The pace that watches you

The pace control in the app is a number you set once and mostly forget. Around it the stream already breathes a little: dwell stretches for long words, numerals and commas, further at sentence ends, and up to 2.8× at a paragraph break. All of that responds to the text. None of it responds to you. Closing that gap is the obvious next move, and it is where a lot of reading products are about to go wrong.

The wrong version is easy to describe because it is nearly always what gets built. Watch the reader for signs of strain. When the signs are absent, speed up. When they appear, slow down. Ship it as an adaptive engine, put a chart in the settings screen, and let the number climb over a fortnight while the user feels increasingly capable. It will demonstrate beautifully. It will also be optimising for the wrong quantity, for reasons we will get to.

First, though, the sensors, because they are what the category is going to argue about. There are four families of signal, they differ enormously in quality, and the strength of the marketing around each is close to inversely related to the strength of the evidence.

Signals you already produce

The cheapest signals are behavioural, and they are already in front of us. A rewind is a comprehension failure with a timestamp on it. A pause before a paragraph break is different from a pause in the middle of a clause. Where a session was abandoned tells you something about where the material got away from the reader. And the recall check at the end of a passage — real targets mixed with plausible distractors, scored next to the speed — is not a proxy for comprehension at all; it is a measurement of it, however coarse.

This family has an unglamorous virtue: it requires no hardware, no permissions, no camera, and nothing that leaves the machine. It is also, we suspect, the most informative per unit of intrusiveness of anything on this list. A rewind at word 4,120 of a contract is a stronger statement about that sentence than any pupil measurement taken at the same moment, and it costs nothing to collect. The reason the industry will chase sensors instead is that sensors demonstrate better in a video.

Eyes, wrists and heads

Eye tracking is the first genuinely new signal, and it is now present in some headsets and in phone hardware. The underlying science is real: fixation behaviour and pupil response do carry information about cognitive load, and the eye-movement literature has been building on that for decades under laboratory conditions. Consumer conditions are not laboratory conditions. Pupil diameter responds to ambient light far more strongly than to mental effort, and it also responds to fatigue, to caffeine, to the emotional content of what is being read, and to the individual sitting in the chair. The signal is there. Extracting it from a phone held at an angle in a lit room is a different problem from measuring it in a chin rest.

Wearables are weaker again. Heart-rate variability and skin conductance are reasonable proxies for arousal, and arousal is related to cognitive effort in a loose and heavily confounded way. A reader whose heart rate rises during a difficult paragraph and a reader whose heart rate rises because someone knocked at the door produce the same trace. Building a pace controller on that is not impossible, but the honest description of what it senses is something changed, not this sentence was hard.

Then EEG, and the consumer brain–computer interface. The laboratory literature here is real and long-standing; electrophysiological work on language comprehension has been productive for decades, under controlled conditions. The gap between that and a headband worn over hair while a person sits on a sofa is enormous: fewer electrodes, worse contact, motion artefacts, and a signal-to-noise ratio that laboratory work manages with repeated trials and averaging, which a live reading session cannot do. The marketing in this space is already well ahead of the evidence, and we would rather say that now than after it becomes a feature checkbox.

The mistake underneath all of them

Suppose the sensor problem were solved. Suppose a system could read your state precisely and adjust the pace accordingly. It would still be built on a target function, and the target function nearly every adaptive system reaches for is comfort — speed up while the reader looks relaxed, slow down when they look strained. That target is wrong, and it is wrong in a specific and expensive way.

Comfort and comprehension come apart exactly where the decision matters. Readers routinely overestimate how much they have understood, and the illusion is strongest when the material is fluent and easy to process. Word-by-word presentation at speed is close to an ideal machine for producing that feeling: the display is smooth, the words keep arriving, the sense of following along is intact, and none of that is necessarily connected to whether a proposition was built. A system optimising for comfort will find the rate at which reading feels best, which is often somewhat faster than the rate at which reading works — the gap we take apart in the comprehension tax.

Optimise for comfort and you will find the rate at which reading feels best. That is not the same rate.

The target function problem

The correct target is not the fastest rate you tolerate. It is the fastest rate at which your recall stays where you want it — and the phrase where you want it is doing real work, because the answer differs for a novel and a lease. Which means the loop has to close on a measurement of recall, not on a proxy for effort. That is unglamorous. It involves a test at the end of a passage, which nobody enjoys, rather than a sensor, which everybody demos. It is also the only version of adaptive pacing that can be checked.

There is a further reason to prefer it: being tested on material produces better long-term retention than spending the same time re-reading it, and the effect is large and heavily replicated. So the measurement that closes the loop is not a tax on the session. It is one of the more useful things in it. An adaptive system built on recall gets its control signal and improves retention with the same action, which is a rare shape for a design to have.

Where our own version falls short

We should be specific about what we have, because it is less than the argument above deserves. The paid tier includes adaptive ramping, and what that means is modest: pace is raised gradually across sessions rather than in response to a live model of your understanding. The recall check is a recognition test — it asks which words genuinely appeared and mixes them with plausible distractors — and recognition is a weak instrument for whether you followed an argument or would notice a contradiction three pages later. A control loop is only as good as the thing it measures, and ours measures a shadow of the target.

There is also a subtler problem we have not solved and do not want to paper over. Closing a loop on recall requires enough tests to distinguish a rate effect from a bad afternoon, and readers will not sit fifty tests to calibrate a slider. Any real system has to work from sparse, noisy, self-selected measurements, and be honest that a single score on a single passage is a data point rather than a finding — the same caution we apply to the research we cite.

What we would actually build first

Not the headband. The first honest version of adaptive pacing uses signals you already generate: rewind density, dwell overruns, where a session stopped, and recall scores across passages of comparable difficulty. It produces one output that is not a number climbing on a chart but a sentence — on material like this, your recall starts falling above roughly this rate — and it lets you decide what to do with it. That is closer to a reading instrument than to an optimiser, and an instrument is what a reader is better served by.

The version everyone else will build measures your pupils and tells you that you are getting faster. It will be more impressive to look at, and it will be right about the pupils. Whether it is right about the reading is a question its own design cannot ask, because a system that never measures comprehension is free to assume it. That freedom is precisely what makes the demo so good, and it is why the boring loop is the one we would rather ship.