diff --git a/README.md b/README.md
index 4dcd880..0adc338 100644
--- a/README.md
+++ b/README.md
@@ -106,7 +106,6 @@ The canonical schema is organized by domain under
 | Area                  | What the contract provides                                                                                        |
 | --------------------- | ----------------------------------------------------------------------------------------------------------------- |
 | Portable generation   | Text or token input, sampling, stopping, transport priorities, multiple sequences, and deterministic seeds        |
-| Non-generative tasks  | Typed embedding, classification, and grouped query/candidate scoring with stable correlation                      |
 | Structured output     | JSON Schema, JSON object, regex, EBNF grammar, structural tags, and fixed choices                                 |
 | Token information     | Prompt and output logprobs, ranks, candidate-token selection, per-token records, and streamed text deltas         |
 | Discovery             | Server identity, deployment capacity, model limits, topology, parsers, and inference capabilities                 |
diff --git a/docs/api.md b/docs/api.md
index 807c068..2226237 100644
--- a/docs/api.md
+++ b/docs/api.md
@@ -23,11 +23,6 @@ import "google/protobuf/struct.proto";
 service Inference {
   // Core inference path.
   rpc Generate(GenerateRequest) returns (stream GenerateResponse);
-
-  // Non-generative inference paths.
-  rpc Embed(EmbedRequest) returns (EmbedResponse);
-  rpc Classify(ClassifyRequest) returns (ClassifyResponse);
-  rpc Score(ScoreRequest) returns (ScoreResponse);
 }
 
 service Control {
@@ -46,13 +41,9 @@ service Control {
   rpc UnloadLora(UnloadLoraRequest) returns (UnloadLoraResponse);
   rpc ListLoras(ListLorasRequest) returns (ListLorasResponse);
 
-  // Disaggregated serving / KV transfer.
-  rpc GetKvConnectorInfo(GetKvConnectorInfoRequest) returns (KvConnectorInfo);
+  // Disaggregated serving / KV transfer. Connector info: ServerInfo.kv_connector.
   rpc GetKvEventSources(GetKvEventSourcesRequest) returns (GetKvEventSourcesResponse);
   rpc SubscribeKvEvents(SubscribeKvEventsRequest) returns (stream SubscribeKvEventsResponse);
-
-  // Structured runtime events for planners/controllers.
-  rpc SubscribeRuntimeEvents(SubscribeRuntimeEventsRequest) returns (stream SubscribeRuntimeEventsResponse);
 }
 ```
 
@@ -212,7 +203,6 @@ message ModelInfo {
 
   string reasoning_parser = 25;
   string tool_call_parser = 26;
-  TaskCapabilities tasks = 27; // Optional non-generative task support for this model.
   google.protobuf.Struct extra = 28; // Engine-specific, non-portable; read opportunistically.
 }
 
@@ -273,351 +263,6 @@ the LoRA lifecycle RPCs on `Control`.
 
 ---
 
-## Non-generative task API
-
-`Embed`, `Classify`, and `Score` are unary inference operations. They share a
-request context and typed inputs, but retain task-specific options and outputs.
-Support is optional per model and advertised through `ModelInfo.tasks`.
-
-```protobuf
-message TaskRequestContext {
-  string request_id = 1;
-  string model = 2;
-  string lora_name = 3;
-  google.protobuf.Struct extra = 4; // Engine-specific, non-portable; may be ignored.
