Update empty tag

moe.py

@@ -146,7 +146,11 @@ def _aggregate_field(
         winner, ratio = _priority(values_by_model, priority)
     else:  # vote
         winner, ratio = _vote(values_by_model, priority)
-    if ratio < 1.0 and len(values_by_model) > 1:
+    # Only flag *task-meaningful* disagreements. `min` aggregator (confidence) and
+    # `priority` aggregator (comment / free-text metadata) always differ across
+    # models — they would drown the user in noise.
+    quiet_aggregators = {"min", "priority"}
+    if ratio < 1.0 and len(values_by_model) > 1 and aggregator not in quiet_aggregators:
         disagreements.append(
             DisagreementCell(
                 token_idx=token_idx,

static/app.js

@@ -61,6 +61,7 @@ function annotator() {
       customTaskName: 'My custom task',
       customTagInput: '',
       customTags: [],
+      includeNone: true,
       includeConfidence: true,
       includeComment: false,
     },
@@ -422,10 +423,14 @@ function annotator() {
 
     buildCustomSchema() {
       const fields = [];
+      const baseTags = this.pasteEditor.customTags.slice();
+      const values = this.pasteEditor.includeNone
+        ? (baseTags.includes('O') ? baseTags : ['O', ...baseTags])
+        : baseTags;
       fields.push({
         name: 'tag',
         type: 'enum',
-        values: this.pasteEditor.customTags.slice(),
+        values,
         nullable: false,
         aggregator: 'vote',
         subfields: [],

static/index.html

@@ -586,6 +586,12 @@
                       :disabled="!(pasteEditor.customTagInput || '').trim()">+ Add</button>
             </div>
             <p class="text-[11px] text-ink-500 mt-1">Press <kbd>Enter</kbd>, <kbd>,</kbd> or <kbd>;</kbd> to add. You can also paste <code>PER, LOC, ORG, WORK</code> and press Enter to add them all at once.</p>
+            <label class="flex items-center gap-2 mt-3 text-sm cursor-pointer">
+              <input type="checkbox" x-model="pasteEditor.includeNone">
+              <span>Add an <strong>"O"</strong> tag for tokens that match <em>none</em> of your labels
+                <span class="text-ink-500">(recommended — without it, the LLM is forced to pick a label for every token)</span>
+              </span>
+            </label>
           </div>
           <div>
             <label class="lbl">Self-reported metadata, per token</label>