V …

How the corpus was built

Every survey that went into the model, every survey that was thrown out and why, what the parser found inside them, and the one threshold the label builder had to choose. Counts are from the corpus build, not recomputed here — see the note under each panel for the code that produced them.

1 · Corpus reader — lineage funnel

Surveys in → excluded, reason-coded → usable.

Of the usable set

reader.py: window filter 0.5–180 min, ≥20 valid respondents to be usable. No clock cross-validation in this version — see the reader.py module docstring for why.

2 · Parser — type mix

Shares of all parsed questions. single is measured at 57.6%; the remaining shares were read from the build chart and are rounded to one decimal, so they sum to 100.0 by construction.

3 · Label builder — pause-threshold sensitivity

Share of items flagged as a pause at each threshold, median across all 13 timing surveys — the evidence for the chosen cutoff, not an assertion of it.

10x chosen: 3x is too aggressive (calls ordinary careful reading a break); 20x misses genuine interruptions; 10x is where breaks separate cleanly from reading time — 28 Aug.

4 · Feature set — deliberately excluded

Signals that were measured, considered, and left out on purpose. A feature absent because it was never tried and one absent because it was tested and rejected are different facts, and only the second is a decision.

    Each exclusion cites the run that settled it. Nothing here was dropped for being inconvenient — where a signal was left out because its definition could not be pinned down, that is what it says.