Merge commit '74a7088' into overnight-jul31

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ

AGENTS.md

@@ -43,14 +43,17 @@ notes. Without it, the floor `HashEmbedder` is used — deterministic but weak
 (Russian recall rarely clears the confidence gate, many commands fall to
 "clarify").
 
-**Download the embedder** (ONNX, ~90 MB):
+**Download the embedder** (ONNX, ~120 MB):
 
 ```sh
 make download-embedder
 ```
 
-This fetches `paraphrase-multilingual-MiniLM-L12-v2` (384-dim, 12-layer,
+This fetches `multilingual-e5-small` (384-dim, 12-layer, Russian and English)
-supports 50+ languages including Russian) to `models/embedder/`.
+to `models/embedder/multilingual-e5-small/`. It is an asymmetric retrieval
+model: the code puts `query: ` in front of a question and `passage: ` in front
+of a stored note, which is how e5 was trained. The quantized file is the one
+that is downloaded, deployed and measured.
 
 **Also need ONNX Runtime** (`libonnxruntime.so`):
 
@@ -64,8 +67,8 @@ sudo cp onnxruntime-linux-x64-1.15.1/lib/libonnxruntime.so* /usr/local/lib/
 ```json
 "voice": {
   "embedder": {
-    "model_path": "models/embedder/model_quantized.onnx",
+    "model_path": "models/embedder/multilingual-e5-small/model_quantized.onnx",
-    "tokenizer_path": "models/embedder/tokenizer.json",
+    "tokenizer_path": "models/embedder/multilingual-e5-small/tokenizer.json",
     "lib_path": "/usr/local/lib/libonnxruntime.so"
   }
 }

MODEL-BAKEOFF-31-07-2026.md

@@ -0,0 +1,101 @@
+# Resident model bake-off — 31-07-2026
+
+**Recommendation: keep Qwen3.5-0.8B.** LFM2.5-1.2B is worse at routing (52.6% vs 60.5%
+intent accuracy), and the loss is almost entirely Russian (18/61 vs 22/61 RU, while EN is a
+wash). It is also 2.4× slower. The Thinking variant is far worse again.
+
+Settles Vikunja **#278 / #250**.
+
+- Same fixture and scorer as `ROUTING-EVAL-31-07-2026.md`: `internal/router/eval/`
+  (`ru_routing_v1.json`, 76 held-out cases).
+- Reproduce: `MAVEN_LLM_URL=http://127.0.0.1:<port> make eval-router`
+  (`TestLLMRouterBaseline`). Note: there is no `make eval-models` target.
+- All three models served by the same `llama-server` flags — `-c 2048 -ngl 99 -t 6`, only
+  `-m` and `--port` differ. One server at a time on an otherwise idle box, so latencies are
+  real and not contention.
+- Measured on top of the router prompt fix (`origin/overnight/router-prompt` merged in), so
+  the Qwen column is directly comparable to the numbers already recorded.
+
+## Results
+
+`llm-only` — the model alone. This is the column that measures the model.
+
+| | Qwen3.5-0.8B | LFM2.5-1.2B Instruct | LFM2.5-1.2B Thinking |
+|---|---|---|---|
+| **intent-only accuracy** | **60.5%** | 52.6% | 36.8% |
+| full accuracy (intent+slots+gate) | **36.8%** | 32.9% | 21.1% |
+| **RU** | **22/61** | 18/61 | 10/61 |
+| EN | 6/15 | **7/15** | 6/15 |
+| route errors | 0 | 0 | 0 |
+| **p50 / p95 latency** | **1.05s / 1.71s** | 2.47s / 3.62s | 2.42s / 3.24s |
+| missed clarify | 6 / 6 | 6 / 6 | 6 / 6 |
+
+`cascade+llm` — stage-0 → model → classifier floor, what #320 would actually ship. Same
+ordering.
+
+| | Qwen3.5-0.8B | LFM2.5-1.2B Instruct | LFM2.5-1.2B Thinking |
+|---|---|---|---|
+| intent-only accuracy | **61.8%** | 55.3% | 38.2% |
+| full accuracy | **46.1%** | 42.1% | 30.3% |
+| RU / EN | **27/61** / 8/15 | 23/61 / **9/15** | 15/61 / 8/15 |
+| route errors | 0 | 0 | 0 |
+| p50 / p95 latency | **1.28s / 1.94s** | 2.18s / 2.72s | 2.27s / 3.19s |
+
+Full logs: the three runs are archived in the session scratchpad
+(`qwen08.txt`, `lfm-instruct.txt`, `lfm-thinking.txt`).
+
+## Russian-specific failures — the owner's worry is confirmed
+
+LFM2.5's Russian loss is not spread out. It has one large, specific failure: **it hears
+almost any Russian imperative or short phrase as `reminder`.**
+
+- `перезапусти докер` → reminder (want act)
+- `включи вытяжку` → reminder (want act)
+- `закрой жалюзи` → reminder (want act)
+- `заметка: продлить домен в августе` → reminder (want note)
+- `запиши что кран на кухне снова капает` → reminder (want note)
+- `доброе утро` → reminder (want chat)
+- `спасибо тебе` → reminder (want note/chat)
+- `переходи в тихий режим` → reminder (want system)
+
+That is `note→reminder ×4`, `act→reminder ×4`, `chat→reminder ×2` in one run. Qwen's
+equivalent failure axis is `query→fact ×8`, which is a narrower and already-understood bug.
+
+Two more Russian-side problems worth naming:
+
+1. **Fact keys come back empty or wrong in Russian.** `воды попил наконец`, `поужинал`,
+   `поспал часов пять` and `отметь что я позавтракал овсянкой` all returned an empty key.
+   `сходил в душ` and `отдохнул минут двадцать` both returned `water`. Qwen does not do this.
+2. **It leaked German.** `slept about seven hours` produced the fact key
+   `"7 Stunden geschlafen"`. Grammar-valid, semantically garbage — a sign the multilingual
+   mix is not anchored where Maven needs it.
+
+The claimed tool-calling advantage did not show up here. `act` is the closest thing this
+fixture has to a tool call, and LFM2.5 got it wrong more often than Qwen, mostly by calling
+it a reminder. It also produced no `fn` slot on any act, same as Qwen.
+
+## The Thinking variant
+
+Not viable. 36.8% intent accuracy, 10/61 Russian, and no latency saving over Instruct — the
+thinking trace costs time without buying accuracy on a short enum classification. With the
+`enable_thinking=false` diagnostic it collapsed further to 28.9% with 2 route errors
+(`query→reminder ×12`). Do not pursue.
+
+## Notes
+
+- Nothing crashed, nothing ignored the GBNF grammar, and no model produced unparseable JSON
+  in the shippable configurations. Zero route errors for both Instruct and Thinking in
+  `llm-only` and `cascade+llm`. The problem with LFM2.5 is what it decides, not whether it
+  can emit the contract.
+- The `6 / 6` missed clarify is unchanged across all three models. No model fixes the missing
+  refusal lane — that is `Confidence: 1.0` hardcoded in `llmrouter.go` (Vikunja #359), not a
+  model property.
+- The report labels every configuration `(0.8B)`; that string is hardcoded in the test, not a
+  reflection of which gguf was loaded. Model identity was confirmed per run via `/v1/models`.
+- No Go code was changed for this measurement, and no bug was found that needed one.
+
+## What this does not settle
+
+Routing only. LFM2.5 might still phrase better, and phrasing is the resident model's other
+job — that needs its own fixture. But routing is the load-bearing path and Maven is
+Russian-first, so on the evidence here the switch is not worth making.

Makefile

@@ -16,7 +16,7 @@ PIPER_BIN     := $(shell pwd)/deps/piper/piper
 PIPER_MODEL   := $(shell pwd)/models/tts/ru_RU-irina-medium.onnx
 PIPER_ESPEAK  := $(shell pwd)/deps/piper/espeak-ng-data
 
-.PHONY: all build build-stt build-tts build-daemon build-client build-waked build-web build-poll build-caldav clean test fmt-check vet run-stt run-tts run-web download-embedder deps-go eval-router eval-recall
+.PHONY: all build build-stt build-tts build-daemon build-client build-waked build-web build-poll build-caldav clean test fmt-check vet run-stt run-tts run-web download-embedder deps-go eval-router eval-recall eval-phrasing eval-models
 
 all: build
 
@@ -103,6 +103,30 @@ eval-router:
 eval-recall:
 	MAVEN_ONNX_LIB="$(MAVEN_ONNX_LIB)" $(GO) test -v -count=1 ./internal/memory/recalleval/
 
+# eval-phrasing -- score nudge phrasing (internal/phraser/eval). Verbose so the
+# report and every generated message land in the terminal. With no environment
+# it scores the deterministic Stub only, which is what CI runs. Set
+# MAVEN_LLM_URL to add the resident model:
+#   MAVEN_LLM_URL=http://127.0.0.1:18099 make eval-phrasing
+# The model run is slow (minutes) -- the timeout is raised to match.
+eval-phrasing:
+	$(GO) test -v -count=1 -timeout 40m ./internal/phraser/eval/
+
+# eval-models — score ONE llama-server against the same fixture, for the
+# resident-model bake-off (#278, #250). Start a server with the gguf you want,
+# then:
+#
+#   make eval-models MAVEN_LLM_URL=http://127.0.0.1:18100
+#
+# The report names carry the model llama-server reports, so runs from two
+# checkpoints stay apart. Only the LLM test runs — the classifier baselines do
+# not depend on the model and take the ONNX runtime with them.
+MAVEN_LLM_URL ?= http://127.0.0.1:18099
+
+eval-models:
+	MAVEN_LLM_URL="$(MAVEN_LLM_URL)" $(GO) test -v -count=1 -timeout 60m \
+		-run TestLLMRouterBaseline ./internal/router/eval/
+
 run-stt: build-stt
 	LD_LIBRARY_PATH="$(shell pwd)/deps/lib" \
 		./mavsttd -socket /tmp/maven/stt.sock -model $(WHISPER_MODEL)
@@ -128,9 +152,13 @@ deps-piper:
 		-o /tmp/piper.tar.gz
 	tar -xzf /tmp/piper.tar.gz -C deps/
 
-EMBEDDER_DIR := $(shell pwd)/models/embedder
+# multilingual-e5-small: an asymmetric retrieval model. It is trained to match
-EMBEDDER_MODEL_URL := https://huggingface.co/Xenova/paraphrase-multilingual-MiniLM-L12-v2/resolve/main/onnx/model_quantized.onnx
+# a short question against a longer passage, which is what note recall is.
-EMBEDDER_TOKENIZER_URL := https://huggingface.co/Xenova/paraphrase-multilingual-MiniLM-L12-v2/resolve/main/tokenizer.json
+# The quantized file is the one we download, deploy and measure — see
+# RECALL-EVAL-31-07-2026.md.
+EMBEDDER_DIR := $(shell pwd)/models/embedder/multilingual-e5-small
+EMBEDDER_MODEL_URL := https://huggingface.co/Xenova/multilingual-e5-small/resolve/main/onnx/model_quantized.onnx
+EMBEDDER_TOKENIZER_URL := https://huggingface.co/Xenova/multilingual-e5-small/resolve/main/tokenizer.json
 
 download-embedder:
 	mkdir -p $(EMBEDDER_DIR)

PHRASING-EVAL-31-07-2026.md

@@ -0,0 +1,138 @@
+# Phrasing evaluation — 31-07-2026
+
+How Maven words a nudge, measured instead of argued. Counterpart to
+`ROUTING-EVAL-31-07-2026.md`.
+
+- Fixture + scorer: `internal/phraser/eval/` (`nudges_v1.json`, 15 cases; `eval.go`, `checks.go`)
+- Reproduce: `MAVEN_LLM_URL=http://127.0.0.1:18099 make eval-phrasing`
+- Model: Qwen3.5-0.8B Q4_K_M, the resident model. Not swapped.
+- Commit: `a40bc55` (prompt fix)
+
+Every check is a string or length test a human can read and disagree with. No model
+grades another model here.
+
+## Result
+
+| | before | after |
+|---|---|---|
+| **cases passing every check** | **0/15** | **13/15** |
+| mood in enum | 6/15 | 15/15 |
+| Russian | 2/15 | 14/15 |
+| length (≤120 chars, ≤16 words) | 13/15 | 15/15 |
+| feminine self-reference | 15/15 | 15/15 |
+| no cringe | 13/15 | 15/15 |
+| on topic | 6/15 | 13/15 |
+| p50 latency | 11.4s | 11.4s |
+
+Latency did not move and is not good. 11s to word one nudge on this box.
+
+## The bug reproduced
+
+Yes, exactly as reported. 7 of 15 messages were the literal string `"..."`, and one was
+`"full voice message"`. Both are text copied straight out of the prompt.
+
+The system prompt said:
+
+```
+Respond ONLY with valid JSON: {"response": "full voice message", "mood": "neutral"}
+```
+
+and the user prompt said:
+
+```
+Respond as JSON: {"response": "...", "mood": "..."}
+```
+
+A 0.8B does not read `"..."` as "put your answer here". It reads it as the answer. The
+prompt was a worked example whose worked part was blank, so the model filled the slot by
+copying. This is the whole of finding 1.
+
+## What else was wrong
+
+Four separate faults, all prompt-side:
+
+1. **Placeholder echo** (7 cases) — above.
+2. **Wrong language** (13/15 failed the language check). The prompt was entirely English
+   and said "in the user's language (Russian or English)". The model picked English. It is
+   never English: the nudge is spoken by a Russian piper voice.
+3. **Rule names are English identifiers.** `netdata_critical`, `service_down`, `break` went
+   into the prompt raw. The model cannot nudge about a topic it has not been told in words,
+   so 9/15 were off topic. The daemon knows what its own rules mean; now it says so.
+4. **Mood invented** (`"warm"`, twice). The enum was listed in a parenthesis at the end of
+   an English sentence. Now it is its own line: "ровно одно из: neutral, happy, thinking,
+   tired, confused."
+
+Plus two non-prompt faults the run exposed:
+
+- **The no-parse fallback was English.** When the model returned nothing usable, the body
+  became `fmt.Sprintf("%s — %s", rule, sev)` — `"water — care"` — and that string went to
+  a Russian TTS. Now it falls back to plain Russian.
+- **Durations were English.** `humanDur` returns "3 hours"; it was landing verbatim inside
+  Russian sentences. Nudges now use a Russian formatter.
+
+## Three iterations, and what each taught
+
+| | score | change |
+|---|---|---|
+| baseline | 0/15 | — |
+| iter 1 | 2/15 | Russian prompt, filled-in examples, Russian durations |
+| iter 2 | 11/15 | required keyword per rule, one example instead of five, Russian fallback |
+| iter 3 | **13/15** | examples moved to topics that are not rules |
+
+The interesting step is 1 → 2. Fixing the placeholder did not fix the disease, it moved it:
+the model stopped copying `"..."` and started copying my first example instead. Five nudges
+in a row came back as `"Ты не пил воду три часа. Налей стакан."` regardless of the rule.
+
+**A small model copies the nearest concrete text in its prompt.** That is one failure mode
+with two symptoms. The fix that stuck was making the examples about laundry and a laptop
+battery — topics no rule ever produces, so copying them is visible in the score rather than
+invisibly passing the water cases.
+
+## Do not oversell 13/15
+
+Seven of the thirteen passes are the **deterministic fallback**, not the model:
+`"Напоминаю: таблетки."`, `"Сервис не отвечает."`, `"Критический алярм: проверь диск."`,
+`"Ты давно не пил воду."`. Those are strings this commit added to Go. The model returned
+nothing parseable and the fallback scored.
+
+So the honest reading is roughly **6/15 from the model, 7/15 from a fallback, 2/15 failing**.
+The prompt fix is real — `"..."` is nearly gone and the language and mood checks are clean —
+but a large part of the jump is that failure now degrades into Russian instead of into
+`"water — care"`. That is a genuine improvement for the operator and a weak one for the model.
+
+The two remaining failures: one `"..."` recurrence (`routine-stretch`) and one meal nudge
+that never says food.
+
+## Broken, found, not fixed
+
+1. **`checkFeminine` only catches half the constraint.** It scans for masculine
+   self-reference and passed 15/15 both runs — but three messages address the *owner* in
+   the feminine: "ты давно не отдыхал**а**", "он не ел". The owner is a man. The check has
+   no second-person gender test, so this scores clean while being exactly the persona
+   failure the constraint exists to prevent. This is the most important gap in the harness.
+2. **Grammar is not checked at all, and it is bad.** `"Он не ел 11 дней"` (it was 11 hours),
+   `"Сонуждились 7 дней"` (not a word), `"Они забыли воду"` (wrong person entirely). Every
+   one of these passes all six checks. The fixture measures properties, not fluency, and at
+   0.8B fluency is the binding constraint.
+3. **Unit confusion.** The model turns hours into days about a third of the time. The
+   prompt now says "11 ч"; it reads it as days.
+4. **11s p50.** Unchanged and untouched here. A nudge the model takes eleven seconds to
+   word has missed its moment. Worth its own task.
+5. **The keyword hint is close to teaching to the test.** `ruleKeywords` names the word the
+   on-topic check looks for. It is defensible — the daemon genuinely knows its rule topics
+   and the model genuinely cannot infer them from `netdata_critical` — but the on-topic
+   number is softer than the others because of it.
+
+## Next steps
+
+1. **Add a second-person gender check** to `checks.go`. Finding 1 above. Until it exists the
+   feminine column means less than it looks like.
+2. **Decide whether the fallback should count as a pass.** Right now `Score` cannot tell a
+   model answer from a fallback. Either mark fallback bodies in `PhrasedNudge` or count them
+   in their own column. Without that, any future prompt change can score well by failing
+   more.
+3. **Attack the 11s.** Nudge phrasing is short and non-interactive; thinking off is the first
+   thing to try, as it was for routing (#376).
+4. **Re-measure when #122 lands.** The CPT'd Qwen3-1.7B is the target. 13/15 with seven
+   fallbacks is the floor it has to beat, and the fluency problems above are the ones a
+   bigger, Russian-trained checkpoint should actually fix.

RECALL-EVAL-31-07-2026.md

@@ -83,14 +83,157 @@ sqlite-backed `store.MemoryStore` and `memory.InMemoryStore` identically — bot
 (`internal/store/memory.go:64`) at ~150µs over 42 rows against a ~59ms query embed. An ANN index is
 not the problem to solve.
 
+## Re-measured after the embedder swap — 31-07-2026, later the same day
+
+Changed: `models/embedder/` is now **multilingual-e5-small** (quantized, 118MB), with `query: ` in
+front of a question and `passage: ` in front of a stored note (Vikunja #371). `deploy/mavend.json`
+and `make download-embedder` now name the same file, and it is the quantized one — that is what the
+column below measures (Vikunja #372). Everything else is unchanged: same fixture, same store, same
+0.55 gate. The old column is the baseline and is left as it was.
