# YazSes v2.10.0-dev.4 — Wave N: WordFind + LoadGuard (developer preview)

**Pre-release · not published** (publish workflows are pinned to `v0.x`/`v1.x`; v2 dev tags do
not ship). **v1.4.1 remains the stable release.** Every feature is **OFF by default**.

Continues Wave N (`design/vision/v2-research/16-wave-n.md`) with the word-finding &
cognitive-load tier.

## New in dev.4 (all off by default)

- **WordFind** (ADR-v2-118, `[wordfind]`) — an offline reverse dictionary for tip-of-the-tongue
  moments: "the word for when water turns to gas" → a ranked shortlist (evaporation, …).
  `rank_candidates` scores lexicon entries by stopword-filtered content-word overlap with your
  spoken description, normalized and deterministically tie-broken. A small demo lexicon ships
  in-tree; your own entries merge over it, and WordNet-scale lexicons / embedding rerankers stay
  optional extras. Anomia/AAC-friendly. Anchor: WordNet reverse-dictionary; TREC 2025
  Tip-of-the-Tongue.
- **LoadGuard** (ADR-v2-121, `[loadguard]`) — cognitive-load-aware guardrails. `estimate_load`
  fuses the load signals the pipeline already computes (pause ratio, filler rate, speech-rate
  slowdown, self-corrections) into a 0..1 estimate; `guard_policy` maps it to `normal` /
  `elevated` (widen confirmations) / `high` (also defer risky, destructive actions). So a
  mis-fired destructive voice command is hardest to trigger exactly when you are most overloaded.
  Anchor: cognitive-load-from-speech literature; ADHD guidance.

`yazses features` now lists **133 capabilities**.

## Wave N remaining
Echo (own-audio replay) · SRPace (screen-reader pacing) — the final tier completes Wave N.

## Quality
1445 tests green across Linux × macOS × Windows on Python 3.11 and 3.12. Both new cores are
unit-tested to 100% with **no new dependency** (pure overlap ranker + pure signal fusion; the
demo lexicon and per-user baselines are later/optional tiers). The default install and the v1
dictation path are unchanged.