-}
-
-message TaskInput {
-  string item_id = 1;
-  oneof input {
-    string text = 2;
-    TokenIds token_ids = 3;
-    MultimodalTaskInput multimodal = 4;
-  }
-}
-
-message MultimodalTaskInput {
-  oneof prompt {
-    string text = 1;
-    TokenIds token_ids = 2;
-  }
-  repeated MediaItem media = 3;
-}
-
-message DenseFloatTensor {
-  repeated uint64 shape = 1;
-  repeated float values = 2;
-}
-
-message SparseFloatTensor {
-  repeated uint64 shape = 1;
-  repeated uint64 indices = 2;
-  repeated float values = 3;
-}
-
-message TaskUsage {
-  uint64 input_tokens = 1;
-  optional uint64 cached_input_tokens = 2;
-}
-
-enum TaskInputType {
-  TASK_INPUT_TYPE_UNSPECIFIED = 0;
-  TASK_INPUT_TYPE_TEXT = 1;
-  TASK_INPUT_TYPE_TOKEN_IDS = 2;
-  TASK_INPUT_TYPE_MULTIMODAL = 3;
-}
-
-enum TaskOutputGranularity {
-  TASK_OUTPUT_GRANULARITY_UNSPECIFIED = 0;
-  TASK_OUTPUT_GRANULARITY_SEQUENCE = 1;
-  TASK_OUTPUT_GRANULARITY_TOKEN = 2;
-}
-
-enum EmbeddingEncoding {
-  EMBEDDING_ENCODING_UNSPECIFIED = 0;
-  EMBEDDING_ENCODING_DENSE = 1;
-  EMBEDDING_ENCODING_SPARSE = 2;
-}
-
-enum TaskValueSemantics {
-  TASK_VALUE_SEMANTICS_UNSPECIFIED = 0;
-  TASK_VALUE_SEMANTICS_LOGITS = 1;
-  TASK_VALUE_SEMANTICS_PROBABILITIES = 2;
-  TASK_VALUE_SEMANTICS_LOG_PROBABILITIES = 3;
-  TASK_VALUE_SEMANTICS_SIMILARITY = 4;
-  TASK_VALUE_SEMANTICS_RELEVANCE = 5;
-  TASK_VALUE_SEMANTICS_REWARD = 6;
-  TASK_VALUE_SEMANTICS_MODEL_DEFINED = 7;
-}
-
-enum ScoreNormalization {
-  SCORE_NORMALIZATION_UNSPECIFIED = 0;
-  SCORE_NORMALIZATION_NONE = 1;
-  SCORE_NORMALIZATION_SOFTMAX = 2;
-}
-```
-
-`TaskRequestContext.request_id` and `model` are required and non-empty. Request
-IDs share the same namespace and abort semantics as generation request IDs.
-A non-empty `lora_name` selects an already loaded adapter. Clients use
-`openengine-priority` only when the corresponding task capability advertises
-priority support, and use `lora_name` only when it advertises LoRA support.
-
-Each request batch must be non-empty. `item_id` is required and unique within
-an embed/classify batch, and every query/candidate item ID is unique within one
-score group. Exactly one `TaskInput.input` variant is set. A multimodal input must
-contain at least one `MediaItem`; its optional prompt is either text or token
-IDs, never both. Media ordering and validation follow `GenerateRequest.media`.
-
-`DenseFloatTensor` is row-major FP32 data. Every dimension is greater than zero,
-the product of `shape` equals `values.size()`, and every value is finite.
-`SparseFloatTensor` uses flattened row-major indices; `indices` and `values`
-have equal length, indices are unique and strictly increasing, and every index
-is smaller than the product of `shape`. A scalar is encoded with shape `[1]`,
-not an empty shape. `cached_input_tokens`, when present, does not exceed
-`input_tokens`.
-
-### Embedding
-
-```protobuf
-message EmbedRequest {
-  TaskRequestContext context = 1;
-  repeated TaskInput inputs = 2;
-  EmbedOptions options = 3;
-}
-
-message EmbedOptions {
-  optional TaskOutputGranularity granularity = 1;
-  optional bool normalize = 2;
-  optional uint32 dimensions = 3;
-  optional EmbeddingEncoding encoding = 4;
-}
-
-message EmbedResponse {
-  string request_id = 1;
-  repeated EmbeddingOutput outputs = 2;
-  TaskUsage usage = 3;
-}
-
-message EmbeddingOutput {
-  string item_id = 1;
-  optional uint32 input_index = 2;
-  TaskOutputGranularity granularity = 3;
-  oneof embedding {
-    DenseFloatTensor dense = 4;
-    SparseFloatTensor sparse = 5;
-  }
-  repeated uint32 token_ids = 6;
-}
-```
-
-Absent embedding options select model defaults. An explicit granularity,
-normalization, or encoding request must be implemented exactly or rejected.
-`dimensions` must be greater than zero and requests dimensionality reduction;
-it does not permit padding a smaller model output.
-
-Sequence embeddings have shape `[dimension]`. Token embeddings have shape
-`[token_count, dimension]`; `token_ids`, when returned, has `token_count`
-entries aligned to the first tensor dimension. Outputs preserve request order,
-and `input_index` is present even for index zero and identifies the matching
-input. A successful response contains exactly one output for every request
-input.