+
+| | recall+onnx, MiniLM (baseline) | recall+onnx, e5-small (new) |
+|---|---|---|
+| **recall@1** | 60.0% (15/25) | **72.0% (18/25)** |
+| recall@3 | 80.0% (20/25) | 84.0% (21/25) |
+| **answered after the 0.55 gate** | 48.0% (12/25) | **72.0% (18/25)** |
+| wrong note on top / tie on top | 10 / 0 | 7 / 0 |
+| ranked first, then silenced by the gate | 3 | 0 |
+| **false recall** | 1/5 (20%) | **5/5 (100%)** |
+| top-1 score when right, min / median | 0.559 / 0.678 | 0.791 / 0.857 |
+| top-1 when it must stay silent, median / max | 0.470 / 0.567 | 0.815 / 0.835 |
+| RU / EN / `hard` cases passed | 13/24 / 3/6 / 2/11 | 14/24 / 4/6 / 5/11 |
+| latency p50 / p95 / max | 59ms / 148ms / 194ms | 18ms / 37ms / 49ms |
+
+### What moved
+
+Ranking got better and got faster. Half the previously-unwinnable `hard` cases now pass (2/11 →
+5/11), the guitar note no longer beats the docker-logs note, and the gate stops silencing notes that
+already ranked first. The quantized e5 is also ~3x quicker than the fp32 MiniLM it replaces.
+
+### What got worse: the gate is now a no-op
+
+e5 packs every cosine into a narrow high band. Right-note scores start at 0.791; must-stay-silent
+scores reach 0.835. **The distributions still overlap, and now they overlap above the gate**, so
+0.55 admits everything and false recall goes from 1/5 to 5/5. The sweep:
+
+```
+gate 0.50–0.70: answered 18/25 (72%)  false recall 5/5
+gate 0.80:      answered 17/25 (68%)  false recall 4/5
+gate 0.90:      answered  0/25 ( 0%)  false recall 0/5
+```
+
+There is no value that keeps real recall and rejects made-up questions — same conclusion as before,
+now with a wider band and no room at all. `query_min_score` was left at 0.55 as instructed. **The
+recommendation is to leave it there and stop tuning it**: any number under ~0.79 is a no-op and
+anything above starts cutting real recall long before it stops the false ones. The fix is a margin
+gate (`top1 − top2 > δ`), next-steps item 3, which is now the top item.
+
+### The prefixes did not do the work
+
+A control run with both prefixes set to the empty string scored the **same** recall@1 (72%), a
+slightly better recall@3 (88%) and the same 5/5 false recall. So on this fixture the gain comes from
+the model, not from the `query:` / `passage:` split. The prefixes are kept because they are how e5
+was trained and the split is the right shape for the read path, but they are not worth defending on
+this evidence — a bigger fixture may say otherwise.
+
+### Stored vectors from the old model are now junk
+
+Cosine between a MiniLM vector and an e5 vector means nothing. Every row already in `notes` and in
+the vector memory table was written by the old model, so after this deploy they will score as noise
+against a new query. A live database needs every note and fact re-embedded before recall works at
+all. Filed as its own task.
+
+## Margin gate — 31-07-2026, third run
+
+Next-steps item 3, done. The absolute gate is replaced by a **margin gate**: answer only when the
+top hit beats the runner-up by more than delta (`top1 − top2 > δ`). Same fixture, same e5 embedder,
+same store as the run above. `internal/memory/gate.go` holds the check; both read paths call it
+(`cmd/mavend/recall.go` and the notes-RAG branch in `voice.go`). New knob `voice.query_min_margin`
+in `deploy/mavend.json`, default 0.008.
+
+### Why the absolute gate could not work, in one line of data
+
+The harness now prints the margin distributions, and they barely overlap where the raw scores
+overlap completely:
+
+| | top-1 score | margin (top1 − top2) |
+|---|---|---|
+| right note first (n=18) | min 0.810, median 0.862, max 0.890 | min 0.001, median 0.029, max 0.053 |
+| must stay silent (n=5) | min 0.795, median 0.815, max 0.835 | min 0.000, median 0.002, **max 0.019** |
+
+Four of the five must-be-silent cases have a margin at or under 0.002 — when there is nothing to
+recall, e5 finds several notes equally close and no clear winner. That is the signal the absolute
+score throws away.
+
+### The delta sweep
+
+Absolute gate held at 0.55 throughout.
+
+```
+delta 0.000: answered 18/25 (72%)  false recall 5/5
+delta 0.002: answered 17/25 (68%)  false recall 3/5
+delta 0.005: answered 17/25 (68%)  false recall 2/5
+delta 0.008: answered 17/25 (68%)  false recall 1/5   <- chosen
+delta 0.010: answered 15/25 (60%)  false recall 1/5
+delta 0.012: answered 14/25 (56%)  false recall 1/5
+delta 0.015: answered 12/25 (48%)  false recall 1/5
+delta 0.020: answered 11/25 (44%)  false recall 0/5
+delta 0.025: answered  9/25 (36%)  false recall 0/5
+delta 0.030: answered  8/25 (32%)  false recall 0/5
+delta 0.040: answered  4/25 (16%)  false recall 0/5
+delta 0.050: answered  2/25 ( 8%)  false recall 0/5
+delta 0.060: answered  0/25 ( 0%)  false recall 0/5
+```
+
+### Chosen: δ = 0.008
+
+It is the best point on the frontier, not a taste call. **0.008 dominates 0.010, 0.012 and 0.015
+outright** — same 1/5 false recall, 8 to 20 points more real recall. Everything below it buys recall
+back only by admitting more false recalls (0.005 → 2/5, 0.002 → 3/5). The next real improvement is
+0.020 at 0/5 false, and it costs 24 points of recall to get there.
+
+The brief's bar was "recall above 60% with false recall at 1/5 or better". 0.008 clears it with room:
+68% and 1/5.
+
+### Before / after
+
+| | absolute gate 0.55 (previous) | margin gate δ=0.008 |
+|---|---|---|
+| recall@1 (ranking, ungated) | 72.0% (18/25) | 72.0% (18/25) — unchanged, the gate does not rank |
+| **answered after the gate** | 72.0% (18/25) | **68.0% (17/25)** |
+| **false recall** | **5/5 (100%)** | **1/5 (20%)** |
+| fixture cases passed | 18/30 | **21/30** |
+
+Four false recalls removed for one real answer. That is the trade the spec asks for — she is not a
+guesser-of-truth. The one survivor is `en-pref-025` ("should i be offered wine"), which recalls a
+filler note at 0.796 with a 0.019 margin: the widest silent-case margin in the fixture, and it sits
+inside the real-recall range, so no delta removes it without taking real answers with it.
+
+### Does the absolute cutoff still earn its keep? Marginally — kept
+
+On this fixture with e5 it is a **no-op**: the lowest right-note score is 0.791, so 0.55 rejects
+nothing the margin does not already reject. It is kept for two reasons, neither glamorous. It still
+does real work for the hash embedder (its own sweep shows answers dropping from 16% to 0% between
+0.30 and 0.50), and it is the only thing standing between the user and a reply built from a store
+where everything is far away but one row happens to be a little less far — a near-empty database, or
+the stale-vector case below. Cheap insurance, no measured cost. If a later embedder makes it bite,
+the sweep is one command.
+
+### Caveat on the numbers
+
+Five must-be-silent cases is a thin basis for a 4-point decision. 1/5 and 2/5 differ by one case.
+The shape of the frontier is trustworthy — margins separate, absolute scores do not — but δ=0.008
+itself should be re-read off a bigger fixture (next-steps item 6) before anyone defends the third
+decimal.
+
 ## Next steps — ordered by value-to-risk; nothing here is a decision
 
 1. **Swap the embedder to `multilingual-e5-small` with `query:`/`passage:` prefixes.** One config
    change plus a prefix in `onnxembedder.go`, re-measurable in one command.
 2. **Re-run `make eval-recall`, then set the gate from the sweep** — not before. Any
    `query_min_score` picked against today's embedder describes a model on its way out.
-3. **Replace the absolute-score gate with a margin gate** (`top1 − top2 > δ`) — as the routing eval
+3. ~~**Replace the absolute-score gate with a margin gate**~~ — done, see the section above.
-   concluded, absolute cosine cannot see a flat distribution.
+   δ=0.008, false recall 5/5 → 1/5.
 4. **Delete or repair the dead `memStore` branch** at `voice.go:776` — search before the gate,
    gate it separately, or restrict it to facts and say so.
 5. **Add a mild time decay to ranking** — the newest statement of a preference is the true one.

ROUTING-EVAL-31-07-2026.md

@@ -40,6 +40,41 @@ model → classifier as failure floor.
 
 Never compare a hash-embedder run to an ONNX one.
 
+## Re-measured after the prompt fix
+
+The table above is the **baseline at commit `46259b4`**, kept as-is. The prompt fix (query
+tested before fact, plus `repeat_penalty` and a bounded grammar string) was then measured on
+an otherwise idle box — no other eval sharing llama-server, so these latencies are real
+rather than contention.
+
+| | llm-only (0.8B) | cascade+llm (0.8B) | llm-only, thinking off |
+|---|---|---|---|
+| **intent-only accuracy** | 48.7% → **61.8%** | 50.0% → **63.2%** | **67.1%** |
+| full accuracy (intent+slots+gate) | 23.7% → **38.2%** | 32.9% → **47.4%** | **42.1%** |
+| route errors | 2 → **0** | 0 → 0 | **0** |
+| p50 / p95 latency | **1.08s / 1.55s** | **1.04s / 1.53s** | **0.93s / 1.41s** |
+
+Three things this run settles:
+
+1. **The prompt fix holds.** An earlier contended run reported 60.5% / 36.8% for llm-only;
+   the quiet run gives 61.8% / 38.2%. Close enough to call the gain real, and the earlier
+   run's 4-5s latency figures were contention, not the model.
+2. **`query→fact` fell from ×15 to ×7**, and both unparseable replies are gone. Zero route
+   errors in every LLM configuration.
+3. **`note→fact ×4` is real, not noise.** It shows up in the quiet run too. The agent that
+   wrote the prompt fix suspected its own change might have caused it by pulling assertive
+   `запиши что…` phrasings toward fact, and that suspicion stands — all five `ru-note-*`
+   cases now land on fact. Tracked as Vikunja #375.
+
+**Thinking off is the best configuration measured so far**, on both accuracy and latency
+(Vikunja #376). That is worth understanding before flipping: routing is a short
+classification into a fixed enum with grammar-constrained output, so there is little to
+reason about, and the thinking trace mostly gives a small model room to talk itself out of
+the right answer. Phrasing is a different job and needs measuring separately.
+
+Still `6 / 6` missed clarify — the router has no way to say "I don't know" (Vikunja #359).
+That is unchanged by anything here.
+
 ## Findings
 
 ### 1. The resident model does route better — 50.0% vs 36.8%

cmd/mavend/recall.go

@@ -8,16 +8,13 @@ import "github.com/kami/maven/internal/memory"
 // that can answer "when did I last …?" from a captured fact). A note hit here
 // is redundant with the notes-RAG path — by design; the two indexes can diverge
 // once the backend is swapped for a persistent/external store. ok=false when
-// there's no hit above the threshold or the hit carries no text.
+// the hit fails the confidence gate (see memory.Confident: an absolute floor
-func bestRecall(results []memory.Result, min float64) (string, bool) {
+// plus a margin over the runner-up) or carries no text.
-	if len(results) == 0 {
+func bestRecall(results []memory.Result, minScore, minMargin float64) (string, bool) {
+	if !memory.Confident(results, minScore, minMargin) {
 		return "", false
 	}
-	top := results[0]
+	text := results[0].Meta["text"]
-	if top.Score < min {
-		return "", false
-	}
-	text := top.Meta["text"]
 	if text == "" {
 		return "", false
 	}

cmd/mavend/recall_test.go

@@ -8,23 +8,24 @@ import (
 
 func TestBestRecall(t *testing.T) {
 	const min = 0.55
+	const margin = 0.008
 
 	t.Run("empty results", func(t *testing.T) {
-		if _, ok := bestRecall(nil, min); ok {
+		if _, ok := bestRecall(nil, min, margin); ok {
 			t.Error("empty results returned ok")
 		}
 	})
 
 	t.Run("top below threshold", func(t *testing.T) {
 		res := []memory.Result{{Score: 0.4, Meta: map[string]string{"text": "выпил воды"}}}
-		if _, ok := bestRecall(res, min); ok {
+		if _, ok := bestRecall(res, min, margin); ok {
 			t.Error("below-threshold hit returned ok")
 		}
 	})
 
 	t.Run("hit without text meta", func(t *testing.T) {
 		res := []memory.Result{{Score: 0.9, Meta: map[string]string{"type": "fact"}}}
-		if _, ok := bestRecall(res, min); ok {
+		if _, ok := bestRecall(res, min, margin); ok {
 			t.Error("textless hit returned ok")
 		}
 	})
@@ -34,7 +35,7 @@ func TestBestRecall(t *testing.T) {
 			{Score: 0.82, Meta: map[string]string{"text": "выпил воды в три часа", "type": "fact"}},
 			{Score: 0.60, Meta: map[string]string{"text": "другое"}},
 		}
-		got, ok := bestRecall(res, min)
+		got, ok := bestRecall(res, min, margin)
 		if !ok {
 			t.Fatal("clearing hit not returned")
 		}
@@ -42,4 +43,16 @@ func TestBestRecall(t *testing.T) {
 			t.Errorf("wrong text: %q", got)
 		}
 	})
+
+	// The runner-up is almost as close, so the embedder cannot tell the two
+	// notes apart. Silence beats reading back a coin flip.
+	t.Run("runner-up too close", func(t *testing.T) {
+		res := []memory.Result{
+			{Score: 0.860, Meta: map[string]string{"text": "выпил воды в три часа"}},
+			{Score: 0.858, Meta: map[string]string{"text": "другое"}},
+		}
+		if _, ok := bestRecall(res, min, margin); ok {
+			t.Error("thin-margin hit returned ok")
+		}
+	})
 }

cmd/mavend/replier_llm.go

@@ -29,7 +29,9 @@ func newLLMReplier(c completer) *llmReplier {
 	return &llmReplier{c: c, stub: voice.NewStubReplier()}
 }
 
-const replySystem = `Ты — Maven, домашняя ассистентка (о себе — в женском роде). Подтверди действие РОВНО ОДНИМ коротким предложением (≤120 символов), тепло и по-русски. Не задавай вопросов, не повторяй слова, не добавляй ничего после точки. Respond ONLY with valid JSON: {"response": "...", "mood": "neutral"}.`
+const replySystem = `Ты — Maven, домашняя ассистентка (о себе — в женском роде). Подтверди действие РОВНО ОДНИМ коротким предложением (≤120 символов), тепло и по-русски. Не задавай вопросов, не повторяй слова, не добавляй ничего после точки. Отвечай ТОЛЬКО одним объектом JSON с полями "response" (текст) и "mood" (ровно одно из: neutral, happy, thinking, tired, confused).
+Пример: {"response": "Записала, что ты выпил стакан воды.", "mood": "neutral"}
+Никогда не пиши "..." в поле response.`
 
 func (r *llmReplier) Reply(d router.Decision) string {
 	if d.Clarify {

cmd/mavend/voice.go

@@ -206,11 +206,11 @@ func wireVoice(cfg *config.Config, coreAPI ipc.CoreAPI, phr phraser.Phraser, mem
 	if threshold <= 0 {
 		threshold = config.DefaultRouterThreshold
 	}
-	// Both routing paths are weak on held-out utterances — the classifier gets
+	// The resident model routes by default: 63.2% of held-out intents right
-	// 36.8% of intents right, the resident model 50.0% and much slower. Off by
+	// against the classifier's 50.0%, at about 1s a turn instead of 30ms (see
-	// default (see config.VoiceConfig.LLMRouter); the classifier always stays
+	// config.VoiceConfig.LLMRouter). The classifier always stays wired as the
-	// wired as the fallback, so a model error never breaks a turn.
+	// fallback, so a model error never breaks a turn.
-	rtr := buildRouter(emb, matcher, threshold, pickLLMRouter(cfg.Voice.LLMRouter, llmClient))
+	rtr := buildRouter(emb, matcher, threshold, pickLLMRouter(cfg.Voice.UseLLMRouter(), llmClient))
 
 	// ----- sessions registry (shared with voicesink) -----
 	sessions := voice.NewSessions()
@@ -257,6 +257,7 @@ func wireVoice(cfg *config.Config, coreAPI ipc.CoreAPI, phr phraser.Phraser, mem
 		clarifyStore:     clarifyStore,
 		extractor:        router.Extractor{Time: timeParser, Acts: matcher, Facts: router.DefaultFactParser{}},
 		queryMinScore:    cfg.Voice.QueryMinScore,
+		queryMinMargin:   cfg.Voice.QueryMinMargin,
 		timeParser:       timeParser,
 		ecosystem:        eco,
 	}
@@ -300,6 +301,9 @@ type reactiveHandler struct {
 	// load-bearing math (same posture as the presence thresholds). Set by
 	// wireVoice from VoiceConfig; default 0.55.
 	queryMinScore float64
+	// queryMinMargin — the second half of that gate: how far the top hit must
+	// beat the runner-up. 0 ⇒ margin off.
+	queryMinMargin float64
 
 	// timeParser — used as a fallback for stage-0 reminder grammar matches
 	// (where the extractor didn't run). Shared with the router's extractor.
@@ -559,7 +563,7 @@ func (h *reactiveHandler) applyAction(ctx context.Context, dec router.Decision)
 		// fail the fact write). Facts aren't in the notes table, so this is the
 		// only recall path for them — "когда я пил воду?" reads back from here.
 		if h.memStore != nil {
-			if vec, err := h.embedder.Embed(ctx, dec.Utterance); err != nil {
+			if vec, err := router.EmbedPassage(ctx, h.embedder, dec.Utterance); err != nil {
 				log.Printf("voice: embed fact for memory: %v", err)
 			} else if err := h.memStore.Insert(ctx, "fact:"+dec.Slots.Key+":"+strconv.FormatInt(now.Unix(), 10), vec, map[string]string{
 				"source": "voice",
@@ -673,7 +677,7 @@ func (h *reactiveHandler) applyAction(ctx context.Context, dec router.Decision)
 		// embed the note text with the same model the classifier uses, persist
 		// via CoreAPI (source=tap:voice). Semantic recall lives in `notes`, not
 		// facts — no predicate reads it (spec's two-memory split).
-		vec, err := h.embedder.Embed(ctx, dec.Utterance)
+		vec, err := router.EmbedPassage(ctx, h.embedder, dec.Utterance)
 		if err != nil {
 			log.Printf("voice: embed note: %v", err)
 			return "не получилось сохранить заметку."
