diff --git a/PROGRESS.md b/PROGRESS.md
index bce2dc5..d32cf65 100644
--- a/PROGRESS.md
+++ b/PROGRESS.md
@@ -177,6 +177,45 @@ Also fixed:
   semantic `query`.
 Earlier: notes/query recall, `/dash` monitoring, `wg_handshake` poller (NO-OP).
 
+### Gaps — why "voice assistant" is still aspirational (2026-07-05)
+
+What separates Maven today from the thing the spec describes. Dealbreakers
+first — these define the category:
+
+1. **No always-on listening.** Voice is push-to-talk in a web page: open the
+   PWA, hold, speak, release. No wake word, no ambient mic, no room hardware.
+   Today it's a dictaphone with a brain, not something you talk to across the
+   kitchen. Highest-leverage gap — a wake word + speaker device changes what
+   Maven *is*, not just how good it is. (Listening modes 2–3 in maven.md.)
+2. **No conversation.** The router classifies one utterance → one reply. No
+   follow-ups, no anaphora ("а когда?"), no dialogue state beyond the 90s
+   destructive-confirm park. The sub-1B phraser only words replies; nothing
+   converses.
+3. **Latency/shape of a turn.** Clip-based STT (record → upload → whisper →
+   route → phrase → piper → play). No streaming either direction, no barge-in;
+   every exchange is a full round trip.
+
+Capability-class gaps — built but thin:
+
+4. **Act surface is a handful of argv allowlist entries.** propose→enable
+   works; until populated, "acts on your homelab" is mostly potential.
+5. **Query answers cover notes + a few built-ins.** No weather, no calendar
+   *querying* (calendar only gates nudges), no general knowledge; the
+   cheatsheet promises more than the router delivers.
+6. **Routing quality depends on the ONNX embedder being configured** — the
+   HashEmbedder floor makes RU recall lexical/weak; many commands fall to
+   "clarify".
+7. **Presence is effectively one signal** (page_heartbeat); desk_active is
+   still an undeployed script — "voice when near" routing runs on a guess.
+8. **Long-term memory per spec (obsidian → chroma) doesn't exist** — sqlite
+   brute-force cosine over notes is recall, not personalization. Persona
+   prompt and custom TTS voice (kami-picked, replaces the irina floor) are
+   still future items.
+
+Ops footnote: in the Docker deploy, voice-over-web needs mavend to bind its
+voice server on 0.0.0.0:9100 — unverified on the target host; until then the
+containers may only do the non-voice surfaces.
+
 ### Future / logged, not now
 
 Personality prompt; custom TTS voice training (kami-picked voice, replaces irina
diff --git a/cmd/mavweb/static/ui.css b/cmd/mavweb/static/ui.css
index dc628d2..e167c8c 100644
--- a/cmd/mavweb/static/ui.css
+++ b/cmd/mavweb/static/ui.css
@@ -53,6 +53,7 @@ button:hover{border-color:var(--accent)}
 button:disabled{opacity:.4;cursor:not-allowed}
 input[type=text]{font:inherit;font-size:.85rem;background:var(--bg);color:var(--fg);
   border:1px solid var(--border);border-radius:3px;padding:.25rem .4rem;max-width:100%}
+input[type=checkbox]{accent-color:var(--accent)}
 label{color:var(--soft)}
 
 /* inline feedback (/history revert, /tools enable, /auth/passkey) */