-
-### Classification
-
-```protobuf
-message ClassifyRequest {
-  TaskRequestContext context = 1;
-  repeated TaskInput inputs = 2;
-  ClassifyOptions options = 3;
-}
-
-message ClassifyOptions {
-  optional TaskOutputGranularity granularity = 1;
-  optional TaskValueSemantics output_semantics = 2;
-}
-
-message ClassifyResponse {
-  string request_id = 1;
-  repeated ClassificationOutput outputs = 2;
-  TaskUsage usage = 3;
-}
-
-message ClassificationOutput {
-  string item_id = 1;
-  optional uint32 input_index = 2;
-  TaskOutputGranularity granularity = 3;
-  TaskValueSemantics semantics = 4;
-  DenseFloatTensor scores = 5;
-  repeated string labels = 6;
-  repeated uint32 token_ids = 7;
-}
-```
-
-Classification normally returns logits, probabilities, log probabilities, or
-model-defined values. The engine reports the actual non-`UNSPECIFIED`
-semantics. Sequence classification has shape `[class_count]`; token
-classification has shape `[token_count, class_count]`. When labels are
-returned, their count equals `class_count` and their order matches the final
-tensor dimension. Token IDs follow the same alignment rule as token
-embeddings. Outputs preserve request order and correlation. A successful
-response contains exactly one output for every request input, and `input_index`
-is present even for index zero.
-
-### Scoring
-
-```protobuf
-message ScoreRequest {
-  TaskRequestContext context = 1;
-  repeated ScoreGroup groups = 2;
-  ScoreOptions options = 3;
-}
-
-message ScoreGroup {
-  string group_id = 1;
-  TaskInput query = 2;
-  repeated TaskInput candidates = 3;
-}
-
-message ScoreOptions {
-  optional TaskOutputGranularity granularity = 1;
-  optional TaskValueSemantics output_semantics = 2;
-  ScoreNormalization normalization = 3;
-  repeated uint32 label_token_ids = 4;
-  string instruction = 5;
-}
-
-message ScoreResponse {
-  string request_id = 1;
-  repeated ScoreGroupOutput groups = 2;
-  TaskOutputGranularity granularity = 3;
-  TaskValueSemantics semantics = 4;
-  optional bool higher_is_better = 5;
-  ScoreNormalization normalization = 6;
-  repeated uint32 label_token_ids = 7;
-  TaskUsage usage = 8;
-}
-
-message ScoreGroupOutput {
-  string group_id = 1;
-  optional uint32 group_index = 2;
-  repeated ScoreCandidateOutput candidates = 3;
-}
-
-message ScoreCandidateOutput {
-  string candidate_id = 1;
-  optional uint32 candidate_index = 2;
-  DenseFloatTensor scores = 3;
-  repeated uint32 token_ids = 4;
-}
-```
-
-Every score request has at least one group. Group IDs are unique, every group
-has one query and at least one candidate, and candidate IDs are unique within
-the group. Repeating groups represents N:N paired scoring; one group with many
-candidates represents the optimized 1:N path used by rerankers and multi-item
-scoring engines.
-
-Absent granularity and output semantics select model defaults. An explicit
-value must be supported. `SCORE_NORMALIZATION_UNSPECIFIED` selects the model
-default, `NONE` requests native unnormalized values, and `SOFTMAX` requests
-normalization across each returned label vector. Duplicate label token IDs are
-invalid. When `label_token_ids` is non-empty, the engine performs causal-model
-label-token scoring and the response repeats those IDs in the same order.
-When it is empty, the engine performs its advertised model-native scoring
-operation. A non-empty `instruction` is valid only when instruction support is
-advertised; template selection and rendering remain engine-owned.
-
-`ScoreResponse` reports the actual granularity, semantics, normalization, and
-label-token order. Group and candidate results preserve request order and carry
-their present zero-based original indexes, including index zero. A successful
-response contains every group and candidate exactly once; partial success is
-not represented. Token-granularity candidate tensors may return aligned token
-IDs. The engine never sorts candidates or echoes source documents.
-
-A gateway may derive reranking only when every candidate has exactly one score
-and `higher_is_better` is present. It stable-sorts using that direction, breaks
-ties by `candidate_index`, applies its external `top_n`, and attaches documents
-from gateway-owned request state. Reranking is response shaping, not a separate
-OpenEngine inference capability.