@@ -750,7 +754,7 @@ func (h *reactiveHandler) applyAction(ctx context.Context, dec router.Decision)
 			return fmt.Sprintf("в %s сейчас %.0f градусов, %s.", w.Location, w.Temperature, w.Condition)
 		}
 
-		vec, err := h.embedder.Embed(ctx, dec.Utterance)
+		vec, err := router.EmbedQuery(ctx, h.embedder, dec.Utterance)
 		if err != nil {
 			log.Printf("voice: embed query: %v", err)
 			return "не получилось найти ответ."
@@ -760,18 +764,23 @@ func (h *reactiveHandler) applyAction(ctx context.Context, dec router.Decision)
 			log.Printf("voice: query notes: %v", err)
 			return "не получилось найти ответ."
 		}
-		// Confidence gate: below threshold, say "I don't know" rather than read
+		// Confidence gate: below it, say "I don't know" rather than read back
-		// back the least-unrelated note — a confident wrong recall is worse than
+		// the least-unrelated note — a confident wrong recall is worse than a
-		// a gap (spec's "not a guesser-of-truth"). Same instinct as the loop's
+		// gap (spec's "not a guesser-of-truth"). Same instinct as the loop's
-		// since(key)==null → don't fire. Tuned for the ONNX embedder; the Hash
+		// since(key)==null → don't fire. Two parts: an absolute cosine floor,
-		// floor scores lexically and may rarely clear it.
+		// and a margin over the runner-up, which is the part that works with
-		if len(notes) == 0 || notes[0].Score < h.queryMinScore {
+		// the e5 embedder's narrow score band. See memory.Confident.
+		noteScores := make([]float64, len(notes))
+		for i, n := range notes {
+			noteScores[i] = n.Score
+		}
+		if !memory.ConfidentScores(noteScores, h.queryMinScore, h.queryMinMargin) {
 			// Long-term memory recall (notes + facts) before general knowledge:
 			// the notes table can't answer fact questions, but the memory store
 			// indexes both. Only runs when notes-RAG already gave up → additive.
 			if h.memStore != nil {
 				if hits, herr := h.memStore.Search(ctx, vec, 3); herr == nil {
-					if text, ok := bestRecall(hits, h.queryMinScore); ok {
+					if text, ok := bestRecall(hits, h.queryMinScore, h.queryMinMargin); ok {
 						return text
 					}
 				}

deploy/mavend.json

@@ -36,11 +36,13 @@
     "stt": { "socket": "/run/maven/stt.sock", "lang": "ru" },
     "tts": { "socket": "/run/maven/tts.sock", "lang": "ru" },
     "embedder": {
-      "model_path": "/opt/maven/models/embedder/model.onnx",
+      "model_path": "/opt/maven/models/embedder/multilingual-e5-small/model_quantized.onnx",
-      "tokenizer_path": "/opt/maven/models/embedder/tokenizer.json",
+      "tokenizer_path": "/opt/maven/models/embedder/multilingual-e5-small/tokenizer.json",
       "lib_path": "/opt/maven/lib/libonnxruntime.so"
     },
-    "llm_router": false,
+    "llm_router": true,
+    "query_min_score": 0.55,
+    "query_min_margin": 0.008,
     "tool_timeout": "30s",
     "tools": [
       { "name": "status",     "cmd": ["systemctl", "status"],         "scope": "homelab", "destructive": false },

internal/config/config.go

@@ -258,18 +258,25 @@ type VoiceConfig struct {
 	RouterThreshold float64 `json:"router_threshold,omitempty"`
 
 	// LLMRouter — route with the resident model instead of the embedding
-	// classifier. Measured on the held-out fixture (ROUTING-EVAL-31-07-2026.md)
+	// classifier. On by default since Vikunja #320.
-	// the model gets 50.0% of intents right against the classifier's 36.8%, but
-	// it costs about 800ms per turn instead of 30ms.
 	//
-	// TODO: the default stays false until two things land.
+	// Measured on the held-out fixture (ROUTING-EVAL-31-07-2026.md): 63.2% of
-	//   1. The LLM router cannot refuse. LLMRouter.Route hardcodes
+	// intents right against the classifier's 50.0%, and no route errors. It
-	//      Confidence: 1.0, so the stage-3 clarify gate never fires and an
+	// costs about 1s per turn instead of 30ms.
-	//      unclear utterance becomes a confident wrong action (Vikunja #359).
+	//
-	//   2. Extractor.Extract never runs on an LLM decision, so acts arrive with
+	// It is safe to leave on. The model can refuse — it answers "unknown" when
-	//      no Fn and reminders with no Time.
+	// it cannot route, and the turn drops to the classifier and its clarify
-	// Turning this on today makes routing more accurate and less safe.
+	// gate. Any LLM error does the same, so a turn never breaks on the model.
-	LLMRouter bool `json:"llm_router,omitempty"`
+	// Slot extraction runs on LLM decisions too, so acts get their Fn and
+	// reminders their Time.
+	//
+	// Set it false to go back to the classifier, e.g. on a box with no
+	// llama-server or when 1s a turn is too slow.
+	//
+	// It is a pointer so that "missing from the file" and "explicitly false"
+	// are different things: missing means on, false means off. Read it with
+	// UseLLMRouter(), not directly.
+	LLMRouter *bool `json:"llm_router,omitempty"`
 
 	// QueryMinScore — the note-recall confidence gate. Top cosine below this
 	// ⇒ "I don't know" instead of a guess. Tuned for the ONNX embedder (0.55);
@@ -277,6 +284,14 @@ type VoiceConfig struct {
 	// default if unset.
 	QueryMinScore float64 `json:"query_min_score,omitempty"`
 
+	// QueryMinMargin — the second half of the recall gate: the top hit must
+	// beat the runner-up by more than this. The absolute score above cannot do
+	// the job on its own, because the e5 embedder puts every cosine in one
+	// narrow high band, so a made-up question scores as high as a real one.
+	// The margin asks whether one note is clearly the best instead.
+	// Negative ⇒ off. 0 ⇒ the default below.
+	QueryMinMargin float64 `json:"query_min_margin,omitempty"`
+
 	// Persona — optional prompt prefix that tunes maven's character. Prepended
 	// to every LLM system prompt (nudge phrasing, note queries, general
 	// knowledge). Empty string ⇒ current hardcoded persona (feminine-gendered
@@ -404,7 +419,14 @@ const (
 	DefaultAutotuneInterval = 10 * time.Minute
 	DefaultRouterThreshold  = 0.55
 	DefaultQueryMinScore    = 0.55
-	DefaultToolTimeout      = 30 * time.Second
+	// Read off the margin sweep in internal/memory/recalleval on the e5
+	// embedder: 0.008 answers 68% of real questions (down from 72%) and cuts
+	// false recall from 5/5 to 1/5. Every larger delta costs real recall
+	// without removing that last one until 0.020, which drops recall to 44%.
+	DefaultQueryMinMargin = 0.008
+	DefaultToolTimeout    = 30 * time.Second
+	// DefaultLLMRouter — route with the resident model unless told otherwise.
+	DefaultLLMRouter = true
 
 	DefaultFactEnrichmentInterval = 30 * time.Second
 )
@@ -486,9 +508,21 @@ func (c *Config) applyDefaults() {
 		if c.Voice.QueryMinScore <= 0 {
 			c.Voice.QueryMinScore = DefaultQueryMinScore
 		}
+		// Unset ⇒ default. Negative is how you turn the margin off on purpose,
+		// so it is clamped to 0 rather than replaced by the default.
+		switch {
+		case c.Voice.QueryMinMargin == 0:
+			c.Voice.QueryMinMargin = DefaultQueryMinMargin
+		case c.Voice.QueryMinMargin < 0:
+			c.Voice.QueryMinMargin = 0
+		}
 		if c.Voice.ToolTimeout <= 0 {
 			c.Voice.ToolTimeout = Duration(DefaultToolTimeout)
 		}
+		if c.Voice.LLMRouter == nil {
+			on := DefaultLLMRouter
+			c.Voice.LLMRouter = &on
+		}
 	}
 
 	// routines: default severity to care-class (1) — the safe floor: a
@@ -507,6 +541,15 @@ func (c *Config) applyDefaults() {
 	}
 }
 
+// UseLLMRouter reports whether to route with the resident model. Unset means
+// on; only an explicit false in the config turns it off.
+func (v *VoiceConfig) UseLLMRouter() bool {
+	if v == nil || v.LLMRouter == nil {
+		return DefaultLLMRouter
+	}
+	return *v.LLMRouter
+}
+
 func (c *Config) validate() error {
 	if c.Phraser != nil {
 		if c.Phraser.ModelPath == "" {

internal/config/config_test.go

@@ -171,14 +171,26 @@ func TestWeatherConfigNilOK(t *testing.T) {
 	}
 }
 
-func TestLLMRouterDefaultsOff(t *testing.T) {
+func TestLLMRouterDefaultsOn(t *testing.T) {
 	p := writeConfig(t, `{"voice":{"enabled":true,"bind":"127.0.0.1:9100"}}`)
 	c, err := Load(p)
 	if err != nil {
 		t.Fatalf("Load: %v", err)
 	}
-	if c.Voice.LLMRouter {
+	if !c.Voice.UseLLMRouter() {
-		t.Error("voice.llm_router absent should mean false")
+		t.Error("voice.llm_router absent should mean on")
+	}
+}
+
+// Missing and explicitly false must not mean the same thing.
+func TestLLMRouterExplicitFalseTurnsItOff(t *testing.T) {
+	p := writeConfig(t, `{"voice":{"enabled":true,"bind":"127.0.0.1:9100","llm_router":false}}`)
+	c, err := Load(p)
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	if c.Voice.UseLLMRouter() {
+		t.Error("voice.llm_router false should turn it off")
 	}
 }
 
@@ -188,7 +200,7 @@ func TestLLMRouterRead(t *testing.T) {
 	if err != nil {
 		t.Fatalf("Load: %v", err)
 	}
-	if !c.Voice.LLMRouter {
+	if !c.Voice.UseLLMRouter() {
 		t.Error("voice.llm_router true was not read")
 	}
 }

internal/memory/gate.go

@@ -0,0 +1,38 @@
+package memory
+
+// Confidence gate for a recall. Two checks, both must pass before Maven says a
+// note back:
+//
+//   - minScore — an absolute cosine floor.
+//   - minMargin — the top hit must beat the runner-up by more than this.
+//
+// The margin is the one that carries the weight. The e5 embedder packs every
+// score into a narrow high band (0.79-0.89 on the recall fixture), so an
+// absolute floor cannot tell a real hit from a confident-looking miss: every
+// value under the band admits everything, every value above it answers nothing.
+// A margin asks a different question — "is this note clearly the best one, or
+// is the whole shelf equally close?" — and a made-up question has no clear best.
+//
+// With one hit and no runner-up there is nothing to compare, so only the floor
+// applies.
+
+// ConfidentScores reports whether the top score clears both gates. scores must
+// be sorted highest first. minMargin <= 0 turns the margin check off.
+func ConfidentScores(scores []float64, minScore, minMargin float64) bool {
+	if len(scores) == 0 || scores[0] < minScore {
+		return false
+	}
+	if minMargin > 0 && len(scores) > 1 && scores[0]-scores[1] <= minMargin {
+		return false
+	}
+	return true
+}
+
+// Confident is ConfidentScores for search results.
+func Confident(results []Result, minScore, minMargin float64) bool {
+	scores := make([]float64, len(results))
+	for i, r := range results {
+		scores[i] = r.Score
+	}
+	return ConfidentScores(scores, minScore, minMargin)
+}

internal/memory/gate_test.go

@@ -0,0 +1,45 @@
+package memory
+
+import "testing"
+
+func TestConfidentScores(t *testing.T) {
+	cases := []struct {
+		name      string
+		scores    []float64
+		minScore  float64
+		minMargin float64
+		want      bool
+	}{
+		{"no hits", nil, 0.55, 0.008, false},
+		{"below the floor", []float64{0.40, 0.10}, 0.55, 0.008, false},
+		{"clear winner", []float64{0.86, 0.70}, 0.55, 0.008, true},
+		{"runner-up too close", []float64{0.860, 0.858}, 0.55, 0.008, false},
+		// The rule is "beats the runner-up by MORE than delta". Not testing an
+		// exactly-equal margin: no pair of these decimals subtracts to exactly
+		// 0.008 in binary float, so such a test would pin rounding, not the rule.
+		{"margin just under delta", []float64{0.8079, 0.8}, 0.55, 0.008, false},
+		{"margin just over delta", []float64{0.8081, 0.8}, 0.55, 0.008, true},
+		// One hit: nothing to compare against, so only the floor applies.
+		{"single hit clears", []float64{0.86}, 0.55, 0.008, true},
+		{"single hit below floor", []float64{0.10}, 0.55, 0.008, false},
+		// Margin off — the old absolute-only behaviour.
+		{"margin off admits a tie", []float64{0.86, 0.86}, 0.55, 0, true},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			if got := ConfidentScores(c.scores, c.minScore, c.minMargin); got != c.want {
+				t.Errorf("got %v, want %v", got, c.want)
+			}
+		})
+	}
+}
+
+func TestConfidentReadsResultScores(t *testing.T) {
+	res := []Result{{ID: "a", Score: 0.86}, {ID: "b", Score: 0.858}}
+	if Confident(res, 0.55, 0.008) {
+		t.Error("thin margin passed the gate")
+	}
+	if !Confident(res, 0.55, 0) {
+		t.Error("margin off should fall back to the floor alone")
+	}
+}

internal/memory/recalleval/recalleval.go

@@ -117,18 +117,40 @@ type cachingEmbedder struct {
 	seen  map[string][]float32
 }
 
+var _ router.AsymmetricEmbedder = (*cachingEmbedder)(nil)
+
 func (c *cachingEmbedder) Dim() int     { return c.inner.Dim() }
 func (c *cachingEmbedder) Close() error { return nil } // the caller owns inner
 
 func (c *cachingEmbedder) Embed(ctx context.Context, text string) ([]float32, error) {
-	if v, ok := c.seen[text]; ok {
+	return c.cached(ctx, "embed:"+text, func() ([]float32, error) {
+		return c.inner.Embed(ctx, text)
+	})
+}
+
+// The two sides of an asymmetric embedder give different vectors for the same
+// string, so the cache key has to say which side asked.
+func (c *cachingEmbedder) EmbedQuery(ctx context.Context, text string) ([]float32, error) {
+	return c.cached(ctx, "query:"+text, func() ([]float32, error) {
+		return router.EmbedQuery(ctx, c.inner, text)
+	})
+}
+
+func (c *cachingEmbedder) EmbedPassage(ctx context.Context, text string) ([]float32, error) {
+	return c.cached(ctx, "passage:"+text, func() ([]float32, error) {
+		return router.EmbedPassage(ctx, c.inner, text)
+	})
+}
+
+func (c *cachingEmbedder) cached(_ context.Context, key string, embed func() ([]float32, error)) ([]float32, error) {
+	if v, ok := c.seen[key]; ok {
 		return v, nil
 	}
-	v, err := c.inner.Embed(ctx, text)
+	v, err := embed()
 	if err != nil {
 		return nil, err
 	}
-	c.seen[text] = v
+	c.seen[key] = v
 	return v, nil
 }
 
@@ -155,6 +177,8 @@ type Outcome struct {
 	Tied    bool
 	TopID   string
 	TopScor float64
+	// Margin — top1 − top2. 0 when fewer than two hits came back.
+	Margin  float64
 	Reasons []string
 }
 
@@ -162,8 +186,10 @@ type Outcome struct {
 // ranks first but is silenced by query_min_score is a threshold problem, and a
 // note that never ranks first is an embedder problem. Those are different fixes.
 type Report struct {
-	Name       string
+	Name     string
-	MinScore   float64
+	MinScore float64
+	// MinMargin — how far the top hit must beat the runner-up. 0 ⇒ off.
+	MinMargin  float64
 	Total      int
 	Answerable int
 	Rank1      int
@@ -190,9 +216,14 @@ type Report struct {
 	// where the right note ranked first, and for the no-answer cases. The gap
 	// between these two distributions is what a defensible query_min_score
 	// would have to sit inside; if they overlap, no threshold separates them.
-	CorrectTop    []float64
+	CorrectTop  []float64
-	NoAnswerTop   []float64
+	NoAnswerTop []float64
-	P50, P95, Max time.Duration
+	// CorrectMargin / NoAnswerMargin — the same two groups, but top1 − top2
+	// instead of top1. This is the pair the margin gate has to separate, and
+	// unlike the absolute scores it is what the sweep reads.
+	CorrectMargin  []float64
+	NoAnswerMargin []float64
+	P50, P95, Max  time.Duration
 }
 
 // TagStat — passed/total for one slice of the fixture.
@@ -223,13 +254,14 @@ func ratio(n, d int) float64 {
 // the run on an embed or search error: an erroring case scores as a miss and is
 // counted in Errors, because "the embedder was down" and "the embedder was
 // wrong" are different numbers.
-func Score(ctx context.Context, name string, emb router.Embedder, newStore NewStore, minScore float64, f Fixture) (Report, error) {
+func Score(ctx context.Context, name string, emb router.Embedder, newStore NewStore, minScore, minMargin float64, f Fixture) (Report, error) {
 	rep := Report{
-		Name:     name,
+		Name:      name,
-		MinScore: minScore,
+		MinScore:  minScore,
-		Total:    len(f.Cases),
+		MinMargin: minMargin,
-		ByTag:    map[string]TagStat{},
+		Total:     len(f.Cases),
-		ByLang:   map[string]TagStat{},
+		ByTag:     map[string]TagStat{},
+		ByLang:    map[string]TagStat{},
 	}
 	lat := make([]time.Duration, 0, len(f.Cases))
 
@@ -239,7 +271,7 @@ func Score(ctx context.Context, name string, emb router.Embedder, newStore NewSt
 		} else {
 			rep.NoAnswer++
 		}
-		o, err := scoreCase(ctx, emb, newStore, minScore, c, f.Filler)
+		o, err := scoreCase(ctx, emb, newStore, minScore, minMargin, c, f.Filler)
 		if err != nil {
 			return Report{}, err
 		}
@@ -263,6 +295,11 @@ func Score(ctx context.Context, name string, emb router.Embedder, newStore NewSt
 			} else if !o.Rank1 {
 				rep.WrongTop++
 			}
+			if o.Rank1 {
+				// Margins are collected on rank, not on the gate, so the
+				// distribution does not move as the sweep changes the gate.
