router: a Russian verb is a whole sentence, not thin evidence

The clarify gate thinned any one-word utterance to 0.3 confidence, which
trips the stage-3 gate and comes back as "не совсем поняла". That is an
English intuition. Russian packs subject, tense and gender into one word,
so "поужинал" is a complete report and "привет" a complete greeting, and
both got clarified.

thinSingleToken keeps the rule for bare nominals, where it is real ("вода"
is a fact-or-query coin flip), and spares two classes: a closed lexicon of
social and control singles, and any token carrying a verb ending. Both
tests are offline.

Fixture: false clarifies 3 → 2, intent-only 74.0% → 75.3%, full accuracy
unchanged at 70.1%, missed clarify still 1.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
This commit is contained in:
kami
2026-08-01 21:29:21 +04:00
parent 47dda97226
commit 29329b5f0e
4 changed files with 138 additions and 6 deletions
+10 -2
View File
@@ -102,8 +102,16 @@ Re-measured on the fixture after the fix: **missed clarify 6/6 → 1**, at the c
clarifies and 2.6pt of full accuracy (72.7% → 70.1%, intent-only 67.5% → 74.0%). Two of the
three false clarifies are acts the model mis-routed and the gate caught — asking beats wrongly
executing, so the fixture and the daemon disagree about what is correct there. The third,
`"поужинал"`, is a real defect: **the single-token rule is an English intuition and does not
transfer to Russian**, where one word is routinely a whole sentence. Narrow or drop it.
`"поужинал"`, was a real defect: the single-token rule was an English intuition and does not
transfer to Russian, where one word is routinely a whole sentence.
Narrowed 01-08-2026. `thinSingleToken` (`internal/router/singletoken.go`) still thins a bare
one-word nominal — "вода", "бэкап" — but spares two classes: a closed lexicon of social and
control singles ("привет", "спасибо", "стоп", "yes"), and any token carrying a Russian verb
ending (past tense, 2nd person, reflexive), because a verb already contains its subject. Both
tests are offline and cost nothing. Re-measured: **false clarifies 3 → 2, intent-only 74.0% →
75.3%, full accuracy unchanged at 70.1%, missed clarify still 1.** The two remaining false
clarifies are the act-with-no-allowlisted-fn arm of the gate, not this rule.
## LLM output contract
+5 -4
View File
@@ -194,12 +194,13 @@ func (lr *LLMRouter) Route(ctx context.Context, utterance string, now time.Time)
return Decision{}, false, nil
}
d := Decision{Utterance: utterance, Stage: 1, Confidence: llmFullConfidence}
// A single-token utterance is thin evidence: the model had nothing to
// A bare one-word nominal is thin evidence: the model had nothing to
// disambiguate on ("вода" is a fact-or-query coin flip, "бэкап" an
// act-or-report one) and stage 0 would already have won on anything
// that pattern-matches cleanly. Flag it now; router.go's stage-3 gate
// (Router.Route) decides whether that trips Clarify.
if len(strings.Fields(utterance)) <= 1 {
// that pattern-matches cleanly. A greeting or an inflected verb is NOT
// thin, however short — see thinSingleToken. Flag it now; router.go's
// stage-3 gate (Router.Route) decides whether that trips Clarify.
if thinSingleToken(utterance) {
d.Confidence = llmThinConfidence
}
switch Intent(a.Intent) {
+90
View File
@@ -0,0 +1,90 @@
package router
import "strings"
// thinSingleToken — is a one-word utterance thin evidence, or is it a whole
// sentence?
//
// The rule this replaces was `len(strings.Fields(u)) <= 1`, an English
// intuition. It does not transfer: Russian packs a subject, a tense and a
// gender into one word, so "поужинал" is a complete report and "привет" a
// complete greeting, yet both got thinned and came back as "не совсем поняла".
// Meanwhile the case the rule exists for is real — a bare noun like "вода" or
// "бэкап" genuinely does not say fact-vs-query or act-vs-report.
//
// So: still one token, but only thin it when the token is a bare nominal.