-
-### Task capability discovery
-
-```protobuf
-message TaskCapabilities {
-  EmbedCapabilities embed = 1;
-  ClassifyCapabilities classify = 2;
-  ScoreCapabilities score = 3;
-}
-
-message EmbedCapabilities {
-  optional bool supported = 1;
-  repeated TaskInputType input_types = 2;
-  repeated TaskOutputGranularity granularities = 3;
-  repeated EmbeddingEncoding encodings = 4;
-  optional uint32 dimension = 5;
-  optional uint32 max_batch_size = 6;
-  optional uint64 max_output_values_per_item = 7;
-  optional bool supports_priority = 8;
-  optional bool supports_lora = 9;
-  optional bool supports_normalization = 10;
-  optional bool supports_dimension_override = 11;
-  repeated Modality modalities = 12;
-}
-
-message ClassifyCapabilities {
-  optional bool supported = 1;
-  repeated TaskInputType input_types = 2;
-  repeated TaskOutputGranularity granularities = 3;
-  repeated TaskValueSemantics semantics = 4;
-  optional uint32 max_batch_size = 5;
-  optional uint64 max_output_values_per_item = 6;
-  optional bool supports_priority = 7;
-  optional bool supports_lora = 8;
-  repeated Modality modalities = 9;
-}
-
-message ScoreCapabilities {
-  optional bool supported = 1;
-  repeated TaskInputType input_types = 2;
-  repeated TaskOutputGranularity granularities = 3;
-  repeated TaskValueSemantics semantics = 4;
-  repeated ScoreNormalization normalizations = 5;
-  optional bool supports_label_token_scoring = 6;
-  optional bool supports_instruction = 7;
-  optional uint32 max_groups = 8;
-  optional uint32 max_candidates_per_group = 9;
-  optional uint64 max_output_values_per_candidate = 10;
-  optional bool supports_priority = 11;
-  optional bool supports_lora = 12;
-  optional bool higher_is_better = 13;
-  repeated Modality modalities = 14;
-  optional uint32 max_label_token_ids = 15;
-}
-```
-
-An absent `ModelInfo.tasks` means the model does not advertise non-generative
-task support. Within it, an absent task capability is unreported; a present
-capability uses `supported` presence to distinguish unreported support from an
-explicit `true` or `false`. Capability lists never contain `UNSPECIFIED`.
-Reported dimensions and limits are greater than zero. `dimension` is present
-only when a model has one fixed native embedding dimension. A rankable native
-score advertises its default direction through `higher_is_better`; the response
-still reports the actual direction for each request. `modalities` is meaningful
-only when the task includes `TASK_INPUT_TYPE_MULTIMODAL`, and never contains
-`MODALITY_UNSPECIFIED`.
-For dense embeddings, `max_output_values_per_item` limits the flattened element
-count; for sparse embeddings, it limits the number of returned nonzero values.
-
-The server rejects an unsupported task with `FAILED_PRECONDITION`, an unknown
-model with `NOT_FOUND`, malformed input or unsupported explicit options with
-`INVALID_ARGUMENT`, and admission/capacity exhaustion with
-`RESOURCE_EXHAUSTED`. A successful unary response is terminal. Client
-cancellation or deadline expiration stops queued or running work, and
-`Abort(request_id)` follows the same semantics as generation.
-
----
-
 ## LoRA lifecycle
 
 ```protobuf
@@ -972,9 +617,11 @@ OpenEngine should support two KV-event modes:
 2. **Compatibility source discovery:** `GetKvEventSources` advertises existing
    engine-native sources such as SGLang/vLLM ZMQ publishers.
 
-```protobuf
-message GetKvConnectorInfoRequest {}
+`KvConnectorInfo` describes the disaggregation transfer connector. It is static
+per deployment and is reported once through `ServerInfo.kv_connector`
+(`GetServerInfo`), not a dedicated RPC.
 
+```protobuf
 message KvConnectorInfo {
   optional bool enabled = 1;
   string transfer_backend = 2;
@@ -1193,86 +840,6 @@ sessions remain.
 
 ---
 
-## Runtime observability
-
-`GetLoad` returns a structured point-in-time load snapshot for schedulers and admission controllers. It is not a replacement for Prometheus metrics; it is the engine-facing control-plane signal for request routing and overload decisions.
-
-```protobuf
-message GetLoadRequest {
-  bool include_per_rank = 1;
-}
-
-message LoadInfo {
-  string instance_id = 1;
-  optional uint64 timestamp_unix_nanos = 2;
-  optional uint32 running_requests = 3;
-  optional uint32 queued_requests = 4;
-  optional uint32 active_kv_sessions = 5;
-  optional uint64 used_kv_blocks = 6;
-  optional uint64 total_kv_blocks = 7;
-  optional uint64 running_tokens = 8;
-  optional uint64 waiting_tokens = 9;
-  optional uint32 prefill_batch_size = 10;
-  optional uint32 decode_batch_size = 11;
-  repeated RankLoadInfo ranks = 20;
-  google.protobuf.Struct attributes = 30;
-}
-
-message RankLoadInfo {
-  optional uint32 data_parallel_rank = 1;
-  optional uint32 running_requests = 2;
-  optional uint32 queued_requests = 3;
-  optional uint64 used_kv_blocks = 4;
-  optional uint64 total_kv_blocks = 5;
-  optional uint32 prefill_batch_size = 6;
-  optional uint32 decode_batch_size = 7;
-}
-```
-
-Every load scalar has explicit presence. Absent means unavailable in that
-engine or snapshot; present zero means the measured load is zero.