+				rep.CorrectMargin = append(rep.CorrectMargin, o.Margin)
+			}
 			if o.Rank1 && o.Recalled != "" {
 				rep.CorrectTop = append(rep.CorrectTop, o.TopScor)
 			}
@@ -271,6 +308,7 @@ func Score(ctx context.Context, name string, emb router.Embedder, newStore NewSt
 				rep.FalseRecall++
 			}
 			rep.NoAnswerTop = append(rep.NoAnswerTop, o.TopScor)
+			rep.NoAnswerMargin = append(rep.NoAnswerMargin, o.Margin)
 		}
 
 		if o.Pass {
@@ -285,6 +323,8 @@ func Score(ctx context.Context, name string, emb router.Embedder, newStore NewSt
 
 	sort.Float64s(rep.CorrectTop)
 	sort.Float64s(rep.NoAnswerTop)
+	sort.Float64s(rep.CorrectMargin)
+	sort.Float64s(rep.NoAnswerMargin)
 	sort.Slice(lat, func(i, j int) bool { return lat[i] < lat[j] })
 	rep.P50, rep.P95 = percentile(lat, 0.50), percentile(lat, 0.95)
 	if len(lat) > 0 {
@@ -296,7 +336,7 @@ func Score(ctx context.Context, name string, emb router.Embedder, newStore NewSt
 // scoreCase inserts the case's notes into a fresh store, then runs the read
 // path the daemon runs. The returned error is fatal (the harness is broken);
 // an embedder or store failure on the query lands in Outcome.Err instead.
-func scoreCase(ctx context.Context, emb router.Embedder, newStore NewStore, minScore float64, c Case, filler []StoredNote) (Outcome, error) {
+func scoreCase(ctx context.Context, emb router.Embedder, newStore NewStore, minScore, minMargin float64, c Case, filler []StoredNote) (Outcome, error) {
 	st, release, err := newStore()
 	if err != nil {
 		return Outcome{}, fmt.Errorf("%s: new store: %w", c.ID, err)
@@ -305,7 +345,7 @@ func scoreCase(ctx context.Context, emb router.Embedder, newStore NewStore, minS
 
 	all := append(append([]StoredNote(nil), c.Notes...), filler...)
 	for _, n := range all {
-		vec, err := emb.Embed(ctx, n.Text)
+		vec, err := router.EmbedPassage(ctx, emb, n.Text)
 		if err != nil {
 			return Outcome{}, fmt.Errorf("%s: embed note %s: %w", c.ID, n.ID, err)
 		}
@@ -317,7 +357,7 @@ func scoreCase(ctx context.Context, emb router.Embedder, newStore NewStore, minS
 
 	o := Outcome{Case: c}
 	start := time.Now()
-	qvec, err := emb.Embed(ctx, c.Query)
+	qvec, err := router.EmbedQuery(ctx, emb, c.Query)
 	if err != nil {
 		o.Latency = time.Since(start)
 		o.Err = err
@@ -335,7 +375,10 @@ func scoreCase(ctx context.Context, emb router.Embedder, newStore NewStore, minS
 
 	if len(hits) > 0 {
 		o.TopID, o.TopScor = hits[0].ID, hits[0].Score
-		o.Recalled = bestRecall(hits, minScore)
+		if len(hits) > 1 {
+			o.Margin = hits[0].Score - hits[1].Score
+		}
+		o.Recalled = bestRecall(hits, minScore, minMargin)
 	}
 	for i, h := range hits {
 		if h.ID != c.Want {
@@ -356,14 +399,14 @@ func scoreCase(ctx context.Context, emb router.Embedder, newStore NewStore, minS
 	switch {
 	case !c.Answerable():
 		if o.Recalled != "" {
-			o.Reasons = append(o.Reasons, fmt.Sprintf("false recall: %q at %.3f, want silence", o.TopID, o.TopScor))
+			o.Reasons = append(o.Reasons, fmt.Sprintf("false recall: %q at %.3f (margin %.3f), want silence", o.TopID, o.TopScor, o.Margin))
 		}
 	case o.Tied:
 		o.Reasons = append(o.Reasons, fmt.Sprintf("tie at %.3f — the right note is on top only by sort order", o.TopScor))
 	case !o.Rank1:
 		o.Reasons = append(o.Reasons, fmt.Sprintf("top hit %q (%.3f), want %q%s", o.TopID, o.TopScor, c.Want, rankNote(o.Rank3)))
 	case o.Recalled == "":
-		o.Reasons = append(o.Reasons, fmt.Sprintf("right note ranked first but scored %.3f < gate %.2f — daemon says \"не знаю\"", o.TopScor, minScore))
+		o.Reasons = append(o.Reasons, fmt.Sprintf("right note ranked first at %.3f (margin %.3f) but the gate silenced it — daemon says \"не знаю\"", o.TopScor, o.Margin))
 	}
 	o.Pass = len(o.Reasons) == 0
 	return o, nil
@@ -379,8 +422,8 @@ func rankNote(inTop3 bool) string {
 // bestRecall mirrors cmd/mavend/recall.go — the gate the daemon actually
 // applies to a memory hit. Duplicated rather than imported because package main
 // is not importable; recalleval_test.go asserts the two agree in behaviour.
-func bestRecall(results []memory.Result, min float64) string {
+func bestRecall(results []memory.Result, minScore, minMargin float64) string {
-	if len(results) == 0 || results[0].Score < min {
+	if !memory.Confident(results, minScore, minMargin) {
 		return ""
 	}
 	return results[0].Meta["text"]
@@ -415,7 +458,7 @@ func percentile(sorted []time.Duration, p float64) time.Duration {
 // the slices that name where the path is weak.
 func (r Report) String() string {
 	var b strings.Builder
-	fmt.Fprintf(&b, "%s: %d/%d cases pass (gate %.2f)\n", r.Name, r.Passed, r.Total, r.MinScore)
+	fmt.Fprintf(&b, "%s: %d/%d cases pass (gate %.2f, margin %.3f)\n", r.Name, r.Passed, r.Total, r.MinScore, r.MinMargin)
 	fmt.Fprintf(&b, "  recall@1 %.1f%% (%d/%d)  recall@3 %.1f%% (%d/%d)  answered after gate %.1f%% (%d/%d)\n",
 		100*r.Recall1(), r.Rank1, r.Answerable,
 		100*r.Recall3(), r.Rank3, r.Answerable,
@@ -426,6 +469,8 @@ func (r Report) String() string {
 		100*r.FalseRecallRate(), r.FalseRecall, r.NoAnswer)
 	fmt.Fprintf(&b, "  top-1 score, right note first: %s\n", spread(r.CorrectTop))
 	fmt.Fprintf(&b, "  top-1 score, must be silent:   %s\n", spread(r.NoAnswerTop))
+	fmt.Fprintf(&b, "  margin top1-top2, right note first: %s\n", spread(r.CorrectMargin))
+	fmt.Fprintf(&b, "  margin top1-top2, must be silent:  %s\n", spread(r.NoAnswerMargin))
 	fmt.Fprintf(&b, "  latency: p50 %s  p95 %s  max %s\n", r.P50, r.P95, r.Max)
 	fmt.Fprintf(&b, "  by lang: %s\n", renderStats(r.ByLang))
 	fmt.Fprintf(&b, "  by tag:  %s\n", renderStats(r.ByTag))

internal/memory/recalleval/recalleval_test.go

@@ -142,21 +142,40 @@ func words(s string) []string {
 // cmd/mavend/recall.go (package main is not importable). This pins the copy to
 // the original's three rules: no hits, below the gate, or no text ⇒ silence.
 func TestBestRecallMatchesDaemon(t *testing.T) {
-	if got := bestRecall(nil, 0.55); got != "" {
+	if got := bestRecall(nil, 0.55, 0); got != "" {
 		t.Errorf("no hits: got %q, want silence", got)
 	}
 	low := []memory.Result{{ID: "a", Score: 0.4, Meta: map[string]string{"text": "чай"}}}
-	if got := bestRecall(low, 0.55); got != "" {
+	if got := bestRecall(low, 0.55, 0); got != "" {
 		t.Errorf("below gate: got %q, want silence", got)
 	}
 	noText := []memory.Result{{ID: "a", Score: 0.9, Meta: map[string]string{}}}
-	if got := bestRecall(noText, 0.55); got != "" {
+	if got := bestRecall(noText, 0.55, 0); got != "" {
 		t.Errorf("no text: got %q, want silence", got)
 	}
 	ok := []memory.Result{{ID: "a", Score: 0.9, Meta: map[string]string{"text": "чай"}}}
-	if got := bestRecall(ok, 0.55); got != "чай" {
+	if got := bestRecall(ok, 0.55, 0); got != "чай" {
 		t.Errorf("above gate: got %q, want %q", got, "чай")
 	}
+	// Margin: a close runner-up means the embedder cannot tell the two apart,
+	// so Maven stays silent even though both clear the absolute floor.
+	close := []memory.Result{
+		{ID: "a", Score: 0.86, Meta: map[string]string{"text": "чай"}},
+		{ID: "b", Score: 0.85, Meta: map[string]string{"text": "кофе"}},
+	}
+	if got := bestRecall(close, 0.55, 0.03); got != "" {
+		t.Errorf("thin margin: got %q, want silence", got)
+	}
+	if got := bestRecall(close, 0.55, 0); got != "чай" {
+		t.Errorf("margin off: got %q, want %q", got, "чай")
+	}
+	clear := []memory.Result{
+		{ID: "a", Score: 0.86, Meta: map[string]string{"text": "чай"}},
+		{ID: "b", Score: 0.70, Meta: map[string]string{"text": "кофе"}},
+	}
+	if got := bestRecall(clear, 0.55, 0.03); got != "чай" {
+		t.Errorf("wide margin: got %q, want %q", got, "чай")
+	}
 }
 
 // TestHashRecallBaseline — the CI ratchet. HashEmbedder, so it needs no model
@@ -171,7 +190,7 @@ func TestHashRecallBaseline(t *testing.T) {
 		t.Fatalf("Load: %v", err)
 	}
 	rep, err := Score(context.Background(), "recall+hash", router.NewHashEmbedder(hashDim), InMemory,
-		config.DefaultQueryMinScore, f)
+		config.DefaultQueryMinScore, config.DefaultQueryMinMargin, f)
 	if err != nil {
 		t.Fatalf("Score: %v", err)
 	}
@@ -201,11 +220,11 @@ func TestPersistentStoreScoresTheSame(t *testing.T) {
 		t.Fatalf("Load: %v", err)
 	}
 	emb := router.NewHashEmbedder(hashDim)
-	inMem, err := Score(context.Background(), "recall+hash+memory", emb, InMemory, config.DefaultQueryMinScore, f)
+	inMem, err := Score(context.Background(), "recall+hash+memory", emb, InMemory, config.DefaultQueryMinScore, config.DefaultQueryMinMargin, f)
 	if err != nil {
 		t.Fatalf("Score in-memory: %v", err)
 	}
-	persistent, err := Score(context.Background(), "recall+hash+sqlite", emb, sqliteStores(t), config.DefaultQueryMinScore, f)
+	persistent, err := Score(context.Background(), "recall+hash+sqlite", emb, sqliteStores(t), config.DefaultQueryMinScore, config.DefaultQueryMinMargin, f)
 	if err != nil {
 		t.Fatalf("Score sqlite: %v", err)
 	}
@@ -246,8 +265,8 @@ func TestONNXRecall(t *testing.T) {
 	if lib == "" {
 		t.Skip("MAVEN_ONNX_LIB unset — see AGENTS.md § Embedder model for intent routing")
 	}
-	model := filepath.Join("../../..", "models/embedder/model.onnx")
+	model := filepath.Join("../../..", "models/embedder/multilingual-e5-small/model_quantized.onnx")
-	tok := filepath.Join("../../..", "models/embedder/tokenizer.json")
+	tok := filepath.Join("../../..", "models/embedder/multilingual-e5-small/tokenizer.json")
 	for _, p := range []string{lib, model, tok} {
 		if _, err := os.Stat(p); err != nil {
 			t.Skipf("missing %s: %v", p, err)
@@ -263,14 +282,16 @@ func TestONNXRecall(t *testing.T) {
 	if err != nil {
 		t.Fatalf("Load: %v", err)
 	}
-	rep, err := Score(context.Background(), "recall+onnx", emb, InMemory, config.DefaultQueryMinScore, f)
+	rep, err := Score(context.Background(), "recall+onnx", emb, InMemory, config.DefaultQueryMinScore, config.DefaultQueryMinMargin, f)
 	if err != nil {
 		t.Fatalf("Score: %v", err)
 	}
 	t.Log("\n" + rep.String() + rep.Failures())
-	// Cached for the sweep only: the headline run above must pay the real
+	// Cached for the sweeps only: the headline run above must pay the real
 	// embedder cost so its latency numbers mean something.
-	t.Log("\ngate sweep:\n" + sweep(t, Cache(emb), f))
+	cached := Cache(emb)
+	t.Log("\ngate sweep (margin off):\n" + sweep(t, cached, f))
+	t.Log("\nmargin sweep (gate 0.55):\n" + marginSweep(t, cached, f))
 }
 
 // sweep scores the fixture at a range of gates and renders one line each. Two
@@ -281,7 +302,7 @@ func sweep(t *testing.T, emb router.Embedder, f Fixture) string {
 	t.Helper()
 	var b strings.Builder
 	for _, gate := range []float64{0.0, 0.30, 0.40, 0.50, 0.55, 0.60, 0.70, 0.80, 0.90} {
-		rep, err := Score(context.Background(), "sweep", emb, InMemory, gate, f)
+		rep, err := Score(context.Background(), "sweep", emb, InMemory, gate, 0, f)
 		if err != nil {
 			t.Fatalf("sweep at %.2f: %v", gate, err)
 		}
@@ -290,3 +311,21 @@ func sweep(t *testing.T, emb router.Embedder, f Fixture) string {
 	}
 	return b.String()
 }
+
+// marginSweep is the same idea for the margin gate (top1 − top2 > delta), with
+// the absolute gate held at its default. The absolute score cannot separate a
+// real hit from a made-up question under e5 — every score lands in one narrow
+// band — so this sweep is the one that picks a number.
+func marginSweep(t *testing.T, emb router.Embedder, f Fixture) string {
+	t.Helper()
+	var b strings.Builder
+	for _, d := range []float64{0, 0.002, 0.005, 0.008, 0.01, 0.012, 0.015, 0.02, 0.025, 0.03, 0.04, 0.05, 0.06} {
+		rep, err := Score(context.Background(), "margin sweep", emb, InMemory, config.DefaultQueryMinScore, d, f)
+		if err != nil {
+			t.Fatalf("margin sweep at %.3f: %v", d, err)
+		}
+		fmt.Fprintf(&b, "  delta %.3f: answered %d/%d (%.0f%%)  false recall %d/%d\n",
+			d, rep.Rank1-rep.Gated, rep.Answerable, 100*rep.Answered(), rep.FalseRecall, rep.NoAnswer)
+	}
+	return b.String()
+}

internal/phraser/eval/checks.go

@@ -0,0 +1,287 @@
+package eval
+
+import (
+	"fmt"
+	"regexp"
+	"strings"
+	"unicode"
+	"unicode/utf8"
+)
+
+// The check names, in report order. Every check is a string or length test — no
+// model grades another model here.
+const (
+	CheckMood     = "mood"     // mood is in the documented enum
+	CheckLang     = "lang"     // the operator's language, not the prompt's
+	CheckLength   = "length"   // a nudge is one sentence, not a paragraph
+	CheckFeminine = "feminine" // her self-reference is feminine (hard constraint)
+	CheckCringe   = "cringe"   // DESIGN.md § Non-goals, "not a relationship"
+	CheckOnTopic  = "ontopic"  // says the thing the rule is about
+)
+
+// CheckNames — report order.
+var CheckNames = []string{CheckMood, CheckLang, CheckLength, CheckFeminine, CheckCringe, CheckOnTopic}
+
+// Result — one check on one message.
+type Result struct {
+	Name   string
+	Pass   bool
+	Detail string
+}
+
+// Moods — the fixed enum from the LLM output contract. Not extended here; the
+// contract lives in the daemon and the harness only reads it.
+var Moods = map[string]bool{
+	"neutral": true, "happy": true, "thinking": true, "tired": true, "confused": true,
+}
+
+// Length ceilings. Justification: the nudge is spoken by piper at roughly 14
+// characters per second, so 120 characters is about 8 seconds of speech. The
+// operator has AuDHD — past one short sentence a nudge stops being a nudge and
+// becomes something to tune out, which is exactly the "not a nag" failure. The
+// word ceiling catches the same thing for languages that pack more per byte.
+const (
+	MaxChars = 120
+	MaxWords = 16
+)
+
+// RunChecks scores one message. Order matches CheckNames.
+func RunChecks(c Case, body, mood string) []Result {
+	return []Result{
+		checkMood(mood),
+		checkLang(body),
+		checkLength(body),
+		checkFeminine(body),
+		checkCringe(body),
+		checkOnTopic(c, body),
+	}
+}
+
+func checkMood(mood string) Result {
+	if Moods[mood] {
+		return Result{CheckMood, true, ""}
+	}
+	return Result{CheckMood, false, fmt.Sprintf("mood %q not in the enum", mood)}
+}
+
+// checkLang — the operator is Russian-speaking and the nudge is spoken aloud by
+// a Russian piper voice. An English nudge is not a tone problem, it is an
+// unusable one.
+func checkLang(body string) Result {
+	cyr, lat := 0, 0
+	for _, r := range body {
+		switch {
+		case unicode.Is(unicode.Cyrillic, r):
+			cyr++
+		case r >= 'a' && r <= 'z', r >= 'A' && r <= 'Z':
+			lat++
+		}
+	}
+	if cyr > lat {
+		return Result{CheckLang, true, ""}
+	}
+	return Result{CheckLang, false, fmt.Sprintf("not Russian (%d cyrillic vs %d latin letters)", cyr, lat)}
+}
+
+func checkLength(body string) Result {
+	chars := utf8.RuneCountInString(body)
+	words := len(strings.Fields(body))
+	if chars <= MaxChars && words <= MaxWords {
+		return Result{CheckLength, true, ""}
+	}
+	return Result{CheckLength, false,
+		fmt.Sprintf("%d chars / %d words, ceiling %d / %d", chars, words, MaxChars, MaxWords)}
+}
+
+// --- feminine self-reference ---------------------------------------------
+//
+// The hard constraint (CLAUDE.md, DESIGN.md § Identity): Maven's Russian
+// self-reference is feminine. The operator is male, so second-person forms
+// addressed to him are MASCULINE and must not be flagged — "ты не пил воду" is
+// correct, "я напомнил" is not. Both directions matter, which is why this is a
+// windowed scan around "я" and not a bare search for masculine endings.
+
+var wordRE = regexp.MustCompile(`[\p{Cyrillic}]+|[,.;:!?…—-]`)
+
+// secondPerson — pronouns that end the self-reference window. Everything after
+// one of these is about him, not about her.