// Two escapes, both cheap and both offline:
//
// - a closed lexicon of social and command singles, which are complete by
// definition ("привет", "спасибо", "стоп", "yes");
// - a suffix test for an inflected predicate — past tense, 2nd person,
// reflexive. Verbs carry their own subject, so a verb IS a sentence.
//
// The suffix test is deliberately loose about nouns that happen to end the
// same way ("канал" reads as past tense here). That direction of error only
// costs a clarify we would not have asked for; the other direction — treating
// a real report as thin — is the bug being fixed.
func thinSingleToken(utterance string) bool {
f := strings.Fields(utterance)
if len(f) != 1 {
return false
}
w := strings.ToLower(strings.Trim(f[0], ".,!?;:—-\"'«»()"))
if w == "" {
return false
}
if completeSingles[w] {
return false
}
return !looksInflected(w)
}
// completeSingles — one-word utterances that need no second half. Greetings,
// acknowledgements and the control words a voice loop has to honour instantly.
var completeSingles = map[string]bool{
// ru: social
"привет": true, "здравствуй": true, "здравствуйте": true, "здорово": true,
"пока": true, "прощай": true, "спокойной": true, "спасибо": true,
"благодарю": true, "извини": true, "прости": true, "пожалуйста": true,
"да": true, "нет": true, "ага": true, "угу": true, "ок": true, "окей": true,
"хорошо": true, "ладно": true, "конечно": true, "верно": true, "точно": true,
// ru: control
"стоп": true, "отмена": true, "отбой": true, "хватит": true, "тихо": true,
"повтори": true, "продолжай": true, "помоги": true, "помощь": true,
// en
"hi": true, "hello": true, "hey": true, "bye": true, "goodbye": true,
"thanks": true, "thank": true, "sorry": true, "please": true,
"yes": true, "no": true, "yep": true, "nope": true, "ok": true, "okay": true,
"sure": true, "right": true, "stop": true, "cancel": true, "help": true,
"repeat": true, "continue": true,
}
// inflectedSuffixes — endings that mark a finite or past-tense Russian verb.
// Ordered longest-first is unnecessary (any match wins), but each entry is
// chosen to be long enough that common nouns rarely collide.
var inflectedSuffixes = []string{
// reflexive — strongly verbal whatever precedes it
"ся", "сь",
// past tense
"ал", "ял", "ил", "ел", "ыл", "ул", "ёл", "ала", "яла", "ила", "ела",
"ыла", "ула", "али", "яли", "или", "ели",
// 2nd person singular
"ешь", "ишь", "ёшь",
// 1st/2nd person plural, 3rd person plural
"аем", "яем", "уем", "аете", "ите", "ают", "яют", "уют", "ат", "ят",
}
// looksInflected — does the word carry a verb ending? Short words are exempt:
// a three-letter token is not enough stem to trust a two-letter suffix on
// ("газ" would otherwise never match, but "нос" and "лес" would).
func looksInflected(w string) bool {
if len([]rune(w)) < 5 {
return false
}
for _, s := range inflectedSuffixes {
if strings.HasSuffix(w, s) {
return true
}
}
return false
}
+33
View File
@@ -0,0 +1,33 @@
package router
import "testing"
func TestThinSingleTokenThinsBareNominals(t *testing.T) {
// The case the rule exists for: one noun, no way to tell what was asked.
for _, w := range []string{"вода", "бэкап", "нексус", "почта", "backup"} {
if !thinSingleToken(w) {
t.Errorf("thinSingleToken(%q) = false, want true", w)
}
}
}
func TestThinSingleTokenSparesCompleteUtterances(t *testing.T) {
// Regression: every one of these used to be answered with
// "не совсем поняла — можешь переформулировать?".
for _, w := range []string{
"привет", "Привет!", "спасибо", "да", "нет", "стоп", "hello", "yes",
"поужинал", "проснулась", "устал", "выспался", "договорились",
} {
if thinSingleToken(w) {
t.Errorf("thinSingleToken(%q) = true, want false", w)
}
}
}
func TestThinSingleTokenIgnoresMultiWord(t *testing.T) {
for _, s := range []string{"выпил воды", "что там с бэкапом", ""} {
if thinSingleToken(s) {
t.Errorf("thinSingleToken(%q) = true, want false", s)
}
}
}