@@ -16,7 +16,8 @@ image paths inside source records resolve relative to their JSONL file.
 | `source_directory` | `null` | Directory preparation settings below. Mutually exclusive with `sources_file`. |
 | `output_dir` | Required | New run directory, or an existing compatible directory when resuming. |
 | `dataset_id` | Required | Nonempty corpus identifier recorded in exports. |
-| `generator`, `judge` | Required | Independent model settings, described below. |
+| `generator`, `judge` | Required | Generator settings and default judge settings, described below. |
+| `judges` | `{}` | Optional complete model settings for individual judgment stages. |
 | `contexts_file` | `null` | Explicit context JSONL; replaces automatic context grouping. |
 | `concurrency` | `8` | Model-call concurrency, from 1 to 128. |
 | `batch_size` | `30` | Request batch size, from 1 to 128. |
@@ -27,7 +28,7 @@ image paths inside source records resolve relative to their JSONL file.
 | `ratios` | `{train: 0.8, validation: 0.0, evaluation: 0.2}` | Query-group split fractions, each from 0 to 1 and summing to 1. |
 | `group_near_duplicates` | `false` | Add conservative lexical query grouping before splitting. |
 
-The plugin groups requests by model role, response schema, and presence of images.
+The plugin groups requests by resolved model settings, response schema, and presence of images.
 Each batch runs a Data Designer structured-response column. Matching requests
 reuse cached responses keyed by their prompt, schema, model settings, and image
 content.
@@ -276,14 +277,91 @@ reasoning models, `max_tokens` can include both reasoning and the final response
 
 To enable local token preflight for a role, set `context_window_tokens`,
 `tokenizer_file`, and `request_overhead_tokens`. Also set
-`image_tokens_per_image` when the corpus contains images. Use values appropriate
+`image_tokens_per_image` for the generator, relevance judge, and localization
+judge when the corpus contains images. Text-only judgment stages do not require
+an image allowance. Use values appropriate
 to the deployment. The check includes the structured-response instructions,
 source input, image allowances, and output reserve.
 
 Without local token preflight, Data Designer's rendered-prompt limit still
 applies, but it is not a model context-window guarantee. Provider context and
 image limits remain applicable.
 
+### Judge Model Overrides
+
+Use `judge` for the default judgment model. Override individual stages under
+`judges` when a different deployment or sampling configuration suits the task:
+
+| Stage | Input | Configuration key |
+| --- | --- | --- |
+| Summary quality | Generated summary and persona | `judges.summary` |
+| Query classification and answer leakage | Query only | `judges.query_metadata` |
+| Self-sufficiency | Query and source text | `judges.self_sufficiency` |
+| Relevance | Query, source text, and available images | `judges.relevance` |
+| Support localization | Query, source text, and available images | `judges.localization` |
+
+An omitted or `null` stage uses `judge`. Each override is a complete
+[Model Settings](#model-settings) object. Required fields must be supplied;
+optional fields use their own defaults rather than inheriting values from
+`judge`. This includes sampling parameters, output limits, credentials, and
+local token budgets. Configure vision-capable deployments for image-bearing
+requests; the plugin does not infer model capabilities or switch models on errors.
+
+This configuration uses one text judge for summary and query checks, and the
+default vision judge for relevance and support localization. Replace the model
+identifiers and endpoints with your serving deployments, and set both credential
+environment variables before running:
+
+```yaml
+sources_file: sources.jsonl
+output_dir: runs/stage-judges
+dataset_id: service-manuals
+generator:
+  model: your-generation-model
+  endpoint: https://generation.example.com/v1
+  credential_env: GENERATION_API_KEY
+judge:
+  model: your-vision-judge-model
+  endpoint: https://vision.example.com/v1
+  credential_env: JUDGE_API_KEY
+  temperature: 0
+  max_tokens: 4096
+judges:
+  summary: &text_judge
+    model: your-text-judge-model
+    endpoint: https://text.example.com/v1
+    credential_env: JUDGE_API_KEY
+    temperature: 0
+    max_tokens: 1024
+  query_metadata: *text_judge
+  self_sufficiency: *text_judge
+```
+
+You can also configure `judges.relevance` and `judges.localization` independently.
+For Python configuration, import `JudgeModels` from
+`data_designer_retrieval_sdg.multimodal`. `config.model_for("judge", "relevance")`
+returns the effective settings for that stage.
+
+Automatic section planning (`context_strategy: sections` without `contexts_file`)
+uses the fallback `judge` for text-only table-of-contents confirmation. This judge
+remains active even when all five named stages are overridden or
+`judge_summaries: false`. Its local token budget is validated before creating the
+run, and its tokenizer content hash is included in the frozen run identity.
+Explicit contexts bypass this confirmation step.
+
+`judge_summaries: false` disables the summary judgment stage, including its
+model-budget validation. The run manifest records the fallback under
+`models.judge` and the effective models for enabled stages under `models.judges`.
+`shared_generator_judge` is true only when every active judge, including the
+fallback used for TOC confirmation, uses the generator's model identifier and endpoint.
+
+Cached responses include the resolved model settings and tokenizer content hash.
+Changing one override cannot reuse that stage's incompatible response. Changing
+settings in an existing run still violates its frozen identity; use a new output
+directory. Omitting overrides preserves the original judge request cache keys.
+Stage configuration does not establish better judgment quality; compare reviewed
+examples when choosing models.
+
 ## Context Planning
 