+var secondPerson = map[string]bool{
+	"ты": true, "тебе": true, "тебя": true, "тобой": true,
+	"вы": true, "вам": true, "вас": true,
+	"он": true, "она": true, "оно": true, "они": true,
+}
+
+// masculinePredicative — short adjectives with no verb ending to key off.
+var masculinePredicative = map[string]bool{
+	"должен": true, "готов": true, "рад": true, "уверен": true,
+	"обязан": true, "сам": true, "занят": true, "прав": true,
+}
+
+// nounsEndingInL — the false positives of "ends in л ⇒ masculine past tense".
+// Small on purpose: it only has to cover nouns a nudge might actually use.
+var nounsEndingInL = map[string]bool{
+	"стол": true, "стул": true, "пол": true, "зал": true, "гол": true,
+	"узел": true, "отдел": true, "файл": true, "канал": true, "угол": true,
+	"футбол": true, "вокзал": true, "металл": true, "интервал": true,
+	"уровень": true, "мускул": true, "апрель": true, "июль": true, "рубль": true,
+}
+
+// masculinePast reports whether a word looks like a masculine past-tense verb.
+// Russian past tense is gendered by suffix: -л (m), -ла (f). A 0.8B with weak
+// Russian defaults to the masculine form, which is the exact drift being
+// measured.
+func masculinePast(w string) bool {
+	if len([]rune(w)) < 3 || nounsEndingInL[w] {
+		return false
+	}
+	return strings.HasSuffix(w, "л") || strings.HasSuffix(w, "лся")
+}
+
+func checkFeminine(body string) Result {
+	words := wordRE.FindAllString(strings.ToLower(body), -1)
+	for i, w := range words {
+		if w != "я" {
+			continue
+		}
+		// Scan the next few words. Stop at punctuation or at a second-person
+		// pronoun: past that point the sentence is about him and masculine is
+		// correct.
+		for j := i + 1; j < len(words) && j <= i+3; j++ {
+			nw := words[j]
+			if len(nw) == 1 && !unicode.Is(unicode.Cyrillic, []rune(nw)[0]) {
+				break
+			}
+			if secondPerson[nw] {
+				break
+			}
+			if masculinePast(nw) || masculinePredicative[nw] {
+				return Result{CheckFeminine, false,
+					fmt.Sprintf("masculine self-reference %q after \"я\"", nw)}
+			}
+		}
+	}
+	// Second pass: self-reference with the pronoun dropped — "напомнил тебе",
+	// "проверил за тебя". A masculine past-tense verb whose object is HIM can
+	// only be her speaking about herself.
+	for i, w := range words {
+		if !masculinePast(w) || i+1 >= len(words) {
+			continue
+		}
+		next := words[i+1]
+		if next == "тебе" || next == "тебя" || next == "за" {
+			return Result{CheckFeminine, false,
+				fmt.Sprintf("masculine self-reference %q before %q", w, next)}
+		}
+	}
+	return Result{CheckFeminine, true, ""}
+}
+
+// --- the cringe checks ---------------------------------------------------
+//
+// "Think Jarvis without the cringe part". DESIGN.md § Non-goals: "Not a
+// relationship — mom-tone is a function that makes nudges land, not emotional
+// company. Names the drift a warm small model falls into." Each pattern below
+// is one shape of that drift. They are deliberately specific: a check that
+// flags any warmth at all would make the nudges robotic, which is the other
+// failure.
+
+type cringePattern struct {
+	// what the pattern is defending against, shown in the failure detail.
+	why string
+	pat *regexp.Regexp
+}
+
+var cringePatterns = []cringePattern{
+	{
+		// Endearments. "Not a relationship" — a pet name reframes a nudge as
+		// intimacy, and the operator asked for mother-like, not girlfriend-like.
+		why: "pet name / endearment",
+		// No \b around the Russian alternatives: Go's RE2 \b is ASCII-only and
+		// never matches at a Cyrillic boundary, so anchoring them would make
+		// this check silently always pass.
+		pat: regexp.MustCompile(`(?i)(милый|дорогой|солнышко|солнце моё|солнце мое|зайчик|котик|сладкий|малыш|дружок|родной|любимый|\bhoney\b|\bsweetie\b|\bdarling\b|\bbuddy\b)`),
+	},
+	{
+		// Emoji. The nudge is spoken aloud; an emoji is either silence or a TTS
+		// artefact. Also the single loudest cringe signal in a small model.
+		why: "emoji",
+		pat: nil, // handled by hasEmoji, ranges don't fit a regexp cleanly
+	},
+	{
+		// Exclamation pileup. One "!" is emphasis; two is a cheerleader.
+		why: "more than one exclamation mark",
+		pat: regexp.MustCompile(`!.*!|!!`),
+	},
+	{
+		// Fake concern. She has no feelings to report, and reporting them makes
+		// the nudge about her instead of about the water.
+		why: "fake concern opener",
+		pat: regexp.MustCompile(`(?i)(я волну|я беспоко|беспокоюсь|переживаю|я забочусь|я тревож|мне тревожно|i'?m worried)`),
+	},
+	{
+		// Apologising. The rule decided she speaks. Apologising for a greenlit
+		// nudge undermines the one thing that makes nudges land.
+		why: "apology",
+		pat: regexp.MustCompile(`(?i)(извини|прости|сожалею|прошу прощения|не хочу мешать|не хочу отвлекать|sorry|apolog)`),
+	},
+	{
+		// Offering emotional support. The explicit "not emotional company" line.
+		why: "offer of emotional support",
+		pat: regexp.MustCompile(`(?i)(я рядом|я здесь для теб|ты не один|всё будет хорошо|все будет хорошо|не переживай|я с тобой|обнимаю|я поддерж|держись)`),
+	},
+	{
+		// Asking how he feels. Turns a one-way nudge into a conversation he now
+		// owes an answer to — the most reliable way to make him mute it.
+		why: "asking how he feels",
+		pat: regexp.MustCompile(`(?i)(как ты\s*[?.!]|как ты себя|как самочувств|как настроение|всё ли в порядке|все ли в порядке|ты в порядке|how are you)`),
+	},
+	{
+		// Praise for compliance. Rewards make the nudge a training exercise;
+		// "not a relationship" again, from the other side.
+		why: "praise / reward framing",
+		pat: regexp.MustCompile(`(?i)(молодец|умница|ты справ|гордюсь|горжусь|отличная работа|так держать|good job|proud of you)`),
+	},
+}
+
+// hasEmoji — the pictographic ranges plus the variation selector. Cyrillic and
+// ordinary punctuation are far below all of these.
+func hasEmoji(s string) bool {
+	for _, r := range s {
+		switch {
+		case r >= 0x1F000 && r <= 0x1FAFF,
+			r >= 0x2600 && r <= 0x27BF,
+			r >= 0x2B00 && r <= 0x2BFF,
+			r == 0xFE0F, r == 0x203C, r == 0x2049:
+			return true
+		}
+	}
+	return false
+}
+
+func checkCringe(body string) Result {
+	for _, c := range cringePatterns {
+		if c.pat == nil {
+			if hasEmoji(body) {
+				return Result{CheckCringe, false, c.why}
+			}
+			continue
+		}
+		if m := c.pat.FindString(body); m != "" {
+			return Result{CheckCringe, false, fmt.Sprintf("%s (%q)", c.why, m)}
+		}
+	}
+	return Result{CheckCringe, true, ""}
+}
+
+// checkOnTopic — the message must name the thing the rule is about. A nudge
+// that never mentions water leaves the operator with a chime and no action.
+func checkOnTopic(c Case, body string) Result {
+	low := strings.ToLower(body)
+	for _, want := range c.WantAny {
+		if strings.Contains(low, strings.ToLower(want)) {
+			return Result{CheckOnTopic, true, ""}
+		}
+	}
+	return Result{CheckOnTopic, false,
+		fmt.Sprintf("mentions none of %v", c.WantAny)}
+}

internal/phraser/eval/eval.go

@@ -0,0 +1,334 @@
+// Package eval scores nudge phrasing — the sentences the operator actually
+// hears. It is the phrasing counterpart to internal/router/eval.
+//
+// Why a separate package from phraser: the fixture must be scorable by BOTH
+// phrasing paths (the deterministic Stub and the resident model) from outside
+// the phraser package, and a _test.go file inside phraser cannot be imported.
+// So the fixture is embedded here and the scorer takes a Nudger interface.
+//
+// Why deterministic checks and not model judgement: the resident model is a
+// 0.8B. It cannot grade its own tone. Every check in checks.go is a string or
+// length test that a human can read and disagree with. A score here is a claim
+// about measurable properties, not about whether a sentence is good.
+//
+// DESIGN.md § "Rules decide, LLM phrases" is why there is no send/veto signal
+// anywhere in this package: the rule already decided she speaks. The phraser
+// only words it, so a nudge the model refuses to write is a failure, never a
+// legitimate outcome.
+package eval
+
+import (
+	"context"
+	_ "embed"
+	"encoding/json"
+	"fmt"
+	"sort"
+	"strings"
+	"time"
+
+	"github.com/kami/maven/internal/delivery"
+	"github.com/kami/maven/internal/loop"
+	"github.com/kami/maven/internal/store"
+)
+
+//go:embed nudges_v1.json
+var fixtureJSON []byte
+
+// SchemaVersion — the version this package understands.
+const SchemaVersion = 1
+
+// Case — one nudge situation, as a real tick would present it. The fields are
+// the (rule, severity, context) input DESIGN.md names, flattened to JSON.
+//
+// WantAny is the on-topic contract: at least one of these lowercased fragments
+// must appear in the message. A water nudge that never mentions water is a
+// failure however charming it reads. Fragments are stems ("вод") so declension
+// does not defeat the check, and they list both languages because the Stub is
+// still English (see the writeup).
+type Case struct {
+	ID       string `json:"id"`
+	Rule     string `json:"rule"`
+	Severity int    `json:"severity"`
+
+	// SinceMinutes — age of the rule's own fact. 0 means "no such fact", which
+	// is the branch where the phraser has no duration to name.
+	SinceMinutes int `json:"since_minutes"`
+
+	// FactKey/FactValue/FactSource — the aggregate fact behind the ops rules.
+	// service_down phrasing reads the key for the service name.
+	FactKey    string `json:"fact_key,omitempty"`
+	FactValue  string `json:"fact_value,omitempty"`
+	FactSource string `json:"fact_source,omitempty"`
+
+	// QuietHours/CalendarBusy — the bad moments. The gate already let this
+	// nudge through (ops outranks quiet hours), so the phrasing still has to be
+	// short and plain rather than apologetic about the timing.
+	QuietHours   bool `json:"quiet_hours,omitempty"`
+	CalendarBusy bool `json:"calendar_busy,omitempty"`
+
+	WantAny []string `json:"want_any"`
+	Tags    []string `json:"tags,omitempty"`
+	Note    string   `json:"note,omitempty"`
+}
+
+// Fixture — the versioned envelope. SchemaVersion gates the loader so an older
+// binary refuses a fixture it would misread instead of reporting a wrong score.
+type Fixture struct {
+	SchemaVersion int      `json:"schema_version"`
+	Name          string   `json:"name"`
+	ReferenceNow  string   `json:"reference_now"`
+	Notes         []string `json:"notes"`
+	Cases         []Case   `json:"cases"`
+}
+
+// Load returns the embedded fixture.
+func Load() (Fixture, error) {
+	var f Fixture
+	if err := json.Unmarshal(fixtureJSON, &f); err != nil {
+		return Fixture{}, fmt.Errorf("parse fixture: %w", err)
+	}
+	if f.SchemaVersion != SchemaVersion {
+		return Fixture{}, fmt.Errorf("fixture schema_version %d, want %d", f.SchemaVersion, SchemaVersion)
+	}
+	if len(f.Cases) == 0 {
+		return Fixture{}, fmt.Errorf("fixture has no cases")
+	}
+	return f, nil
+}
+
+// Now — the fixture's reference clock, so fact ages are reproducible.
+func (f Fixture) Now() (time.Time, error) {
+	t, err := time.Parse(time.RFC3339, f.ReferenceNow)
+	if err != nil {
+		return time.Time{}, fmt.Errorf("parse reference_now %q: %w", f.ReferenceNow, err)
+	}
+	return t, nil
+}
+
+// Candidate rebuilds the loop.Candidate a tick would hand the phraser.
+func (c Case) Candidate(now time.Time) loop.Candidate {
+	state := loop.State{
+		Now:          now,
+		Facts:        map[string]store.Fact{},
+		QuietHours:   c.QuietHours,
+		CalendarBusy: c.CalendarBusy,
+	}
+	if c.SinceMinutes > 0 {
+		state.Facts[c.Rule] = store.Fact{
+			Key: c.Rule,
+			Ts:  now.Add(-time.Duration(c.SinceMinutes) * time.Minute),
+		}
+	}
+	if c.FactKey != "" {
+		state.Facts[c.Rule] = store.Fact{
+			Key:    c.FactKey,
+			Value:  c.FactValue,
+			Source: c.FactSource,
+			Ts:     now.Add(-time.Duration(c.SinceMinutes) * time.Minute),
+		}
+	}
+	sev := loop.Severity(c.Severity)
+	return loop.Candidate{
+		Rule:     loop.Rule{Name: c.Rule, Severity: sev},
+		Severity: sev,
+		State:    state,
+	}
+}
+
+// Nudger — the one thing a phrasing path must do to be scorable. Both
+// *phraser.Stub and *phraser.LLMPhraser satisfy it.
+type Nudger interface {
+	PhraseNudge(ctx context.Context, c loop.Candidate) (delivery.PhrasedNudge, error)
+}
+
+// Outcome — one scored case. Failed lists the check names that did not pass,
+// Reasons the human-readable detail. Failed is empty exactly when Pass is true.
+type Outcome struct {
+	Case    Case
+	Body    string
+	Mood    string
+	Err     error
+	Latency time.Duration
+	Pass    bool
+	Failed  []string
+	Reasons []string
+}
+
+// Report — the aggregate. ByCheck is the useful part: one composite percentage
+// hides which property broke, and tuning a prompt needs to know.
+type Report struct {
+	Name     string
+	Total    int
+	Passed   int
+	Errors   int
+	ByCheck  map[string]int
+	ByRule   map[string]TagStat
+	Outcomes []Outcome
+	P50      time.Duration
+	P95      time.Duration
+	Max      time.Duration
+}
+
+// TagStat — passed/total for one slice of the fixture.
+type TagStat struct{ Passed, Total int }
+
+// Accuracy — fraction of cases that passed every check.
+func (r Report) Accuracy() float64 {
+	if r.Total == 0 {
+		return 0
+	}
+	return float64(r.Passed) / float64(r.Total)
+}
+
+// Score runs every case through p and aggregates. A phrasing error scores as a
+// miss and is counted in Errors — "the model was down" and "the model wrote
+// something bad" are different numbers and a prompt change must not be able to
+// hide behind the first one.
+//
+// Latency is wall-clock per PhraseNudge call. On the CPU/iGPU target a nudge
+// the model takes a minute to word has already missed its moment.
+func Score(ctx context.Context, name string, p Nudger, f Fixture) (Report, error) {
+	now, err := f.Now()
+	if err != nil {
+		return Report{}, err
+	}
+	rep := Report{
+		Name:    name,
+		Total:   len(f.Cases),
+		ByCheck: map[string]int{},
+		ByRule:  map[string]TagStat{},
+	}
+	for _, name := range CheckNames {
+		rep.ByCheck[name] = 0
+	}
+	lat := make([]time.Duration, 0, len(f.Cases))
+
+	for _, c := range f.Cases {
+		start := time.Now()
+		pn, err := p.PhraseNudge(ctx, c.Candidate(now))
+		o := Outcome{Case: c, Body: pn.Body, Mood: pn.Mood, Err: err, Latency: time.Since(start)}
+		lat = append(lat, o.Latency)
+
+		if err != nil {
+			rep.Errors++
+			o.Failed = append(o.Failed, "call")
+			o.Reasons = append(o.Reasons, fmt.Sprintf("phrase error: %v", err))
+		} else {
+			for _, res := range RunChecks(c, pn.Body, pn.Mood) {
+				if res.Pass {
+					rep.ByCheck[res.Name]++
+					continue
+				}
+				o.Failed = append(o.Failed, res.Name)
+				o.Reasons = append(o.Reasons, res.Name+": "+res.Detail)
+			}
+		}
+
+		o.Pass = len(o.Failed) == 0
+		if o.Pass {
+			rep.Passed++
+		}
+		bump(rep.ByRule, ruleFamily(c.Rule), o.Pass)
+		rep.Outcomes = append(rep.Outcomes, o)
+	}
+
+	sort.Slice(lat, func(i, j int) bool { return lat[i] < lat[j] })
+	rep.P50, rep.P95 = percentile(lat, 0.50), percentile(lat, 0.95)
+	if len(lat) > 0 {
+		rep.Max = lat[len(lat)-1]
+	}
+	return rep, nil
+}
+
+// ruleFamily collapses "routine:зарядка" to "routine" so the per-rule table
+// stays readable however many routines the operator configures.
+func ruleFamily(rule string) string {
+	if i := strings.IndexByte(rule, ':'); i > 0 {
+		return rule[:i]
+	}
+	return rule
+}
+
+func bump(m map[string]TagStat, key string, pass bool) {
+	if key == "" {
+		return
+	}
+	s := m[key]
+	s.Total++
+	if pass {
+		s.Passed++
+	}
+	m[key] = s
+}
+
+// percentile — nearest-rank on a pre-sorted slice. No interpolation: with ~15
+// samples an interpolated p95 invents a latency no call actually took.
+func percentile(sorted []time.Duration, p float64) time.Duration {
+	if len(sorted) == 0 {
+		return 0
+	}
+	i := int(p * float64(len(sorted)))
+	if i >= len(sorted) {
+		i = len(sorted) - 1
+	}
+	return sorted[i]
+}
+
+// String renders the comparison table — composite score, then per-check so a
+// regression names the property it broke, then latency.
+func (r Report) String() string {
+	var b strings.Builder
+	fmt.Fprintf(&b, "%s: %d/%d cases pass every check (%.1f%%), %d errors\n",
+		r.Name, r.Passed, r.Total, 100*r.Accuracy(), r.Errors)
+	for _, name := range CheckNames {
+		fmt.Fprintf(&b, "  %-9s %d/%d\n", name, r.ByCheck[name], r.Total)
+	}
+	fmt.Fprintf(&b, "  latency: p50 %s  p95 %s  max %s\n", r.P50, r.P95, r.Max)
+	fmt.Fprintf(&b, "  by rule: %s\n", renderStats(r.ByRule))
+	return b.String()
+}
+
+// Failures — per-case detail, sorted by ID so two runs diff cleanly.
