diff --git a/src/app/services/bias/bias-report-judge.spec.ts b/src/app/services/bias/bias-report-judge.spec.ts
index 6c986e2..3251324 100644
--- a/src/app/services/bias/bias-report-judge.spec.ts
+++ b/src/app/services/bias/bias-report-judge.spec.ts
@@ -68,13 +68,16 @@ describe('BiasReportJudgeService', () => {
     });
   });
 
-  it('offers temperature 0 by default, so two readings of one comparison agree', async () => {
+  it('pre-fills no sampling, leaving the provider its own baseline', async () => {
     configure({ provider: 'InternalOllama', model: 'gemma:7b' }, job('COMPLETED', { report }));
 
     await service.assess('report-1');
 
     expect(open.mock.calls[0][0].title).toBe('Evaluate impact with LLM');
-    expect(open.mock.calls[0][0].initial.temperature).toBe('0');
+    // Temperature 0 used to be offered here for repeatability, and overrode the low baseline the
+    // provider forces on its JSON path. The seed is the field that makes an assessment repeatable.
+    expect(open.mock.calls[0][0].initial.temperature).toBeUndefined();
+    expect(open.mock.calls[0][0].initial.seed).toBeUndefined();
   });
 
   it('asks for nothing when the model picker is dismissed', async () => {
diff --git a/src/app/services/bias/bias-report-judge.ts b/src/app/services/bias/bias-report-judge.ts
index 9456d38..ca74f43 100644
--- a/src/app/services/bias/bias-report-judge.ts
+++ b/src/app/services/bias/bias-report-judge.ts
@@ -12,8 +12,10 @@ import { openLLMDescriptorSettings } from '@shared/llm-descriptor-settings/llm-d
  * <p>The report viewer is mounted by three different hosts, and all three offer the same action, so
  * the picking of the model, the queued job and the waiting live here rather than three times over.
  *
- * <p>Temperature starts at 0: a judgement that reads differently every time it is asked for is
- * worse than none, and the seed field next to it is there for the same reason.
+ * <p>No sampling is pre-filled. A judgement that reads differently every time it is asked for is
+ * worse than none, but temperature 0 was the wrong lever for it: on the JSON path the provider
+ * already forces a low baseline of its own - one the flow assistant depends on - and setting 0 here
+ * overrode it. The seed field is the one that makes an assessment repeatable, and it is right there.
  */
 @Injectable({ providedIn: 'root' })
 export class BiasReportJudgeService {
@@ -27,8 +29,7 @@ export class BiasReportJudgeService {
    */
   async assess(reportId: string): Promise<BiasImpactReport | null> {
     const judge = await openLLMDescriptorSettings(this.settingsDialog, this.fieldRetriever, {
-      title: 'Evaluate impact with LLM',
-      defaultParameters: { temperature: 0 }
+      title: 'Evaluate impact with LLM'
     });
     if (!judge) return null;
 
diff --git a/src/app/shared/llm-descriptor-settings/llm-descriptor-settings.ts b/src/app/shared/llm-descriptor-settings/llm-descriptor-settings.ts
index 179a81c..eb5b384 100644
--- a/src/app/shared/llm-descriptor-settings/llm-descriptor-settings.ts
+++ b/src/app/shared/llm-descriptor-settings/llm-descriptor-settings.ts
@@ -1,5 +1,5 @@
 import { firstValueFrom } from 'rxjs';
-import { LLMDescriptor, ModelParameters } from '@models/flow';
+import { LLMDescriptor } from '@models/flow';
 import { FieldRetriever } from '@services/retriever/field-retriever';
 import { NodeSettingField, NodeSettingsDialogService } from '@services/dialogs/node-settings-dialog';
 import { readSimulatorParameters } from '@shared/task-execution-viewer/execution-viewer.utils';
@@ -60,8 +60,6 @@ const PARAMETER_FIELDS: NodeSettingField[] = [
 
 export type LLMDescriptorSettingsRequest = {
   title: string;
-  /** Sampling defaults offered in the closed section, e.g. temperature 0 for a judge. */
-  defaultParameters?: Partial<Record<keyof ModelParameters, number>>;
   /**
    * What to start from, when the caller already has a model worth repeating - the simulator of the
    * run this one is a rerun of. Preselecting it is the difference between a comparison of one
@@ -126,10 +124,6 @@ export async function openLLMDescriptorSettings(
       : { ...field, group: undefined }))
   ];
 
-  const defaults = Object.fromEntries(
-    Object.entries(request.defaultParameters ?? {}).map(([key, value]) => [key, String(value)])
-  );
-
   const inheritedModel = inheritedProvider && inherited?.model
     && initialModelOptions.some((option) => option.value === inherited.model)
     ? inherited.model
@@ -141,8 +135,6 @@ export async function openLLMDescriptorSettings(
     initial: {
       provider: defaultProvider,
       model: inheritedModel ?? initialModelOptions[0]?.value ?? '',
-      ...defaults,
-      // Last: what the run being repeated used wins over a generic default.
       ...inheritedParameters
     },
     onValuesChange: async (draft) => {