 A context specifies the units available for one generation request. These fields
 
@@ -16,7 +16,8 @@ image paths inside source records resolve relative to their JSONL file.
 | `source_directory` | `null` | Directory preparation settings below. Mutually exclusive with `sources_file`. |
 | `output_dir` | Required | New run directory, or an existing compatible directory when resuming. |
 | `dataset_id` | Required | Nonempty corpus identifier recorded in exports. |
-| `generator`, `judge` | Required | Independent model settings, described below. |
+| `generator`, `judge` | Required | Generator settings and default judge settings, described below. |
+| `judges` | `{}` | Optional complete model settings for individual judgment stages. |
 | `contexts_file` | `null` | Explicit context JSONL; replaces automatic context grouping. |
 | `concurrency` | `8` | Model-call concurrency, from 1 to 128. |
 | `batch_size` | `30` | Request batch size, from 1 to 128. |
@@ -27,7 +28,7 @@ image paths inside source records resolve relative to their JSONL file.
 | `ratios` | `{train: 0.8, validation: 0.0, evaluation: 0.2}` | Query-group split fractions, each from 0 to 1 and summing to 1. |
 | `group_near_duplicates` | `false` | Add conservative lexical query grouping before splitting. |
 
-The plugin groups requests by model role, response schema, and presence of images.
+The plugin groups requests by resolved model settings, response schema, and presence of images.
 Each batch runs a Data Designer structured-response column. Matching requests
 reuse cached responses keyed by their prompt, schema, model settings, and image
 content.
@@ -276,14 +277,91 @@ reasoning models, `max_tokens` can include both reasoning and the final response
 
 To enable local token preflight for a role, set `context_window_tokens`,
 `tokenizer_file`, and `request_overhead_tokens`. Also set
-`image_tokens_per_image` when the corpus contains images. Use values appropriate
+`image_tokens_per_image` for the generator, relevance judge, and localization
+judge when the corpus contains images. Text-only judgment stages do not require
+an image allowance. Use values appropriate
 to the deployment. The check includes the structured-response instructions,
 source input, image allowances, and output reserve.
 
 Without local token preflight, Data Designer's rendered-prompt limit still
 applies, but it is not a model context-window guarantee. Provider context and
 image limits remain applicable.
 
+### Judge Model Overrides
+
+Use `judge` for the default judgment model. Override individual stages under
+`judges` when a different deployment or sampling configuration suits the task:
+
+| Stage | Input | Configuration key |
+| --- | --- | --- |
+| Summary quality | Generated summary and persona | `judges.summary` |
+| Query classification and answer leakage | Query only | `judges.query_metadata` |
+| Self-sufficiency | Query and source text | `judges.self_sufficiency` |
+| Relevance | Query, source text, and available images | `judges.relevance` |
+| Support localization | Query, source text, and available images | `judges.localization` |
+
+An omitted or `null` stage uses `judge`. Each override is a complete
+[Model Settings](#model-settings) object. Required fields must be supplied;
+optional fields use their own defaults rather than inheriting values from
+`judge`. This includes sampling parameters, output limits, credentials, and
+local token budgets. Configure vision-capable deployments for image-bearing
+requests; the plugin does not infer model capabilities or switch models on errors.
+
+This configuration uses one text judge for summary and query checks, and the
+default vision judge for relevance and support localization. Replace the model
+identifiers and endpoints with your serving deployments, and set both credential
+environment variables before running:
+
+```yaml
+sources_file: sources.jsonl
+output_dir: runs/stage-judges
+dataset_id: service-manuals
+generator:
+  model: your-generation-model
+  endpoint: https://generation.example.com/v1
+  credential_env: GENERATION_API_KEY
+judge:
+  model: your-vision-judge-model
+  endpoint: https://vision.example.com/v1
+  credential_env: JUDGE_API_KEY
+  temperature: 0
+  max_tokens: 4096
+judges:
+  summary: &text_judge
+    model: your-text-judge-model
+    endpoint: https://text.example.com/v1
+    credential_env: JUDGE_API_KEY
+    temperature: 0
+    max_tokens: 1024
+  query_metadata: *text_judge
+  self_sufficiency: *text_judge
+```
+
+You can also configure `judges.relevance` and `judges.localization` independently.
+For Python configuration, import `JudgeModels` from
+`data_designer_retrieval_sdg.multimodal`. `config.model_for("judge", "relevance")`
+returns the effective settings for that stage.
+
+Automatic section planning (`context_strategy: sections` without `contexts_file`)
+uses the fallback `judge` for text-only table-of-contents confirmation. This judge
+remains active even when all five named stages are overridden or
+`judge_summaries: false`. Its local token budget is validated before creating the
+run, and its tokenizer content hash is included in the frozen run identity.
+Explicit contexts bypass this confirmation step.
+
+`judge_summaries: false` disables the summary judgment stage, including its
+model-budget validation. The run manifest records the fallback under
+`models.judge` and the effective models for enabled stages under `models.judges`.
+`shared_generator_judge` is true only when every active judge, including the
+fallback used for TOC confirmation, uses the generator's model identifier and endpoint.
+
+Cached responses include the resolved model settings and tokenizer content hash.
+Changing one override cannot reuse that stage's incompatible response. Changing
+settings in an existing run still violates its frozen identity; use a new output
+directory. Omitting overrides preserves the original judge request cache keys.
+Stage configuration does not establish better judgment quality; compare reviewed
+examples when choosing models.
+
 ## Context Planning
 
