@2kw/ai 7.4.0-dev.4 → 7.4.0-dev.6

package/dist/commands/experiments.js

@@ -17,13 +17,22 @@ function formatComparisonCell(result) {
     const scores = result.scores ?? [];
     if (scores.length > 0) {
         // A SKIP carries a placeholder 0.0 that is not a rating, and a comment-only
-        // score has no number; neither takes part in the average (#693).
-        const rated = scores.filter((s) => s.label !== "SKIP" && s.score != null);
+        // score rates nothing; neither is picked (#693). Of the rest, one score per
+        // evaluator, as the console's pickCellScore: HUMAN, then LLM_JUDGE, then the
+        // first, so an override replaces the score it supersedes (#1410).
+        const byEvaluator = new Map();
+        for (const s of scores) {
+            if (s.label === "SKIP" || (s.score == null && !s.label))
+                continue;
+            byEvaluator.set(s.evaluatorId, [...(byEvaluator.get(s.evaluatorId) ?? []), s]);
+        }
+        const picked = [...byEvaluator.values()].map((group) => group.find((s) => s.source === "HUMAN") ?? group.find((s) => s.source === "LLM_JUDGE") ?? group[0]);
+        const rated = picked.filter((s) => s.score != null);
         if (rated.length > 0) {
             return (rated.reduce((sum, s) => sum + s.score, 0) / rated.length).toFixed(2);
         }
         // A label-only verdict (#1163) has no number; show its label, as the console does.
-        const verdict = scores.find((s) => s.label && s.label !== "SKIP");
+        const verdict = picked[0];
         if (verdict)
             return String(verdict.label);
         if (scores.every((s) => s.label === "SKIP"))

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@2kw/ai",
-  "version": "7.4.0-dev.4",
+  "version": "7.4.0-dev.6",
   "description": "CLI for 2kw.ai — schema-driven document extraction, an OpenAI-compatible EU LLM gateway, transcription, prompts, datasets, and experiments from your terminal or agentic workflows. Ships as 2kw, backbone, and bb.",
   "keywords": [
     "cli",