+func (r Report) Failures() string {
+	var b strings.Builder
+	for _, o := range r.sorted() {
+		if o.Pass {
+			continue
+		}
+		fmt.Fprintf(&b, "  %s %q\n    %s\n", o.Case.ID, o.Body, strings.Join(o.Reasons, "; "))
+	}
+	return b.String()
+}
+
+// Messages — every generated message verbatim, pass or fail. This is what a
+// human reads to judge tone; the score only says which checks fired.
+func (r Report) Messages() string {
+	var b strings.Builder
+	for _, o := range r.sorted() {
+		mark := "ok  "
+		if !o.Pass {
+			mark = "FAIL"
+		}
+		fmt.Fprintf(&b, "  %s %-22s [%s] %q\n", mark, o.Case.ID, o.Mood, o.Body)
+	}
+	return b.String()
+}
+
+func (r Report) sorted() []Outcome {
+	out := append([]Outcome(nil), r.Outcomes...)
+	sort.Slice(out, func(i, j int) bool { return out[i].Case.ID < out[j].Case.ID })
+	return out
+}
+
+func renderStats(m map[string]TagStat) string {
+	keys := make([]string, 0, len(m))
+	for k := range m {
+		keys = append(keys, k)
+	}
+	sort.Strings(keys)
+	parts := make([]string, 0, len(keys))
+	for _, k := range keys {
+		parts = append(parts, fmt.Sprintf("%s %d/%d", k, m[k].Passed, m[k].Total))
+	}
+	return strings.Join(parts, "  ")
+}

internal/phraser/eval/eval_test.go

@@ -0,0 +1,179 @@
+package eval
+
+import (
+	"context"
+	"strings"
+	"testing"
+
+	"github.com/kami/maven/internal/loop"
+	"github.com/kami/maven/internal/phraser"
+)
+
+func TestLoadFixture(t *testing.T) {
+	f, err := Load()
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	if _, err := f.Now(); err != nil {
+		t.Fatalf("Now: %v", err)
+	}
+	seen := map[string]bool{}
+	for _, c := range f.Cases {
+		if seen[c.ID] {
+			t.Errorf("duplicate case id %q", c.ID)
+		}
+		seen[c.ID] = true
+		if c.Rule == "" || c.Severity < 1 || c.Severity > 4 {
+			t.Errorf("%s: rule %q severity %d", c.ID, c.Rule, c.Severity)
+		}
+		if len(c.WantAny) == 0 {
+			t.Errorf("%s: no want_any, the on-topic check would always pass", c.ID)
+		}
+	}
+	// Coverage floor: all five loop rules plus both minted families, or the
+	// fixture measures a subset and the score does not mean what it says.
+	for _, rule := range []string{"water", "meal", "break", "service_down", "netdata_critical", "routine", "morning"} {
+		found := false
+		for _, c := range f.Cases {
+			if ruleFamily(c.Rule) == rule {
+				found = true
+			}
+		}
+		if !found {
+			t.Errorf("no case for rule family %q", rule)
+		}
+	}
+}
+
+// TestStubBaseline is the CI ratchet: the deterministic Stub, no model, no
+// network. The floor is low on purpose — the Stub is English template phrasing,
+// so it fails `lang` on every case by construction. The point of the ratchet is
+// that the checks keep running and the Stub does not get worse, not that the
+// Stub is good.
+func TestStubBaseline(t *testing.T) {
+	f, err := Load()
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	rep, err := Score(context.Background(), "stub (deterministic floor)", phraser.NewStub(), f)
+	if err != nil {
+		t.Fatalf("Score: %v", err)
+	}
+	t.Log("\n" + rep.String() + rep.Messages())
+
+	if rep.Errors != 0 {
+		t.Errorf("stub returned %d errors — the deterministic path must never fail", rep.Errors)
+	}
+	// Per-check ratchets rather than one composite: the Stub's composite is 0
+	// (it never passes `lang`), so a composite floor would catch nothing.
+	floors := map[string]int{
+		CheckMood: 15,
+		// 12, not 15: the Stub's `break` template genuinely runs past the
+		// ceiling ("you've been at your desk for 4 hours without a break — step
+		// away for a bit." is 76 chars but 16 words). Left failing rather than
+		// raising the ceiling to hide it.
+		CheckLength:   12,
+		CheckFeminine: 15,
+		CheckCringe:   15,
+		CheckOnTopic:  12,
+	}
+	for name, floor := range floors {
+		if rep.ByCheck[name] < floor {
+			t.Errorf("check %s: %d/%d, below ratchet %d — phrasing regressed",
+				name, rep.ByCheck[name], rep.Total, floor)
+		}
+	}
+}
+
+// TestChecksCatchWhatTheyClaim — the checks are the measurement, so they get
+// their own tests. Without these, a bad regexp would silently make every
+// phrasing run look clean.
+func TestChecksCatchWhatTheyClaim(t *testing.T) {
+	water := Case{Rule: "water", WantAny: []string{"вод"}}
+
+	cases := []struct {
+		name string
+		body string
+		want string // the check that must fail, "" for a clean message
+	}{
+		{"clean", "уже четыре часа без воды — попей.", ""},
+		{"long", "уже четыре часа без воды, а это довольно много, и вообще пить надо регулярно, иначе будет плохо совсем", CheckLength},
+		{"english", "you haven't had water in 4 hours, drink something", CheckLang},
+		{"masculine self", "я напомнил про воду.", CheckFeminine},
+		{"masculine dropped pronoun", "напомнил тебе про воду.", CheckFeminine},
+		{"masculine predicative", "я должен сказать: попей воды.", CheckFeminine},
+		// The other direction: HE is male, so second-person masculine is right.
+		{"second person masculine ok", "ты не пил воду четыре часа.", ""},
+		{"feminine self ok", "я заметила: воды не было четыре часа.", ""},
+		{"pet name", "милый, попей воды.", CheckCringe},
+		{"emoji", "попей воды 💧", CheckCringe},
+		{"exclamations", "попей воды!!", CheckCringe},
+		{"fake concern", "я беспокоюсь: воды не было четыре часа.", CheckCringe},
+		{"apology", "извини, что отвлекаю — попей воды.", CheckCringe},
+		{"emotional support", "я рядом, ты не один. попей воды.", CheckCringe},
+		{"asks how he feels", "как ты себя чувствуешь? попей воды.", CheckCringe},
+		{"praise", "молодец! теперь попей воды.", CheckCringe},
+		{"off topic", "пора бы уже что-то сделать.", CheckOnTopic},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			var failed []string
+			for _, r := range RunChecks(water, tc.body, "neutral") {
+				if !r.Pass {
+					failed = append(failed, r.Name+"("+r.Detail+")")
+				}
+			}
+			joined := strings.Join(failed, " ")
+			switch {
+			case tc.want == "" && len(failed) > 0:
+				t.Errorf("clean message flagged: %s", joined)
+			case tc.want != "" && !strings.Contains(joined, tc.want+"("):
+				t.Errorf("want %s to fail, got %q", tc.want, joined)
+			}
+		})
+	}
+}
+
+func TestMoodCheckUsesTheEnum(t *testing.T) {
+	if r := checkMood("cheerful"); r.Pass {
+		t.Error("mood outside the enum passed")
+	}
+	if r := checkMood(""); r.Pass {
+		t.Error("empty mood passed")
+	}
+	for m := range Moods {
+		if r := checkMood(m); !r.Pass {
+			t.Errorf("enum mood %q failed", m)
+		}
+	}
+}
+
+// TestCandidateCarriesTheContext — the whole harness is worthless if the
+// Candidate it builds does not carry the duration the prompt is supposed to
+// name.
+func TestCandidateCarriesTheContext(t *testing.T) {
+	f, err := Load()
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	now, _ := f.Now()
+	for _, c := range f.Cases {
+		cand := c.Candidate(now)
+		if cand.Rule.Name != c.Rule || cand.Severity != loop.Severity(c.Severity) {
+			t.Errorf("%s: candidate lost rule or severity", c.ID)
+		}
+		if c.SinceMinutes > 0 {
+			d, ok := cand.State.Since(c.Rule)
+			if !ok || int(d.Minutes()) != c.SinceMinutes {
+				t.Errorf("%s: since %v ok=%v, want %d minutes", c.ID, d, ok, c.SinceMinutes)
+			}
+		}
+		if c.FactKey != "" {
+			fact, ok := cand.State.Fact(c.Rule)
+			if !ok || fact.Key != c.FactKey {
+				t.Errorf("%s: fact key %q, want %q", c.ID, fact.Key, c.FactKey)
+			}
+		}
+	}
+}

internal/phraser/eval/llmphraser_test.go

@@ -0,0 +1,73 @@
+package eval
+
+import (
+	"context"
+	"os"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/kami/maven/internal/phraser"
+)
+
+// TestLLMPhrasingBaseline — the resident model wording real nudges. Opt-in,
+// same shape as internal/router/eval's MAVEN_LLM_URL gate, because CI has no
+// model and a phrasing run costs minutes on the CPU target.
+//
+//	llama-server -m /mnt/hdd1/llms/qwen3.5/Qwen3.5-0.8B.Q4_K_M.gguf \
+//	  --host 127.0.0.1 --port 18099 -c 2048 -ngl 99
+//	MAVEN_LLM_URL=http://127.0.0.1:18099 make eval-phrasing
+//
+// It reports and does not assert a quality bar. The numbers are the input to
+// tuning the persona prompt; an assertion here would be the test inventing the
+// bar rather than measuring against it. The one thing worth failing on is a
+// harness fault — every case erroring means the run measured infrastructure.
+func TestLLMPhrasingBaseline(t *testing.T) {
+	base := os.Getenv("MAVEN_LLM_URL")
+	if base == "" {
+		t.Skip("MAVEN_LLM_URL unset — point it at a running llama-server (see doc comment)")
+	}
+	// A local llama-server must not go through an HTTP proxy. This box proxies
+	// loopback through a SOCKS bridge that answers 503, which would score every
+	// case as a phrasing error and read as "the model cannot phrase".
+	noProxyLoopback(t)
+
+	f, err := Load()
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+
+	cfg := phraser.DefaultConfig("")
+	// Generous: an unconstrained 0.8B can spend a minute thinking before it
+	// writes a word, and a timeout would be scored as a model failure.
+	cfg.Timeout = 5 * time.Minute
+	p := phraser.NewLLMPhraserAt(base, cfg)
+	defer p.Close()
+
+	rep, err := Score(context.Background(), "llm (0.8B, built-in persona)", p, f)
+	if err != nil {
+		t.Fatalf("Score: %v", err)
+	}
+	t.Log("\n" + rep.String() + "\nmessages:\n" + rep.Messages() + "\nfailures:\n" + rep.Failures())
+
+	if rep.Errors == rep.Total {
+		t.Errorf("all %d cases errored — harness fault, not a measurement", rep.Total)
+	}
+}
+
+// noProxyLoopback appends the loopback host to no_proxy before any request, so
+// http.ProxyFromEnvironment (which caches the environment on first use) sees it.
+func noProxyLoopback(t *testing.T) {
+	t.Helper()
+	for _, key := range []string{"no_proxy", "NO_PROXY"} {
+		cur := os.Getenv(key)
+		if strings.Contains(cur, "127.0.0.1") {
+			continue
+		}
+		if cur == "" {
+			t.Setenv(key, "127.0.0.1,localhost")
+			continue
+		}
+		t.Setenv(key, cur+",127.0.0.1,localhost")
+	}
+}

internal/phraser/eval/nudges_v1.json

@@ -0,0 +1,158 @@
+{
+  "schema_version": 1,
+  "name": "nudge phrasing v1",
+  "reference_now": "2026-07-31T21:40:00+03:00",
+  "notes": [
+    "The five loop rules (water, meal, break, service_down, netdata_critical) plus one routine: and one morning: case, at the severities they actually ship with.",
+    "The bad-moment cases (quiet_hours, calendar_busy) are here because the gate already let them through — ops outranks quiet hours. The phrasing must stay short and plain, not apologise for the timing.",
+    "want_any lists stems, not whole words, so declension does not defeat the on-topic check. English stems are included because the deterministic Stub is still English.",
+    "There is no expected message. The checks measure properties, not similarity to a reference sentence — a fixed golden string would just freeze one arbitrary phrasing."
+  ],
+  "cases": [
+    {
+      "id": "water-3h",
+      "rule": "water",
+      "severity": 1,
+      "since_minutes": 190,
+      "want_any": ["вод", "попей", "пить", "выпей", "напит", "water", "drink"],
+      "tags": ["care"],
+      "note": "The base case. Just over the 3h predicate."
+    },
+    {
+      "id": "water-7h",
+      "rule": "water",
+      "severity": 1,
+      "since_minutes": 430,
+      "want_any": ["вод", "попей", "пить", "выпей", "напит", "water", "drink"],
+      "tags": ["care"],
+      "note": "Long overdue. Severity is unchanged, so the phrasing must not escalate into alarm."
+    },
+    {
+      "id": "water-busy",
+      "rule": "water",
+      "severity": 1,
+      "since_minutes": 240,
+      "calendar_busy": true,
+      "want_any": ["вод", "попей", "пить", "выпей", "напит", "water", "drink"],
+      "tags": ["care", "bad-moment"],
+      "note": "Mid-meeting. A bad moment invites an apology, which is the check that should catch it."
+    },
+    {
+      "id": "meal-7h",
+      "rule": "meal",
+      "severity": 1,
+      "since_minutes": 420,
+      "want_any": ["ешь", "еда", "еды", "поешь", "перекус", "обед", "ужин", "покуш", "food", "eat"],
+      "tags": ["care"]
+    },
+    {
+      "id": "meal-11h-quiet",
+      "rule": "meal",
+      "severity": 1,
+      "since_minutes": 660,
+      "quiet_hours": true,
+      "want_any": ["ешь", "еда", "еды", "поешь", "перекус", "обед", "ужин", "покуш", "food", "eat"],
+      "tags": ["care", "bad-moment"],
+      "note": "Quiet hours suppresses care nudges in the gate, so this one only reaches the phraser via an explicit override. Included because it is the shape most likely to draw a hedge."
+    },
+    {
+      "id": "break-90m",
+      "rule": "break",
+      "severity": 2,
+      "since_minutes": 95,
+      "want_any": ["перерыв", "разомн", "встань", "отдохн", "пауз", "отойд", "размин", "break", "step away"],
+      "tags": ["care"]
+    },
+    {
+      "id": "break-4h",
+      "rule": "break",
+      "severity": 2,
+      "since_minutes": 240,
+      "want_any": ["перерыв", "разомн", "встань", "отдохн", "пауз", "отойд", "размин", "break", "step away"],
+      "tags": ["care"]
+    },
+    {
+      "id": "break-busy",
+      "rule": "break",
+      "severity": 2,
+      "since_minutes": 150,
+      "calendar_busy": true,
+      "want_any": ["перерыв", "разомн", "встань", "отдохн", "пауз", "отойд", "размин", "break", "step away"],
+      "tags": ["care", "bad-moment"]
+    },
+    {
+      "id": "service-down",
+      "rule": "service_down",
+      "severity": 4,
+      "since_minutes": 3,
+      "fact_key": "vaultwarden",
+      "fact_value": "\"down\"",
+      "fact_source": "poll:uptimekuma",
+      "want_any": ["vaultwarden", "сервис", "упал", "не отвеч", "лежит", "недоступ", "down", "service"],
+      "tags": ["ops"],
+      "note": "The service name is in the fact key, not the value. A nudge that says 'a service' without naming it is on-topic but useless — the on-topic check cannot catch that, a human reading Messages() can."
+    },
+    {
+      "id": "service-down-night",
+      "rule": "service_down",
+      "severity": 4,
+      "since_minutes": 2,
+      "quiet_hours": true,
+      "fact_key": "nextcloud",
+      "fact_value": "\"down\"",
+      "fact_source": "poll:uptimekuma",
+      "want_any": ["nextcloud", "сервис", "упал", "не отвеч", "лежит", "недоступ", "down", "service"],
+      "tags": ["ops", "bad-moment"],
+      "note": "03:00-shaped. Sev4 outranks quiet hours by design, so she speaks — plainly, without softening it into a maybe."
+    },
+    {
+      "id": "netdata-disk",
+      "rule": "netdata_critical",
+      "severity": 3,
+      "since_minutes": 5,
+      "fact_key": "netdata_alarm",
+      "fact_value": "\"critical\"",
+      "fact_source": "poll:netdata",
+      "want_any": ["netdata", "диск", "критич", "алярм", "аларм", "тревог", "место", "памят", "critical", "alarm", "disk"],
+      "tags": ["ops"]
+    },
+    {
+      "id": "netdata-busy",
+      "rule": "netdata_critical",
+      "severity": 3,
+      "since_minutes": 12,
+      "calendar_busy": true,
+      "fact_key": "netdata_alarm",
+      "fact_value": "\"critical\"",
+      "fact_source": "poll:netdata",
+      "want_any": ["netdata", "диск", "критич", "алярм", "аларм", "тревог", "место", "памят", "critical", "alarm", "disk"],
+      "tags": ["ops", "bad-moment"]
+    },
+    {
+      "id": "routine-pills",
+      "rule": "routine:таблетки",
+      "severity": 2,
+      "since_minutes": 0,
+      "want_any": ["таблетк", "приня", "лекарств", "pill"],
+      "tags": ["routine"],
+      "note": "A routine: rule has no fact of its own, so there is no duration to name. The rule name is the only context."
+    },
+    {
+      "id": "routine-stretch",
+      "rule": "routine:зарядка",
+      "severity": 1,
+      "since_minutes": 0,
+      "want_any": ["зарядк", "размин", "упражн", "потянис", "разомн", "stretch", "exercise"],
+      "tags": ["routine"]
+    },
+    {
+      "id": "morning-checklist",
+      "rule": "morning:утро",
+      "severity": 2,
+      "since_minutes": 0,
+      "want_any": ["утр", "чеклист", "список", "не сделан", "осталось", "morning"],
+      "tags": ["routine"],
+      "note": "In production cmd/mavend/tick.go phrases morning routines deterministically and never calls the LLM. Scored anyway: the phraser is reachable with this rule name, and a fallback that garbles it is still a bug."
+    }
+  ]
+}

internal/phraser/llmphraser.go

@@ -67,6 +67,22 @@ func NewLLMPhraser(ctx context.Context, cfg Config) (*LLMPhraser, error) {
 	return p, nil
 }
 
+// NewLLMPhraserAt wires a phraser to a llama-server that someone else started
+// and owns. It spawns nothing, so Close does not kill anything.
+//
+// This exists for the phrasing scorer (internal/phraser/eval), which must
+// measure the phrasing against a shared llama-server without taking the model
+// load hit per run or killing a server another process depends on. The daemon
+// still uses NewLLMPhraser and still owns its own child process.
