admin/frontend/controls.html

@@ -877,9 +877,12 @@ function classifierModeFieldsHtml(mode, data) {
           </select>
         </div>
         <div class="col-md-3">
-          <input class="form-control form-control-sm" type="number" step="0.05" min="0" max="1"
+          <div class="input-group input-group-sm">
-                 id="classifier-encoder-threshold" placeholder="confidence_threshold"
+            <input class="form-control form-control-sm" type="number" step="1" min="0" max="100"
-                 value="${enc.confidence_threshold != null ? enc.confidence_threshold : ''}">
+                   id="classifier-encoder-threshold" placeholder="confidence % (0-100)"
+                   value="${enc.confidence_threshold != null ? Math.round(enc.confidence_threshold * 100) : ''}">
+            <span class="input-group-text">%</span>
+          </div>
         </div>
       </div>`;
   }
@@ -967,8 +970,10 @@ function collectClassifierConfigBody() {
     const model = document.getElementById('classifier-encoder-model').value.trim();
     if (model) encoder.model = model;
     encoder.device = document.getElementById('classifier-encoder-device').value;
-    const threshold = document.getElementById('classifier-encoder-threshold').value;
+    const thresholdPct = document.getElementById('classifier-encoder-threshold').value;
-    if (threshold !== '') encoder.confidence_threshold = parseFloat(threshold);
+    // The field is 0-100 for a human to type ("80" meaning 80%); the backend
+    // wants the 0.0-1.0 probability classify_zero_shot actually returns.
+    if (thresholdPct !== '') encoder.confidence_threshold = parseFloat(thresholdPct) / 100;
     body.encoder = encoder;
   }
   return body;

src/config.py

@@ -796,9 +796,27 @@ class LocalEncoderConfig(StrictModel):
     # Below this, the classification is treated as a FAILURE, not a low-
     # confidence answer -- the caller cascades exactly as it would for a
     # local-LLM parse failure, rather than confidently mis-routing on a
-    # guess the encoder itself was unsure about.
+    # guess the encoder itself was unsure about. classify_zero_shot returns
+    # a 0.0-1.0 probability, so this must be too -- a percent-style value
+    # (e.g. 80 meaning "80%") silently makes every real confidence score
+    # read as below-threshold, since no probability exceeds 1.0. Caught live
+    # 2026-09-06: the admin UI took a raw number with no conversion or
+    # bound, so typing the intuitive "80" broke classification on every
+    # request. The UI now converts 0-100 to 0.0-1.0 before saving; this
+    # validator is the fail-closed backstop for any other caller.
     confidence_threshold: float = 0.5
 
+    @field_validator("confidence_threshold")
+    @classmethod
+    def confidence_threshold_in_range(cls, v: float) -> float:
+        if not (0.0 <= v <= 1.0):
+            raise ValueError(
+                f"classifier.encoder.confidence_threshold must be in [0.0, 1.0], "
+                f"got {v!r} -- classify_zero_shot returns a 0-1 probability, "
+                f"not a percent"
+            )
+        return v
+
 
 class ClassifierConfig(StrictModel):
     provider: str

tests/test_admin_frontend.py

@@ -133,6 +133,21 @@ def test_admin_controls_has_a_dedicated_classifier_card(admin_client):
     assert "(baseUrl && model) ? { base_url: baseUrl, model } : null" in text
 
 
+def test_admin_controls_confidence_threshold_field_is_percent_with_conversion(admin_client):
+    """The encoder confidence_threshold field takes 0-100 (a human types "80"
+    meaning 80%) and converts to the 0.0-1.0 probability classify_zero_shot
+    actually returns -- typing the intuitive percent value used to be stored
+    literally, so no real confidence score could ever clear the threshold
+    and every classification silently failed (live 2026-09-06)."""
+    resp = admin_client.get("/admin/controls")
+    text = resp.text
+    assert 'id="classifier-encoder-threshold"' in text
+    assert 'max="100"' in text
+    assert "parseFloat(thresholdPct) / 100" in text
+    # Loading back a stored 0.0-1.0 value must display it as a percent.
+    assert "Math.round(enc.confidence_threshold * 100)" in text
+
+
 def test_admin_models_returns_html_with_availability_marker(admin_client):
     """GET /admin/models returns 200, text/html, and contains the Model
     Availability card title."""

tests/test_classifier_modes_config.py

@@ -139,6 +139,36 @@ def test_local_encoder_rejects_unknown_device(raw):
         RouterConfig(**cfg)
 
 
+def test_local_encoder_rejects_percent_style_confidence_threshold(raw):
+    """classify_zero_shot returns a 0.0-1.0 probability, so a percent-style
+    value (e.g. 80 meaning "80%") must be rejected -- otherwise no real
+    confidence score can ever clear the threshold and every classification
+    silently fails. Caught live 2026-09-06 via the admin UI taking a raw
+    number with no conversion."""
+    cfg = copy.deepcopy(raw)
+    cfg["classifier"]["mode"] = "local_encoder"
+    cfg["classifier"]["encoder"] = {"confidence_threshold": 80}
+    with pytest.raises(ValueError, match=r"must be in \[0.0, 1.0\]"):
+        RouterConfig(**cfg)
+
+
+def test_local_encoder_rejects_negative_confidence_threshold(raw):
+    cfg = copy.deepcopy(raw)
+    cfg["classifier"]["mode"] = "local_encoder"
+    cfg["classifier"]["encoder"] = {"confidence_threshold": -0.1}
+    with pytest.raises(ValueError, match=r"must be in \[0.0, 1.0\]"):
+        RouterConfig(**cfg)
+
+
+def test_local_encoder_accepts_confidence_threshold_at_bounds(raw):
+    cfg = copy.deepcopy(raw)
+    cfg["classifier"]["mode"] = "local_encoder"
+    cfg["classifier"]["encoder"] = {"confidence_threshold": 0.0}
+    assert RouterConfig(**cfg).classifier.encoder.confidence_threshold == 0.0
+    cfg["classifier"]["encoder"] = {"confidence_threshold": 1.0}
+    assert RouterConfig(**cfg).classifier.encoder.confidence_threshold == 1.0
+
+
 def test_local_encoder_unaffected_by_the_cloud_llm_validator(raw):
     """A local_encoder config leaving cloud_primary/auto both unset must not
     trip the cloud_llm validator -- it's scoped to mode == cloud_llm."""