diff --git a/README.md b/README.md
index 4f83e9b..6dbc420 100644
--- a/README.md
+++ b/README.md
@@ -249,8 +249,12 @@ answerlayer optimize start \
   --mode recommendation
 ```
 
-By default the proposal model also reviews bounded failed-case traces during
-failure analysis. Use `--analysis-model MODEL` to select a dedicated reviewer.
+The shared `--model` defaults the eval agent, judge, trace analyzer, and
+component author. Use `--eval-model`, `--judge-model`, `--analysis-model`, or
+`--proposal-model` for explicit role overrides. If `--analysis-model` is
+omitted, trace analysis uses the explicit `--proposal-model` before falling
+back to the shared model. Without a shared model or `--judge-model`, each case
+keeps its configured judge.
 
 Inspect and operate the registry with `answerlayer optimize list`,
 `get RUN_ID`, `pause RUN_ID`, `resume RUN_ID`, `stop RUN_ID`, `clone RUN_ID`,
diff --git a/src/cli.js b/src/cli.js
index 19af96e..f0d5395 100644
--- a/src/cli.js
+++ b/src/cli.js
@@ -785,6 +785,7 @@ async function handleOptimization(client, command, positionals, parsed, io) {
       promotion_policy: firstValue(parsed.flags.promotionPolicy) || "manual",
       models: {
         eval_execution: firstValue(parsed.flags.evalModel) || sharedModel,
+        judge: firstValue(parsed.flags.judgeModel) || sharedModel,
         proposal: firstValue(parsed.flags.proposalModel) || sharedModel,
         analysis: firstValue(parsed.flags.analysisModel)
           || firstValue(parsed.flags.proposalModel)
@@ -1646,6 +1647,7 @@ function normalizeFlagName(rawName) {
     "--review-policy": "reviewPolicy",
     "--promotion-policy": "promotionPolicy",
     "--eval-model": "evalModel",
+    "--judge-model": "judgeModel",
     "--analysis-model": "analysisModel",
     "--proposal-model": "proposalModel",
     "--base-hash": "baseHash",
@@ -1904,7 +1906,8 @@ Data products:
     and --use-semantic-layer or --no-semantic-layer
   Optimization:
     answerlayer optimize start --name NAME --connection ID --suite ID --model MODEL
-    [--eval-model MODEL] [--proposal-model MODEL] [--analysis-model MODEL]
+    [--eval-model MODEL] [--judge-model MODEL] [--analysis-model MODEL]
+    [--proposal-model MODEL]
     [--mode recommendation|supervised|fully_automatic]
     [--component entities,relationships,measures,metrics,dimensions,filters]
     [--review-policy every_sweep|exceptions_only|end_of_run]
diff --git a/test/cli.test.js b/test/cli.test.js
index 23c20d9..87e3636 100644
--- a/test/cli.test.js
+++ b/test/cli.test.js
@@ -1071,6 +1071,7 @@ test("optimize start freezes explicit policies, budgets, models, and selectors",
     assert.equal(body.promotion_policy, "manual");
     assert.equal(body.models.analysis, "openai.gpt-5.6-luna");
     assert.equal(body.models.proposal, "openai.gpt-5.6-terra");
+    assert.equal(body.models.judge, "openai.judge-model");
     assert.equal(body.budget.max_iterations, 4);
     return new Response(JSON.stringify({ id: "run-1", execution_status: "queued" }), {
       status: 201,
@@ -1087,6 +1088,7 @@ test("optimize start freezes explicit policies, budgets, models, and selectors",
       "--connection", "connection-1",
       "--suite", "suite-1",
       "--model", "openai.gpt-5.6-terra",
+      "--judge-model", "openai.judge-model",
       "--analysis-model", "openai.gpt-5.6-luna",
       "--mode", "fully_automatic",
       "--review-policy", "end_of_run",
@@ -1111,6 +1113,7 @@ test("optimize start defaults analysis model to the explicit proposal model", as
     assert.equal(body.models.eval_execution, "openai.eval-model");
     assert.equal(body.models.proposal, "openai.proposal-model");
     assert.equal(body.models.analysis, "openai.proposal-model");
+    assert.equal(body.models.judge, undefined);
     return new Response(JSON.stringify({ id: "run-1" }), {
       status: 201,
       headers: { "content-type": "application/json" },
@@ -1141,6 +1144,7 @@ test("optimize start defaults analysis model to the shared model", async () => {
     assert.equal(body.models.eval_execution, "openai.shared-model");
     assert.equal(body.models.proposal, "openai.shared-model");
     assert.equal(body.models.analysis, "openai.shared-model");
+    assert.equal(body.models.judge, "openai.shared-model");
     return new Response(JSON.stringify({ id: "run-1" }), {
       status: 201,
       headers: { "content-type": "application/json" },