+func NewLLMPhraserAt(baseURL string, cfg Config) *LLMPhraser {
+	return &LLMPhraser{
+		cfg:    cfg,
+		client: &http.Client{Timeout: cfg.Timeout},
+		port:   strings.TrimSuffix(baseURL, "/"),
+		cancel: func() {},
+	}
+}
+
 func (p *LLMPhraser) start(ctx context.Context) error {
 	args := []string{
 		"-m", p.cfg.ModelPath,
@@ -168,7 +184,10 @@ func (p *LLMPhraser) PhraseNudge(ctx context.Context, c loop.Candidate) (deliver
 		body, _ = parsePhrase(resp)
 	}
 	if body == "" {
-		body = fmt.Sprintf("%s — %s", c.Rule.Name, sevLabel(c.Severity))
+		// The model said nothing usable. Say it in Russian anyway — this text
+		// goes straight to a Russian piper voice, so the old "water — care"
+		// fallback was unspeakable.
+		body = fallbackNudge(c)
 	}
 	if mood == "" {
 		mood = "neutral"
@@ -412,8 +431,33 @@ func (p *LLMPhraser) chatWithSystem(ctx context.Context, system, user string, ma
 	return stripThink(content), nil
 }
 
+// nudgeSystem — the phrasing contract for nudges.
+//
+// Written as filled-in examples, not as a schema with "..." in it. A 0.8B
+// copies whatever sits in the response slot, so a literal placeholder there
+// teaches it to answer with the placeholder. Measured: 7/15 nudges came back
+// as "..." before this. See PHRASING-EVAL-31-07-2026.md.
+//
+// Russian only, feminine self-reference, second person masculine (the owner is
+// a man). One short sentence — the nudge is spoken aloud.
+const nudgeSystem = `Ты — Maven, домашняя ассистентка. О себе говоришь в женском роде ("я проверила", "я записала"). Владелец — мужчина, обращайся к нему в мужском роде ("ты пил", "ты забыл").
+
+Пиши ОДНО короткое напоминание по-русски: не больше 120 символов и не больше 16 слов. Только по делу.
+
+Запрещено: обращения ("дорогой", "милый"), эмодзи, извинения ("прости", "извини"), вопросы о самочувствии, похвала, больше одного восклицательного знака, английские слова кроме имён сервисов.
+
+Отвечай ТОЛЬКО одним объектом JSON с полями "response" и "mood".
+"response" — сам текст напоминания.
+"mood" — ровно одно из: neutral, happy, thinking, tired, confused.
+
+Так выглядит правильный ответ по форме. Темы здесь посторонние — их в запросе не будет:
+{"response": "Стиральная машина закончила. Развесь бельё.", "mood": "neutral"}
+{"response": "Ноутбук на трёх процентах. Я поставила его на зарядку.", "mood": "confused"}
+
+Это примеры ФОРМЫ, а не темы. Пиши только про ту ситуацию, которую тебе дали в запросе. Не копируй примеры и никогда не пиши "..." в поле response.`
+
 func (p *LLMPhraser) systemPrompt() string {
-	base := `You are maven, a self-hosted personal assistant. Generate brief, natural nudge messages in the user's language (Russian or English). Respond ONLY with valid JSON: {"response": "full voice message", "mood": "neutral"}. "response" is what the user hears; "mood" reflects maven's tone (neutral/happy/thinking/tired/confused).`
+	base := nudgeSystem
 	if p.cfg.Persona != "" {
 		base = p.cfg.Persona + "\n\n" + base
 	}
@@ -430,22 +474,117 @@ func (p *LLMPhraser) querySystemPrompt() string {
 	return base
 }
 
+// ruleTopics — Russian gloss for each built-in rule name. The rule names are
+// English identifiers; a 0.8B asked to nudge about "netdata_critical" writes
+// about nothing. The daemon knows what its own rules mean, so it says so.
+var ruleTopics = map[string]string{
+	"water":            "он давно не пил воду",
+	"meal":             "он давно не ел",
+	"break":            "он давно без перерыва, пора встать и размяться",
+	"service_down":     "сервис не отвечает, лежит",
+	"netdata_critical": "критический алярм в netdata, проблема с диском или местом",
+}
+
+// ruleKeywords — the word the message must contain. The 0.8B drifts to
+// whatever topic it saw last unless the required word is named outright.
+var ruleKeywords = map[string]string{
+	"water":            "воду",
+	"meal":             "поешь",
+	"break":            "перерыв",
+	"service_down":     "сервис",
+	"netdata_critical": "диск",
+}
+
+// ruleTopic turns a rule name into a Russian description of the situation.
+// "routine:зарядка" and "morning:утро" carry their own Russian suffix.
+func ruleTopic(rule string) string {
+	if t, ok := ruleTopics[rule]; ok {
+		return t
+	}
+	if i := strings.IndexByte(rule, ':'); i > 0 && i+1 < len(rule) {
+		switch rule[:i] {
+		case "morning":
+			return "утро, пора начать день: " + rule[i+1:]
+		default:
+			return "пора сделать по распорядку: " + rule[i+1:]
+		}
+	}
+	return rule
+}
+
+// ruleKeyword — the word the nudge must contain, or "" when the rule name's
+// own Russian suffix already is that word.
+func ruleKeyword(rule string) string {
+	if k, ok := ruleKeywords[rule]; ok {
+		return k
+	}
+	if i := strings.IndexByte(rule, ':'); i > 0 && i+1 < len(rule) {
+		return rule[i+1:]
+	}
+	return ""
+}
+
+// ruDur — duration in Russian. humanDur is English and its output was landing
+// verbatim in the message.
+func ruDur(d time.Duration) string {
+	if d < 0 {
+		d = 0
+	}
+	h, m := int(d.Hours()), int(d.Minutes())%60
+	switch {
+	case h >= 2:
+		return fmt.Sprintf("%d ч", h)
+	case h == 1 && m >= 30:
+		return "полтора часа"
+	case h == 1:
+		return "час"
+	default:
+		return fmt.Sprintf("%d мин", m)
+	}
+}
+
+// fallbackNudge — plain Russian for when the model returns nothing parseable.
+var fallbackNudges = map[string]string{
+	"water":            "Ты давно не пил воду.",
+	"meal":             "Ты давно не ел, поешь.",
+	"break":            "Пора сделать перерыв.",
+	"service_down":     "Сервис не отвечает.",
+	"netdata_critical": "Критический алярм: проверь диск.",
+}
+
+func fallbackNudge(c loop.Candidate) string {
+	if s, ok := fallbackNudges[c.Rule.Name]; ok {
+		return s
+	}
+	if kw := ruleKeyword(c.Rule.Name); kw != "" {
+		return "Напоминаю: " + kw + "."
+	}
+	return "Напоминаю о деле."
+}
+
 func buildNudgePrompt(c loop.Candidate) string {
 	var ctxParts []string
-	ctxParts = append(ctxParts, fmt.Sprintf("Rule: %s", c.Rule.Name))
+	ctxParts = append(ctxParts, "Ситуация: "+ruleTopic(c.Rule.Name))
-	ctxParts = append(ctxParts, fmt.Sprintf("Severity: %s", sevLabel(c.Severity)))
+	if f, ok := c.State.Facts[c.Rule.Name]; ok && f.Key != "" && f.Key != c.Rule.Name {
-
+		ctxParts = append(ctxParts, "Что именно: "+f.Key)
+	}
 	if d, ok := c.State.Since(c.Rule.Name); ok {
-		ctxParts = append(ctxParts, fmt.Sprintf("Duration since last event: %s", humanDur(d)))
+		ctxParts = append(ctxParts, "Прошло: "+ruDur(d))
+	}
+	switch sevLabel(c.Severity) {
+	case "alarm":
+		ctxParts = append(ctxParts, "Срочно, скажи прямо.")
+	case "ops":
+		ctxParts = append(ctxParts, "Это про сервер, не про здоровье.")
+	}
+	tail := "Напиши напоминание про эту ситуацию. Одно предложение, по-русски, в JSON."
+	if kw := ruleKeyword(c.Rule.Name); kw != "" {
+		// Last line on purpose: a 0.8B weights the end of the prompt hardest,
+		// and without the required word it drifts back to the examples.
+		tail += " Ответ ДОЛЖЕН содержать слово «" + kw + "»."
 	}
 
-	return fmt.Sprintf(
+	return strings.Join(ctxParts, "\n") + "\n\n" + tail
-		`Generate a nudge message. Context:
-%s
-
-Respond as JSON: {"response": "...", "mood": "..."}`,
-		strings.Join(ctxParts, "\n"),
-	)
 }
 
 type responseMood struct {

internal/phraser/phraser.go

@@ -92,7 +92,11 @@ func (s *Stub) Close() error { return nil }
 // the predicate fire (the same State the predicate saw).
 func (s *Stub) PhraseNudge(_ context.Context, c loop.Candidate) (delivery.PhrasedNudge, error) {
 	body, summary := phraseNudge(c)
-	return delivery.PhrasedNudge{Candidate: c, Body: body, Summary: summary}, nil
+	// "neutral" rather than empty: Mood is part of the documented output
+	// contract and the Stub is a production fallback, so it must satisfy the
+	// contract too. Template phrasing has no tone to report, and neutral is the
+	// enum's own default.
+	return delivery.PhrasedNudge{Candidate: c, Body: body, Summary: summary, Mood: "neutral"}, nil
 }
 
 // PhraseReminder — extracts the user's text from the reminder payload (raw

internal/router/embedder.go

@@ -19,6 +19,37 @@ type Embedder interface {
 	Close() error
 }
 
+// AsymmetricEmbedder — an embedder that wants to know whether a text is a
+// search query or a stored passage. Recall is asymmetric: a short question
+// goes in, a longer note comes out. The e5 family is trained for exactly that
+// and needs the side written into the text ("query: " / "passage: ").
+//
+// Optional on purpose: HashEmbedder has no such notion, so callers go through
+// EmbedQuery and EmbedPassage below, which fall back to plain Embed.
+type AsymmetricEmbedder interface {
+	Embedder
+	EmbedQuery(ctx context.Context, text string) ([]float32, error)
+	EmbedPassage(ctx context.Context, text string) ([]float32, error)
+}
+
+// EmbedQuery embeds text that is being searched WITH — a question.
+func EmbedQuery(ctx context.Context, e Embedder, text string) ([]float32, error) {
+	if a, ok := e.(AsymmetricEmbedder); ok {
+		return a.EmbedQuery(ctx, text)
+	}
+	return e.Embed(ctx, text)
+}
+
+// EmbedPassage embeds text that is being searched FOR — a note or a fact on
+// its way into the store. Store and lookup must use these two calls, not one
+// of them twice, or the asymmetry buys nothing.
+func EmbedPassage(ctx context.Context, e Embedder, text string) ([]float32, error) {
+	if a, ok := e.(AsymmetricEmbedder); ok {
+		return a.EmbedPassage(ctx, text)
+	}
+	return e.Embed(ctx, text)
+}
+
 // HashEmbedder — a deterministic bag-of-words embedder used for tests and as a
 // non-zero default floor. NOT semantically meaningful across languages; the
 // real classifier swaps in the multilingual ONNX model wholesale.

internal/router/embedder_test.go

@@ -37,3 +37,60 @@ func TestHashEmbedderCyrillic(t *testing.T) {
 		t.Fatalf("cosine(shared)=%.3f not > cosine(disjoint)=%.3f", cosine(a, b), cosine(a, c))
 	}
 }
+
+// recordingEmbedder — an asymmetric embedder that only remembers which side
+// was asked for. Enough to pin the dispatch; real vectors need the model.
+type recordingEmbedder struct{ calls []string }
+
+func (r *recordingEmbedder) Dim() int     { return 2 }
+func (r *recordingEmbedder) Close() error { return nil }
+
+func (r *recordingEmbedder) Embed(_ context.Context, _ string) ([]float32, error) {
+	r.calls = append(r.calls, "embed")
+	return []float32{1, 0}, nil
+}
+
+func (r *recordingEmbedder) EmbedQuery(_ context.Context, _ string) ([]float32, error) {
+	r.calls = append(r.calls, "query")
+	return []float32{1, 0}, nil
+}
+
+func (r *recordingEmbedder) EmbedPassage(_ context.Context, _ string) ([]float32, error) {
+	r.calls = append(r.calls, "passage")
+	return []float32{0, 1}, nil
+}
+
+// TestEmbedQueryAndPassageSplit — a question and a stored note must not take
+// the same path. If both ended up on the same call the asymmetric model buys
+// nothing, which is the whole reason for the swap.
+func TestEmbedQueryAndPassageSplit(t *testing.T) {
+	rec := &recordingEmbedder{}
+	if _, err := EmbedQuery(context.Background(), rec, "где логи?"); err != nil {
+		t.Fatalf("EmbedQuery: %v", err)
+	}
+	if _, err := EmbedPassage(context.Background(), rec, "логи в /var/log"); err != nil {
+		t.Fatalf("EmbedPassage: %v", err)
+	}
+	if len(rec.calls) != 2 || rec.calls[0] != "query" || rec.calls[1] != "passage" {
+		t.Errorf("calls %v, want [query passage]", rec.calls)
+	}
+}
+
+// TestEmbedFallsBackToPlainEmbed — HashEmbedder has no sides, so both helpers
+// must still work and give the same vector.
+func TestEmbedFallsBackToPlainEmbed(t *testing.T) {
+	h := NewHashEmbedder(64)
+	q, err := EmbedQuery(context.Background(), h, "text")
+	if err != nil {
+		t.Fatalf("EmbedQuery: %v", err)
+	}
+	p, err := EmbedPassage(context.Background(), h, "text")
+	if err != nil {
+		t.Fatalf("EmbedPassage: %v", err)
+	}
+	for i := range q {
+		if q[i] != p[i] {
+			t.Fatalf("hash embedder gave two different vectors for the same text")
+		}
+	}
+}

internal/router/eval/eval_test.go

@@ -185,8 +185,8 @@ func TestONNXBaseline(t *testing.T) {
 	if lib == "" {
 		t.Skip("MAVEN_ONNX_LIB unset — see AGENTS.md § Embedder model for intent routing")
 	}
-	model := filepath.Join("../../..", "models/embedder/model.onnx")
+	model := filepath.Join("../../..", "models/embedder/multilingual-e5-small/model_quantized.onnx")
-	tok := filepath.Join("../../..", "models/embedder/tokenizer.json")
+	tok := filepath.Join("../../..", "models/embedder/multilingual-e5-small/tokenizer.json")
 	for _, p := range []string{lib, model, tok} {
 		if _, err := os.Stat(p); err != nil {
 			t.Skipf("missing %s: %v", p, err)

internal/router/eval/llmrouter_test.go

@@ -23,7 +23,11 @@ import (
 //
 //	llama-server -m /mnt/hdd1/llms/qwen3.5/Qwen3.5-0.8B.Q4_K_M.gguf \
 //	  --host 127.0.0.1 --port 18099 -c 2048 -ngl 99
-//	MAVEN_LLM_URL=http://127.0.0.1:18099 make eval-router
+//	MAVEN_LLM_URL=http://127.0.0.1:18099 make eval-models
+//
+// Every report name carries the model llama-server reports over /v1/models, so
+// a bake-off across checkpoints (#278, #250) produces tables you can tell
+// apart. Point the variable at one server at a time.
 //
 // Three configurations, because "the LLM router" is ambiguous and the three
 // numbers answer different questions:
@@ -53,6 +57,14 @@ func TestLLMRouterBaseline(t *testing.T) {
 	}
 
 	ctx := context.Background()
+	model, err := ModelID(ctx, base)
+	if err != nil {
+		// Not fatal: an unlabelled score is still a score. But say so loudly,
+		// because an unlabelled row in a bake-off table is worthless.
+		t.Logf("could not read model id from %s: %v — reports will say %q", base, err, "unknown-model")
+		model = "unknown-model"
+	}
+	t.Logf("scoring model %s at %s", model, base)
 	lr := router.NewLLMRouter(client)
 
 	// llm-only: the LLM stage in isolation. Route returns (Decision, ok, err);
@@ -68,7 +80,7 @@ func TestLLMRouterBaseline(t *testing.T) {
 		}
 		return d, nil
 	})
-	repLLM, err := Score(ctx, "llm-only (0.8B, as deployed)", llmOnly, f)
+	repLLM, err := Score(ctx, "llm-only ("+model+", as deployed)", llmOnly, f)
 	if err != nil {
 		t.Fatalf("Score llm-only: %v", err)
 	}
@@ -77,7 +89,7 @@ func TestLLMRouterBaseline(t *testing.T) {
 	// cascade+llm: stage-0 grammar → LLM → classifier fallback, the wiring #320
 	// proposes. Hash embedder for the fallback so the classifier contribution is
 	// the deterministic floor and any lift is attributable to the model.
-	repCascade, err := Score(ctx, "cascade+llm (0.8B) + hash fallback",
+	repCascade, err := Score(ctx, "cascade+llm ("+model+") + hash fallback",
 		newBaselineRouter(t, router.NewHashEmbedder(1024), lr), f)
 	if err != nil {
 		t.Fatalf("Score cascade: %v", err)
@@ -96,7 +108,7 @@ func TestLLMRouterBaseline(t *testing.T) {
 	// either way. Kept so the question stays answered instead of being
 	// re-asked, and so internal/llm does NOT grow a chat_template_kwargs field
 	// for a problem that does not exist.
-	repNoThink, err := Score(ctx, "llm-only (0.8B, thinking off) [diagnostic]",
+	repNoThink, err := Score(ctx, "llm-only ("+model+", thinking off) [diagnostic]",
 		RouterFunc(func(ctx context.Context, u string, now time.Time) (router.Decision, error) {
 			d, ok, err := router.NewLLMRouter(&noThinkCompleter{base: base, http: &http.Client{Timeout: 60 * time.Second}}).Route(ctx, u, now)
 			if err != nil {

internal/router/eval/modelid.go

@@ -0,0 +1,51 @@
+package eval
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"strings"
+)
+
+// ModelID asks llama-server which model it has loaded, so a scoring run can
+// label itself. Without this a bake-off between two models produces two tables
+// that look identical, and the operator has to remember which server was up.
+//
+// Read from the server rather than passed in on purpose: a hand-typed label
+// goes stale the moment someone restarts the server with a different -m.
+func ModelID(ctx context.Context, base string) (string, error) {
+	req, err := http.NewRequestWithContext(ctx, "GET", strings.TrimSuffix(base, "/")+"/v1/models", nil)
+	if err != nil {
+		return "", err
+	}
+	resp, err := http.DefaultClient.Do(req)
+	if err != nil {
+		return "", err
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != 200 {
+		return "", fmt.Errorf("models: status %d", resp.StatusCode)
+	}
+	var out struct {
+		Data []struct {
+			ID string `json:"id"`
+		} `json:"data"`
+	}
+	if err := json.NewDecoder(resp.Body).Decode(&out); err != nil {
+		return "", err
+	}
+	if len(out.Data) == 0 {
+		return "", fmt.Errorf("models: empty list")
+	}
+	return shortModelID(out.Data[0].ID), nil
+}
+
+// shortModelID trims the path and the .gguf suffix — llama-server reports the
+// file name it was started with, which is too long for a table header.