-
-`LoadInfo.attributes` and `RuntimeEvent.attributes` carry engine-specific metrics as a
-`google.protobuf.Struct`, so numeric, boolean, and list values keep their JSON
-type on the wire instead of being flattened to strings. `Struct` numbers are
-IEEE-754 doubles (exact only to 2^53); carry larger integers as strings.
-
-Runtime event stream:
-
-```protobuf
-message SubscribeRuntimeEventsRequest {
-  repeated RuntimeEventType types = 1;
-}
-
-message SubscribeRuntimeEventsResponse {
-  oneof event {
-    RuntimeEvent runtime_event = 1;
-    EngineError error = 2;
-  }
-}
-
-enum RuntimeEventType {
-  RUNTIME_EVENT_TYPE_UNSPECIFIED = 0;
-  RUNTIME_EVENT_TYPE_FORWARD_PASS = 1;
-  RUNTIME_EVENT_TYPE_BATCH = 2;
-  RUNTIME_EVENT_TYPE_QUEUE = 3;
-  RUNTIME_EVENT_TYPE_TRANSFER = 4;
-}
-
-message RuntimeEvent {
-  string event_id = 1;
-  uint64 timestamp_unix_nanos = 2;
-  RuntimeEventType type = 3;
-  google.protobuf.Struct attributes = 4;
-}
-```
-
-After subscription acceptance, an application failure is the final
-`SubscribeRuntimeEventsResponse` with `error` set. No runtime event may follow
-it, and the server closes the stream with gRPC `OK`.
-
----
-
 ## Standard errors
 
 ```protobuf
diff --git a/proto/openengine/v1/README.md b/proto/openengine/v1/README.md
index f9f8a0c..fcc090f 100644
--- a/proto/openengine/v1/README.md
+++ b/proto/openengine/v1/README.md
@@ -15,14 +15,10 @@ share the same package and together define the API.
 | [`server.proto`](server.proto) | Server identity, deployment capacity, engine roles, and parallelism |
 | [`model.proto`](model.proto) | Model metadata and inference capabilities |
 | [`generation.proto`](generation.proto) | Generation requests, parameters, streamed events, and usage |
-| [`tasks.proto`](tasks.proto) | Shared non-generative task request and output vocabulary |
-| [`embedding.proto`](embedding.proto) | Dense and sparse embedding inference |
-| [`classification.proto`](classification.proto) | Sequence and token classification inference |
-| [`scoring.proto`](scoring.proto) | Grouped query and candidate scoring inference |
 | [`lora.proto`](lora.proto) | LoRA adapter lifecycle |
 | [`kv.proto`](kv.proto) | KV sessions, connector discovery, and cache events |
 | [`lifecycle.proto`](lifecycle.proto) | Health, abort, and drain operations |
-| [`observability.proto`](observability.proto) | Load snapshots and runtime events |
+| [`observability.proto`](observability.proto) | Load snapshots |
 | [`error.proto`](error.proto) | Terminal errors for accepted streaming requests |
 
 Generate bindings from every `.proto` file in this directory. Compiling only
diff --git a/proto/openengine/v1/classification.proto b/proto/openengine/v1/classification.proto
deleted file mode 100644
index 20b65af..0000000
--- a/proto/openengine/v1/classification.proto
+++ /dev/null
@@ -1,50 +0,0 @@
-// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
-// SPDX-License-Identifier: Apache-2.0
-
-// Portable sequence and token classification inference.
