Project · Reading Trainer
The goal is a better reader — and proof that holds.
Reading is a motor skill before it is anything else: several eye movements a second, each landing, holding, suppressing the rest of the page — and, in the same instant, a bet on what comes next. Fluent readers make that bet cheaply and are usually right. For many people it never becomes cheap, and the cost shows as slowness, fatigue, and comprehension that leaks. The ambition of this project is to lower that cost — to make reading faster and lighter for the person doing it, and to know, not hope, that it worked.
Where it stands: phase one is built and tested. The reading corpus is being scored; the phase-two drills come next. Training claims come only after, and only if the numbers earn them.
The standard
Reading-improvement claims are everywhere and almost none can be checked. So this is built the other way round: measurement first, claim last. The reading outcome is sealed off from the software that runs the sessions, so no result can be quietly invented; every phase ends on a measured plateau, not a calendar. If a gain appears, it will be one that survives.
Phase one: the eye
Fixation, the targeted jump, the suppressed distraction. A small, cluttered symbol is unreadable in peripheral vision. That failure is the instrument. If a person correctly names which way the gap in a ring points, their eye was pointed at it — no inference required.
The symbol changes state every fifth of a second, which turns identity into a clock: the state someone reports says when their eye arrived, with reaction time algebraically removed. It measures where the eye was and when it got there, using only a keyboard.
What phase one tracks
Peak scores are the least informative thing about a skill. We measure how performance decays across a sustained block, and how far it falls when a second task competes for attention. If control genuinely improves, it should become cheaper, not merely higher. Training advances only when the effort of holding steady has dropped — and steadiness is then tracked through everything that follows, so that a reading gain can never be mistaken for the eye simply settling.
Phase two: the reader
Once the eye is steady and stays steady, training turns to what a fluent reader does without thinking: betting on what comes next. A study of 225,000 words of parsed English found where that bet can be won — not in words, which are an open set with no end to them, but in the closed sets. About five phrase shapes cover most of the language, and their first word announces them. The role of the last word read narrows what can follow from eleven possibilities to five. A verb's frame settles most of what comes after it. The ending of an unfamiliar word fixes its class four times in five.
Phase two drills exactly these — phrase families from their openers, word classes from morphology, the frames of the fifty commonest verbs, the two readings of to and that as two different words — and leaves out what the data says is not learnable: modifier order, adjuncts, the open sea of vocabulary. The same map grades the text, so that a harder passage is a number rather than a feeling, and gains can be measured on difficulty that stays fixed.
Read the findings: what a reader can predict from the sentence so far →
How we avoid fooling ourselves
The program that presents sessions cannot import the analysis code — a separation enforced by an automated test that fails the build if it is ever breached.
Phases advance when measured performance plateaus, not on a schedule, and only when the effort measures have flattened rather than the scores alone.
An internal check requires that a person holding still and guessing scores no better than chance; if they do not, the stimulus is wrong and every result downstream of it is void.
The reading-outcome field in our data model is deliberately empty, unwritable by the current software, so that no outcome can be quietly invented to match a hope.
What this can and cannot show
This is a single-subject study under tightly controlled conditions. At best it can demonstrate that something changed for one person, in a way that survives the checks above. It cannot show that a training method works in general, and it will not be described as if it could. It may well show nothing at all — the design is built to make that outcome visible rather than deniable.