+func shortModelID(id string) string {
+	if i := strings.LastIndexAny(id, "/\\"); i >= 0 {
+		id = id[i+1:]
+	}
+	return strings.TrimSuffix(id, ".gguf")
+}

internal/router/llmrouter.go

@@ -29,7 +29,7 @@ func NewLLMRouter(c Completer) *LLMRouter { return &LLMRouter{c: c} }
 const routeGrammar = `
 root   ::= "[" ws action ("," ws action)* ws "]"
 action ::= "{" ws "\"intent\"" ws ":" ws intent ("," ws field)* ws "}"
-intent ::= "\"fact\"" | "\"reminder\"" | "\"note\"" | "\"query\"" | "\"act\"" | "\"chat\"" | "\"system\""
+intent ::= "\"fact\"" | "\"reminder\"" | "\"note\"" | "\"query\"" | "\"act\"" | "\"chat\"" | "\"system\"" | "\"unknown\""
 field  ::= key ws ":" ws string
 key    ::= "\"key\"" | "\"value\"" | "\"text\"" | "\"verb\""
 string ::= "\"" ([^"\\] | "\\" .){0,120} "\""
@@ -41,12 +41,17 @@ ws     ::= [ \t\n]*
 // question naming a fact key ("сколько воды я выпил с утра") matched the fact
 // rule first and was stored as an assertion — 15 of 76 fixture cases.
 //
+// Changed again 31-07-2026: added the "unknown" escape hatch so the model can
+// admit it cannot route (Vikunja #359).
+//
 // The training workspace keeps its own copy of this prompt for relabelling, and
 // `llm/check_prompt_parity.py` there compares the two. That copy is in another
-// repo and was not touched, so parity will fail until it gets the same edit.
+// repo and was not touched, so parity will fail until it gets the same edits —
+// both the rule reorder and the "unknown" wording (Vikunja #362).
 const routeSystem = `Классифицируй ровно одно сообщение пользователя. Верни ОДИН JSON-массив действий.
 
 Ровно одно намерение: fact, reminder, note, query, act, chat, system.
+Есть восьмое значение unknown — только для случаев, когда просьбу невозможно понять.
 
 Классифицируй по цели пользователя. Порядок решения:
 1. Хочет напоминание в будущем → reminder
@@ -56,12 +61,14 @@ const routeSystem = `Классифицируй ровно одно сообще
 5. Утверждает: сообщает или обновляет текущее состояние/событие → fact
 6. Просит выполнить работу → act
 7. Про ассистента, настройки или память → system
-8. Иначе → chat
+8. Реплика — обрывок или указание на неназванное («это», «то», «потом»), и без него непонятно, что именно нужно сделать → unknown
+9. Иначе → chat
 
 Различия:
 - note — сохранить информацию, без напоминания. text = суть.
 - reminder — уведомить позже. text = что напомнить.
 - fact — неявное обновление: пользователь сообщает, что что-то в мире изменилось (текущее/изменённое состояние, случившееся событие). key/value.
+- unknown — редкий случай. Ставь его, только если в самой реплике нет ни предмета, ни действия. Короткая, простая или незнакомая тема — это не причина для unknown: приветствие и болтовня — это chat, вопрос на любую тему — это query, просьба сделать что-то названное — это act.
 - query против fact — решает форма реплики, а не тема. Вопрос о состоянии — это query, даже если названо то же самое, что бывает в fact. Только утверждение — это fact.
 
 Примеры:
@@ -76,6 +83,13 @@ const routeSystem = `Классифицируй ровно одно сообще
 "напиши письмо" → {"intent":"act","verb":"написать письмо"}
 "очисти память" → {"intent":"system"}
 "привет" → {"intent":"chat","text":"привет"}
+"сделай это" → {"intent":"unknown"}
+"ну это" → {"intent":"unknown"}
+"потом" → {"intent":"unknown"}
+Но не путай — здесь unknown не нужен:
+"сделай кофе" → {"intent":"act","verb":"сделать кофе"}
+"что такое кватернион?" → {"intent":"query","text":"что такое кватернион"}
+"ага" → {"intent":"chat","text":"ага"}
 
 Ответ — JSON-массив: по одному объекту на каждую просьбу. Обычно один. Если в реплике несколько просьб — по объекту на каждую. "напомни купить молоко, и запиши что кофе кончился" → [{"intent":"reminder","text":"купить молоко"},{"intent":"note","text":"кофе кончился"}]. Только JSON, без пояснений.`
 
@@ -84,6 +98,11 @@ const routeSystem = `Классифицируй ровно одно сообще
 // the loop without hurting short slot values.
 const routeRepeatPenalty = 1.15
 
+// routeIntentUnknown — the model's way of saying "I could not route this".
+// It is a wire value only: it never becomes a router.Intent, it just makes
+// Route return ok=false so the caller drops to the classifier cascade.
+const routeIntentUnknown = "unknown"
+
 type routeAction struct {
 	Intent string `json:"intent"`
 	Key    string `json:"key"`
@@ -92,6 +111,10 @@ type routeAction struct {
 	Verb   string `json:"verb"`
 }
 
+// Route asks the model for one decision. The bool is false when there is no
+// decision to use: either the model failed (err set) or it refused with the
+// "unknown" intent (err nil). Both mean the same thing to the caller — use the
+// classifier instead.
 func (lr *LLMRouter) Route(ctx context.Context, utterance string, now time.Time) (Decision, bool, error) {
 	raw, err := lr.c.Complete(ctx, llm.Req{
 		System:        routeSystem,
@@ -115,6 +138,13 @@ func (lr *LLMRouter) Route(ctx context.Context, utterance string, now time.Time)
 	// with the engine turn-on (Router.Route → []Decision, both voice.go handlers
 	// loop). Until then only the first ask is honored.
 	a := acts[0]
+	// The model refused. Report "no decision" without an error, which is the
+	// same fall-through the caller already uses for a parse failure — the
+	// classifier cascade gets the turn and its own confidence gate decides
+	// whether to ask. Better a slower second opinion than a confident guess.
+	if a.Intent == routeIntentUnknown {
+		return Decision{}, false, nil
+	}
 	d := Decision{Utterance: utterance, Stage: 1, Confidence: 1.0}
 	switch Intent(a.Intent) {
 	case IntentFact:

internal/router/llmrouter_test.go

@@ -100,8 +100,10 @@ func TestLLMRouterReminderMapping(t *testing.T) {
 	}
 }
 
+// An intent name that is not in the contract at all (as opposed to "unknown",
+// which is a real refusal) still defaults to chat.
 func TestLLMRouterChatFallback(t *testing.T) {
-	lr := NewLLMRouter(mockLLM{out: `{"intent":"unknown"}`})
+	lr := NewLLMRouter(mockLLM{out: `{"intent":"banana"}`})
 	d, ok, err := lr.Route(context.Background(), "как дела?", time.Now())
 	if err != nil || !ok {
 		t.Fatalf("ok=%v err=%v", ok, err)
@@ -111,6 +113,61 @@ func TestLLMRouterChatFallback(t *testing.T) {
 	}
 }
 
+// The model must be able to say "I could not route this".
+func TestRouteGrammarAllowsUnknown(t *testing.T) {
+	if !strings.Contains(routeGrammar, `"\"unknown\""`) {
+		t.Fatal("grammar cannot express a refusal")
+	}
+}
+
+// If the prompt does not tell the model when to refuse, it never will.
+func TestRoutePromptExplainsUnknown(t *testing.T) {
+	if !strings.Contains(routeSystem, "unknown") {
+		t.Fatal("prompt never mentions the unknown intent")
+	}
+	if !strings.Contains(routeSystem, `"сделай это" → {"intent":"unknown"}`) {
+		t.Fatal("prompt lost its worked refusal example")
+	}
+	// A refusal-only router is useless, so the prompt must also show cases that
+	// look ambiguous but are not.
+	if !strings.Contains(routeSystem, "здесь unknown не нужен") {
+		t.Fatal("prompt lost its counter-examples")
+	}
+}
+
+// A refusal is not an error. It reports "no decision" so the cascade moves on.
+func TestLLMRouterUnknownRefuses(t *testing.T) {
+	lr := NewLLMRouter(mockLLM{out: `{"intent":"unknown"}`})
+	_, ok, err := lr.Route(context.Background(), "сделай это", time.Now())
+	if ok {
+		t.Fatal("a refusal must not produce a usable decision")
+	}
+	if err != nil {
+		t.Fatalf("a refusal is not an error, got %v", err)
+	}
+}
+
+// The whole point of the refusal: the turn keeps going on the classifier, the
+// same way it does when the model returns garbage.
+func TestRouterFallsBackWhenLLMRefuses(t *testing.T) {
+	c := NewClassifier(NewHashEmbedder(1024))
+	seedClassifier(t, c)
+	r := New(Config{
+		Classifier: c,
+		Extractor:  Extractor{Time: StubDateTimeParser{}, Facts: DefaultFactParser{}},
+		Threshold:  0.4,
+		LLM:        NewLLMRouter(mockLLM{out: `{"intent":"unknown"}`}),
+	})
+	d, err := r.Route(context.Background(), "напомни позвонить маме", refNow())
+	if err != nil {
+		t.Fatalf("route: %v", err)
+	}
+	// Stage 1 is the LLM's own answer; the classifier lands on stage 2 or 3.
+	if d.Stage < 2 {
+		t.Fatalf("want the classifier to decide, got stage %d (%+v)", d.Stage, d)
+	}
+}
+
 func TestLLMRouterLLMError(t *testing.T) {
 	lr := NewLLMRouter(mockLLM{out: "", err: fmt.Errorf("llm down")})
 	_, ok, err := lr.Route(context.Background(), "x", time.Now())
@@ -118,3 +175,87 @@ func TestLLMRouterLLMError(t *testing.T) {
 		t.Fatal("want ok=false, err!=nil on llm error")
 	}
 }
+
+// --- slot extraction on top of an LLM decision --------------------------------
+
+// newLLMTestRouter — a router whose route always comes from the mock model.
+func newLLMTestRouter(t *testing.T, out string) *Router {
+	t.Helper()
+	c := NewClassifier(NewHashEmbedder(1024))
+	seedClassifier(t, c)
+	acts := DefaultActMatcher{Fns: []string{"restart", "stop", "run", "backup"}}
+	return New(Config{
+		Classifier: c,
+		Extractor:  Extractor{Time: StubDateTimeParser{}, Acts: acts, Facts: DefaultFactParser{}},
+		Threshold:  0.4,
+		LLM:        NewLLMRouter(mockLLM{out: out}),
+	})
+}
+
+// The model cannot produce a fire time, so without extraction every LLM-routed
+// reminder was dropped as "no time".
+func TestLLMDecisionGetsReminderTime(t *testing.T) {
+	r := newLLMTestRouter(t, `{"intent":"reminder","text":"позвонить маме"}`)
+	d, err := r.Route(context.Background(), "напомни позвонить маме через 2 часа", refNow())
+	if err != nil {
+		t.Fatalf("route: %v", err)
+	}
+	if d.Intent != IntentReminder {
+		t.Fatalf("want reminder, got %v", d.Intent)
+	}
+	if !d.Slots.HasTime || !d.Slots.Time.Equal(refNow().Add(2*time.Hour)) {
+		t.Fatalf("want time now+2h, got %+v", d.Slots)
+	}
+	if d.Slots.Text != "позвонить маме" {
+		t.Fatalf("extraction overwrote the model's text: %q", d.Slots.Text)
+	}
+}
+
+// No time in the utterance ⇒ no time in the slots. Do not invent one; the
+// daemon says it could not read the time.
+func TestLLMReminderWithoutTimeStaysEmpty(t *testing.T) {
+	r := newLLMTestRouter(t, `{"intent":"reminder","text":"позвонить маме"}`)
+	d, err := r.Route(context.Background(), "напомни позвонить маме", refNow())
+	if err != nil {
+		t.Fatalf("route: %v", err)
+	}
+	if d.Slots.HasTime {
+		t.Fatalf("invented a time: %v", d.Slots.Time)
+	}
+}
+
+// An act decision arrived with no Fn, so the tool never ran.
+func TestLLMDecisionGetsActFn(t *testing.T) {
+	r := newLLMTestRouter(t, `{"intent":"act","verb":"restart nginx"}`)
+	d, err := r.Route(context.Background(), "слушай, restart nginx пожалуйста", refNow())
+	if err != nil {
+		t.Fatalf("route: %v", err)
+	}
+	if !d.Slots.HasFn || d.Slots.Fn != "restart" || len(d.Slots.Args) != 1 || d.Slots.Args[0] != "nginx" {
+		t.Fatalf("want fn=restart args=[nginx], got %+v", d.Slots)
+	}
+}
+
+// The model's own slots win; extraction only fills gaps.
+func TestLLMSlotsWinOverExtraction(t *testing.T) {
+	r := newLLMTestRouter(t, `{"intent":"fact","key":"hydration","value":"выпил"}`)
+	d, err := r.Route(context.Background(), "я выпил воду", refNow())
+	if err != nil {
+		t.Fatalf("route: %v", err)
+	}
+	if d.Slots.Key != "hydration" {
+		t.Fatalf("extraction overwrote the model's key: %q", d.Slots.Key)
+	}
+}
+
+// A fact the model left keyless still gets one from the parser.
+func TestLLMFactGetsKeyFromParser(t *testing.T) {
+	r := newLLMTestRouter(t, `{"intent":"fact","text":"я выпил воду"}`)
+	d, err := r.Route(context.Background(), "я выпил воду", refNow())
+	if err != nil {
+		t.Fatalf("route: %v", err)
+	}
+	if !d.Slots.HasKey || d.Slots.Key != "water" {
+		t.Fatalf("want key=water, got %+v", d.Slots)
+	}
+}

internal/router/onnxembedder.go

@@ -12,6 +12,15 @@ import (
 	"golang.org/x/text/unicode/norm"
 )
 
+// The deployed model is multilingual-e5-small. e5 was trained with these two
+// words glued to the front of every text, and it scores badly without them —
+// they are part of the model, not a style choice. Swapping back to a symmetric
+// paraphrase model means dropping them again.
+const (
+	queryPrefix   = "query: "
+	passagePrefix = "passage: "
+)
+
 const (
 	padTokenID = 1
 	unkTokenID = 3
@@ -54,7 +63,25 @@ func NewONNXEmbedder(modelPath, tokenizerPath, libPath string) (*onnxEmbedder, e
 
 func (e *onnxEmbedder) Dim() int { return embedDim }
 
+// Embed treats the text as a query. The classifier compares one short
+// utterance to another short seed phrase, so both sides get the same prefix
+// and the comparison stays fair. The recall path must call EmbedQuery and
+// EmbedPassage instead.
 func (e *onnxEmbedder) Embed(ctx context.Context, text string) ([]float32, error) {
+	return e.embed(ctx, queryPrefix+text)
+}
+
+// EmbedQuery — the question the user just asked.
+func (e *onnxEmbedder) EmbedQuery(ctx context.Context, text string) ([]float32, error) {
+	return e.embed(ctx, queryPrefix+text)
+}
+
+// EmbedPassage — a note or fact being stored, or re-scored at lookup time.
+func (e *onnxEmbedder) EmbedPassage(ctx context.Context, text string) ([]float32, error) {
+	return e.embed(ctx, passagePrefix+text)
+}
+
+func (e *onnxEmbedder) embed(ctx context.Context, text string) ([]float32, error) {
 	inputIDs, attentionMask, _ := e.tokenizer.Encode(text)
 
 	inputShape := ort.NewShape(1, int64(maxLength))
@@ -313,4 +340,4 @@ func preTokenize(text string) []string {
 	return out
 }
 
-var _ Embedder = (*onnxEmbedder)(nil)
+var _ AsymmetricEmbedder = (*onnxEmbedder)(nil)

internal/router/router.go

@@ -88,6 +88,7 @@ func (r *Router) Route(ctx context.Context, utterance string, now time.Time) (De
 	if r.llm != nil {
 		if d, ok, err := r.llm.Route(ctx, utterance, now); err == nil && ok {
 			d.Utterance = utterance
+			r.fillSlots(ctx, &d, now)
 			return d, nil
 		} else if err != nil {
 			log.Printf("router: llm route fell back to classifier: %v", err)
@@ -118,6 +119,39 @@ func (r *Router) Route(ctx context.Context, utterance string, now time.Time) (De
 	return d, nil
 }
 
+// fillSlots — run stage-2 extraction on an LLM decision and fill only the slots
+// the model left empty. The LLM wins where it answered: it saw the sentence, the
+// parsers are keyword tables. Extraction covers what the model cannot produce at
+// all — a parsed reminder time and an allowlist fn.
+//
+// If a reminder still has no time, leave it missing. The daemon then says it
+// could not read the time; inventing one would set a wrong alarm.
+func (r *Router) fillSlots(ctx context.Context, d *Decision, now time.Time) {
+	ex := r.extractor.Extract(ctx, d.Intent, d.Utterance, now)
+	if !d.Slots.HasTime && ex.HasTime {
+		d.Slots.Time, d.Slots.HasTime = ex.Time, ex.HasTime
+	}
+	if !d.Slots.HasKey && ex.HasKey {
+		d.Slots.Key, d.Slots.Value, d.Slots.HasKey = ex.Key, ex.Value, ex.HasKey
+	}
+	if !d.Slots.HasFn && ex.HasFn {
+		d.Slots.Fn, d.Slots.Args, d.Slots.HasFn = ex.Fn, ex.Args, ex.HasFn
+	}
+	// For an act the model returns the verb in Text ("restart nginx"), which is
+	// often cleaner than the raw utterance ("maven, could you restart nginx").
+	// Try it too when the utterance did not match the allowlist.
+	if d.Intent == IntentAct && !d.Slots.HasFn && r.extractor.Acts != nil &&
+		d.Slots.Text != "" && d.Slots.Text != d.Utterance {
+		if fn, args, ok := r.extractor.Acts.Match(d.Slots.Text); ok {
+			d.Slots.Fn, d.Slots.Args, d.Slots.HasFn = fn, args, true
+		}
+	}
+	if d.Slots.Text == "" {
+		d.Slots.Text = ex.Text
+	}
+	// Stage stays 1: it says who decided the route, and that was the LLM.
+}
+
 // CorrectMisroute — the user corrected a bad classification. Appends a new
 // example for the corrected intent (append-only — grows the classifier, no
 // retrain). Same shape as nudges.outcome tuning cooldowns: more reliable over