Selection (access-core; SPEC §5.4, proposed): a prior over the options on the
board — usage frequency now, the Laya decision model later — times the
person's own switch-timing likelihood, after Nomon (Broderick & MacKay):
- press_timing.rs learns a damped kernel-density estimate of each press's
offset from the intended highlight, from a population prior scaled to the
person's scan interval, and only from selections the person keeps: presses
are staged per attempt, dropped when abandoned or undone (including an Undo
reached by scanning), and learned only after three newer selections.
- selection.rs computes the posterior (late presses explained as late, a
2% lapse term), selects at 99:1 odds with the prior capped so presses must
favour the choice 4:1, and otherwise keeps scanning — a row gets up to two
extra passes. Without opt-in, or before 10 kept presses, it behaves exactly
like the conventional scanner and only measures.
- scan_order.rs orders rows, groups and cells by prior mass (fewest expected
steps) without moving any; options keep stable position-order ids, and a
new order applies from the next pass. dwell.rs weights per-cell dwell from
0.8x to 1.5x within hard bounds. Every adaptation is opt-in and previewable.
Conformance C8–C11 (order never moves a position, opt-in required, dwell
bounds, selection follows presses): 370/370 with the original 236.
Prediction (aac-core): per-language packs — word frequencies, word pairs,
sentence starts and a letter model, front-coded and CRC-checked — built
reproducibly by tools/lang/build_packs.py from pinned wordfreq 3.2.0
(CC-BY-SA-4.0), FrequencyWords 2018 (CC-BY-SA-4.0) and Common Voice sentences
(CC0), into data/lang/ (43 of 45 languages; gu and kn have no open source
large enough and report "adaptive only"). PredictionEngine blends a pack with
what the person has said (hierarchical Dirichlet smoothing): next word,
completion, and letter prediction for letter scanning, deterministically.
A per-language pair filter exists; its lists still need curating before the
app shows next-word suggestions to children.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
edcf02b350