reading rate against text difficulty and your own baseline, scroll regressions, selection, copy, tab blur, idle time. Precise, always available, no permission needed. This is the default path and what most readers run on.

Authority 3 · optional, coarse

Webcam gaze

Presence and a rough screen region only (several lines of text wide). It can assert that you've stepped away, or corroborate a state telemetry already suspects. It can never assert a reading state on its own.

Read left to right: an answer overrides telemetry, telemetry overrides the camera, and the camera never gets the final word by itself.

The state vocabulary

Six states. Every one is an observation, not a feeling

None of this claims to read what you're thinking. Each state names something measurable: a pace, a pattern of movement, a period without signal. Never an internal state like confusion or boredom, which no browser can see.

StateWhat it observes
on paceReading speed is consistent with the paragraph's difficulty and your own baseline.
skimmingMoving faster than the text's difficulty would predict for careful reading.
strugglingMoving markedly slower than baseline, or re-reading the same lines repeatedly.
driftingScroll and dwell patterns that don't match active reading of the visible text.
awayThe tab lost focus, or. with the camera on. no one is at the screen.
unknownNot enough signal to say anything. Common, normal, and nothing interrupts on it.

Limitations, stated plainly

States are inferred from behaviour, and inference is imperfect. Someone can read carefully at an unusual pace and get misread as skimming; someone can stare at a page while thinking about something else and get read as on pace. That's exactly why the system doesn't act on an inference alone. it asks a question and treats the answer as the only thing that settles it. When there isn't enough signal to guess responsibly, the honest answer is unknown, and the system does nothing with it. That's a design choice, not a gap: guessing confidently from thin signal would be worse than saying nothing.

Why a question, not a summary

Summaries feel like understanding. Recall produces it

A summary is something you're shown. A question is something you have to answer, using the passage you just read rather than the assistant's paraphrase of it. That distinction is the entire mechanism, and it isn't ours. it's retrieval practice, one of the more replicated findings in reading research.

D'Mello et al. (2016) ran a randomised trial interrupting drifting readers with just-in-time questions and found it recovered comprehension losses (d = 0.47). alcoia's interruption policy. one question at most every three minutes, five a session, never twice on the same paragraph, never on unknown. is built around not spending that effect on questions the reader doesn't need.

Curious what the site can't claim, and why: /research

Add to Chrome. free