-
-syntax = "proto3";
-
-package openengine.v1;
-
-import "openengine/v1/input.proto";
-import "openengine/v1/tasks.proto";
-
-message ClassifyCapabilities {
-  optional bool supported = 1;
-  repeated TaskInputType input_types = 2;
-  repeated TaskOutputGranularity granularities = 3;
-  repeated TaskValueSemantics semantics = 4;
-  optional uint32 max_batch_size = 5;
-  optional uint64 max_output_values_per_item = 6;
-  optional bool supports_priority = 7;
-  optional bool supports_lora = 8;
-  repeated Modality modalities = 9;
-}
-
-message ClassifyRequest {
-  TaskRequestContext context = 1;
-  repeated TaskInput inputs = 2;
-  ClassifyOptions options = 3;
-}
-
-message ClassifyOptions {
-  optional TaskOutputGranularity granularity = 1;
-  optional TaskValueSemantics output_semantics = 2;
-}
-
-message ClassifyResponse {
-  string request_id = 1;
-  repeated ClassificationOutput outputs = 2;
-  TaskUsage usage = 3;
-}
-
-message ClassificationOutput {
-  string item_id = 1;
-  optional uint32 input_index = 2;
-  TaskOutputGranularity granularity = 3;
-  TaskValueSemantics semantics = 4;
-  DenseFloatTensor scores = 5;
-  repeated string labels = 6;
-  repeated uint32 token_ids = 7;
-}
diff --git a/proto/openengine/v1/embedding.proto b/proto/openengine/v1/embedding.proto
deleted file mode 100644
index 70b7fe6..0000000
--- a/proto/openengine/v1/embedding.proto
+++ /dev/null
@@ -1,69 +0,0 @@
-// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
-// SPDX-License-Identifier: Apache-2.0
-
-// Portable dense and sparse embedding inference.
-
-syntax = "proto3";
-
-package openengine.v1;
-
-import "openengine/v1/input.proto";
-import "openengine/v1/tasks.proto";
-
-enum EmbeddingEncoding {
-  EMBEDDING_ENCODING_UNSPECIFIED = 0;
-  EMBEDDING_ENCODING_DENSE = 1;
-  EMBEDDING_ENCODING_SPARSE = 2;
-}
-
-// Row-major sparse FP32 tensor with flattened, strictly increasing indices.
-message SparseFloatTensor {
-  repeated uint64 shape = 1;
-  repeated uint64 indices = 2;
-  repeated float values = 3;
-}
-
-message EmbedCapabilities {
-  optional bool supported = 1;
-  repeated TaskInputType input_types = 2;
-  repeated TaskOutputGranularity granularities = 3;
-  repeated EmbeddingEncoding encodings = 4;
-  optional uint32 dimension = 5;
-  optional uint32 max_batch_size = 6;
-  optional uint64 max_output_values_per_item = 7;
-  optional bool supports_priority = 8;
-  optional bool supports_lora = 9;
-  optional bool supports_normalization = 10;
-  optional bool supports_dimension_override = 11;
-  repeated Modality modalities = 12;
-}
-
-message EmbedRequest {
-  TaskRequestContext context = 1;
-  repeated TaskInput inputs = 2;
-  EmbedOptions options = 3;
-}
-
-message EmbedOptions {
-  optional TaskOutputGranularity granularity = 1;
-  optional bool normalize = 2;
-  optional uint32 dimensions = 3;
-  optional EmbeddingEncoding encoding = 4;
-}
-
-message EmbedResponse {
-  string request_id = 1;
-  repeated EmbeddingOutput outputs = 2;
-  TaskUsage usage = 3;
-}
-
-message EmbeddingOutput {
-  string item_id = 1;
-  optional uint32 input_index = 2;
-  TaskOutputGranularity granularity = 3;
-  oneof embedding {
-    DenseFloatTensor dense = 4;
-    SparseFloatTensor sparse = 5;
-  }
-  repeated uint32 token_ids = 6;
-}
diff --git a/proto/openengine/v1/kv.proto b/proto/openengine/v1/kv.proto
index 95a2295..af180dd 100644
--- a/proto/openengine/v1/kv.proto
+++ b/proto/openengine/v1/kv.proto
@@ -29,8 +29,6 @@ message KvEndpoint {
   string protocol = 3; // grpc, nixl, ucx, tcp, shm, etc.
 }
 
-message GetKvConnectorInfoRequest {}
-
 message KvConnectorInfo {
   optional bool enabled = 1;
   string transfer_backend = 2;
diff --git a/proto/openengine/v1/model.proto b/proto/openengine/v1/model.proto
index de92175..5995a20 100644
--- a/proto/openengine/v1/model.proto
+++ b/proto/openengine/v1/model.proto
@@ -8,9 +8,6 @@ syntax = "proto3";
 package openengine.v1;
 
 import "google/protobuf/struct.proto";
-import "openengine/v1/classification.proto";
-import "openengine/v1/embedding.proto";
-import "openengine/v1/scoring.proto";
 
 message GetModelInfoRequest {
   string model = 1;
@@ -36,8 +33,6 @@ message ModelInfo {
   string reasoning_parser = 25;
   string tool_call_parser = 26;
 
-  TaskCapabilities tasks = 27; // Optional non-generative task support for this model.