 A context specifies the units available for one generation request. These fields
 
@@ -7,6 +7,7 @@
 
 from data_designer_retrieval_sdg.multimodal.models import (
     GenerationContext,
+    JudgeModels,
     ModelSettings,
     MultimodalSDGConfig,
     QueryInstruction,
@@ -18,6 +19,7 @@
 
 __all__ = [
     "GenerationContext",
+    "JudgeModels",
     "ModelSettings",
     "MultimodalSDGConfig",
     "QueryInstruction",
 
@@ -40,12 +40,11 @@ def validate_model_budget(model) -> None:
 
 def validate_deployment(config, has_images: bool) -> None:
     """Validate model budgets/assets before creating an immutable run directory."""
-    for role in ("generator", "judge"):
-        model = getattr(config, role)
+    for role, model in {"generator": config.generator, **config.active_judges()}.items():
         if model.context_window_tokens is None:
             continue
         validate_model_budget(model)
-        if has_images and model.image_tokens_per_image is None:
+        if has_images and role in ("generator", "relevance", "localization") and model.image_tokens_per_image is None:
             raise ValueError(f"Configure image_tokens_per_image for {role} before running image SDG")
         tokenizer(str(model.tokenizer_file), digest(model.tokenizer_file))
 
@@ -63,7 +62,7 @@ def check_request(request, config) -> None:
     """
     if len(request.text) > environment.MAX_RENDERED_LEN:
         raise RequestTooLarge(f"Rendered prompt exceeds DD limit ({environment.MAX_RENDERED_LEN} characters)")
-    model = getattr(config, request.role)
+    model = config.model_for(request.role, request.judge_stage)
     if model.context_window_tokens is not None:
         validate_model_budget(model)
         if request.images and model.image_tokens_per_image is None:
@@ -80,7 +79,7 @@ def check_request(request, config) -> None:
         )
         if required > model.context_window_tokens:
             raise RequestTooLarge(
-                f"{request.role} request needs {required} tokens including images, overhead and output; "
+                f"{request.judge_stage or request.role} request needs {required} tokens including images, overhead and output; "
                 f"configured context window is {model.context_window_tokens}"
             )
 
 
@@ -346,6 +346,7 @@ def write_multimodal_bundle(
                     exclude={
                         "generator",
                         "judge",
+                        "judges",
                         "sources_file",
                         "source_directory",
                         "contexts_file",
@@ -357,9 +358,12 @@ def write_multimodal_bundle(
             "models": {
                 "generator": config.generator.model_dump(mode="json"),
                 "judge": config.judge.model_dump(mode="json"),
+                "judges": {stage: model.model_dump(mode="json") for stage, model in config.resolved_judges().items()},
             },
-            "shared_generator_judge": config.generator.model == config.judge.model
-            and config.generator.endpoint == config.judge.endpoint,
+            "shared_generator_judge": all(
+                config.generator.model == model.model and config.generator.endpoint == model.endpoint
+                for model in config.active_judges().values()
+            ),
             "dataset_id": config.dataset_id,
             "split_protocol": "grouped_query_disjoint",
             "corpus_scope": "full_collection",