What we know, and what we are still guessing.
Attention estimation is an empirical claim, and ours is young. This page states where the evidence is thin, because a tool that misreads you should say so before you trust it.
Three things the data has already told us.
All from a small sample: one reader, two sessions. Enough to falsify a guess, nowhere near enough to fit a model. We say which is which.
Device class changes the signal basis, not just the numbers.
Applied to a phone session, desktop thresholds labelled 65% of an engaged reader's time as drifting or inactive. Retuning a constant was never going to fix that; the channels themselves behave differently.
On a large screen, interaction silence is blind rather than weak.
64% of sampled windows on an unfolded foldable held no scroll events at all, and the keyboard never fired once, because the reader had nothing to scroll. Silence there is an artefact of the layout, not evidence about the mind.
Our first mobile threshold was set too tight, and we can prove it.
45s sat at roughly the 97th percentile of observed engaged silence, against a maximum of 48.2s, so it fired on someone who was reading. We have not retuned it on one reader's data, and it stays marked provisional in the product until we can.
Four questions we can answer with data instead of opinion.
Each is computable from derived state alone. None of them require knowing what you read or wrote.
Recovery time after drift
How long until engagement returns, with an intervention versus without. The core claim of the whole instrument.
False-positive rate on drift
How often we interrupt someone who was fine. The costs are asymmetric: a missed drift is cheap, a wrong intervention is insulting.
Distribution of engaged silence
Per device class: the actual basis for a threshold, replacing the numbers we currently defend as provisional.
Does directing outperform occupying?
An ambient bed responds; a pacer leads. Given the asymmetry above, leading may read as coercive on a false positive. Measurable, so we would rather not argue it.
Before the numbers, we want the words.
Open answers, in your language rather than a survey’s. Where focus breaks, what pulls you away, what has and hasn’t worked. The answers shape what gets built and how it gets explained, and we publish what we learn back here.
No pitch at the end.