-
   // Engine-specific model metadata with no portable field (e.g. kv-cache dtype,
   // quantization, capabilities not yet standardized). NOT part of the portable
   // contract: clients read it opportunistically and never depend on it for
@@ -45,11 +40,6 @@ message ModelInfo {
   google.protobuf.Struct extra = 28;
 }
 
-message TaskCapabilities {
-  EmbedCapabilities embed = 1;
-  ClassifyCapabilities classify = 2;
-  ScoreCapabilities score = 3;
-}
 
 message GenerationCapabilities {
   LogprobCapabilities prompt_logprobs = 1;
diff --git a/proto/openengine/v1/observability.proto b/proto/openengine/v1/observability.proto
index ed18f11..d254abd 100644
--- a/proto/openengine/v1/observability.proto
+++ b/proto/openengine/v1/observability.proto
@@ -1,14 +1,13 @@
 // SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 // SPDX-License-Identifier: Apache-2.0
 
-// Load snapshots and structured runtime events.
+// Load snapshots.
 
 syntax = "proto3";
 
 package openengine.v1;
 
 import "google/protobuf/struct.proto";
-import "openengine/v1/error.proto";
 
 message GetLoadRequest {
   bool include_per_rank = 1;
@@ -39,29 +38,3 @@ message RankLoadInfo {
   optional uint32 prefill_batch_size = 6;
   optional uint32 decode_batch_size = 7;
 }
-
-message SubscribeRuntimeEventsRequest {
-  repeated RuntimeEventType types = 1;
-}
-
-message SubscribeRuntimeEventsResponse {
-  oneof event {
-    RuntimeEvent runtime_event = 1;
-    EngineError error = 2; // Terminal.
-  }
-}
-
-enum RuntimeEventType {
-  RUNTIME_EVENT_TYPE_UNSPECIFIED = 0;
-  RUNTIME_EVENT_TYPE_FORWARD_PASS = 1;
-  RUNTIME_EVENT_TYPE_BATCH = 2;
-  RUNTIME_EVENT_TYPE_QUEUE = 3;
-  RUNTIME_EVENT_TYPE_TRANSFER = 4;
-}
-
-message RuntimeEvent {
-  string event_id = 1;
-  uint64 timestamp_unix_nanos = 2;
-  RuntimeEventType type = 3;
-  google.protobuf.Struct attributes = 4;
-}
diff --git a/proto/openengine/v1/openengine.proto b/proto/openengine/v1/openengine.proto
index 8a50f7c..101099d 100644
--- a/proto/openengine/v1/openengine.proto
+++ b/proto/openengine/v1/openengine.proto
@@ -16,15 +16,12 @@ syntax = "proto3";
 
 package openengine.v1;
 
-import "openengine/v1/classification.proto";
-import "openengine/v1/embedding.proto";
 import "openengine/v1/generation.proto";
 import "openengine/v1/kv.proto";
 import "openengine/v1/lifecycle.proto";
 import "openengine/v1/lora.proto";
 import "openengine/v1/model.proto";
 import "openengine/v1/observability.proto";
-import "openengine/v1/scoring.proto";
 import "openengine/v1/server.proto";
 
 // Standard ASCII gRPC request metadata:
@@ -42,11 +39,6 @@ import "openengine/v1/server.proto";
 service Inference {
   // Core inference path.
   rpc Generate(GenerateRequest) returns (stream GenerateResponse);
-
-  // Non-generative inference paths.
-  rpc Embed(EmbedRequest) returns (EmbedResponse);
-  rpc Classify(ClassifyRequest) returns (ClassifyResponse);
-  rpc Score(ScoreRequest) returns (ScoreResponse);
 }
 
 service Control {
@@ -65,11 +57,7 @@ service Control {
   rpc UnloadLora(UnloadLoraRequest) returns (UnloadLoraResponse);
   rpc ListLoras(ListLorasRequest) returns (ListLorasResponse);
 
-  // Disaggregated serving / KV transfer.
-  rpc GetKvConnectorInfo(GetKvConnectorInfoRequest) returns (KvConnectorInfo);
+  // Disaggregated serving / KV transfer. Connector info: ServerInfo.kv_connector.
   rpc GetKvEventSources(GetKvEventSourcesRequest) returns (GetKvEventSourcesResponse);
   rpc SubscribeKvEvents(SubscribeKvEventsRequest) returns (stream SubscribeKvEventsResponse);
-
-  // Structured runtime events for planners/controllers.
-  rpc SubscribeRuntimeEvents(SubscribeRuntimeEventsRequest) returns (stream SubscribeRuntimeEventsResponse);
 }
diff --git a/proto/openengine/v1/scoring.proto b/proto/openengine/v1/scoring.proto
deleted file mode 100644
index 4887913..0000000
--- a/proto/openengine/v1/scoring.proto
+++ /dev/null
@@ -1,80 +0,0 @@
-// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
-// SPDX-License-Identifier: Apache-2.0
-
-// Portable grouped query and candidate scoring inference.
-
-syntax = "proto3";
-
-package openengine.v1;
-
-import "openengine/v1/input.proto";
-import "openengine/v1/tasks.proto";
-
-enum ScoreNormalization {
-  SCORE_NORMALIZATION_UNSPECIFIED = 0;
-  SCORE_NORMALIZATION_NONE = 1;
-  SCORE_NORMALIZATION_SOFTMAX = 2;
-}
-
-message ScoreCapabilities {
-  optional bool supported = 1;
-  repeated TaskInputType input_types = 2;
-  repeated TaskOutputGranularity granularities = 3;
-  repeated TaskValueSemantics semantics = 4;
-  repeated ScoreNormalization normalizations = 5;
-  optional bool supports_label_token_scoring = 6;
-  optional bool supports_instruction = 7;
-  optional uint32 max_groups = 8;
-  optional uint32 max_candidates_per_group = 9;
-  optional uint64 max_output_values_per_candidate = 10;
-  optional bool supports_priority = 11;
-  optional bool supports_lora = 12;
-  optional bool higher_is_better = 13;
-  repeated Modality modalities = 14;
-  optional uint32 max_label_token_ids = 15;
-}
-
-message ScoreRequest {
-  TaskRequestContext context = 1;
-  repeated ScoreGroup groups = 2;
-  ScoreOptions options = 3;
-}
-
-// A query and one or more candidates scored as one correlation group.
-message ScoreGroup {
-  string group_id = 1;
-  TaskInput query = 2;
-  repeated TaskInput candidates = 3;
-}
-
-message ScoreOptions {
-  optional TaskOutputGranularity granularity = 1;
-  optional TaskValueSemantics output_semantics = 2;
-  ScoreNormalization normalization = 3;
-  repeated uint32 label_token_ids = 4;
-  string instruction = 5;
-}
-
-message ScoreResponse {
-  string request_id = 1;
-  repeated ScoreGroupOutput groups = 2;
-  TaskOutputGranularity granularity = 3;
-  TaskValueSemantics semantics = 4;
-  optional bool higher_is_better = 5;
-  ScoreNormalization normalization = 6;
-  repeated uint32 label_token_ids = 7;
-  TaskUsage usage = 8;
-}
-
-message ScoreGroupOutput {
-  string group_id = 1;
-  optional uint32 group_index = 2;
-  repeated ScoreCandidateOutput candidates = 3;
-}
-
-message ScoreCandidateOutput {
-  string candidate_id = 1;
-  optional uint32 candidate_index = 2;
-  DenseFloatTensor scores = 3;
-  repeated uint32 token_ids = 4;
-}
diff --git a/proto/openengine/v1/tasks.proto b/proto/openengine/v1/tasks.proto
deleted file mode 100644
index b5fcadc..0000000
--- a/proto/openengine/v1/tasks.proto
+++ /dev/null
@@ -1,52 +0,0 @@
-// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
-// SPDX-License-Identifier: Apache-2.0
-
-// Shared request and output vocabulary for non-generative inference tasks.
-
-syntax = "proto3";
-
-package openengine.v1;
-
-import "google/protobuf/struct.proto";
-
-// Request identity and model selection shared by every non-generative task.
-message TaskRequestContext {
-  string request_id = 1;
-  string model = 2;
-  string lora_name = 3;
-
-  // Engine-specific task parameters that have no portable field. NOT part of
-  // the portable contract: an engine MAY ignore keys it does not recognize, and
-  // a request MUST remain valid with this field empty. Clients treat it as
-  // best-effort. Carried as a Struct so JSON types survive the wire.
-  google.protobuf.Struct extra = 4;
-}
-
-// Row-major FP32 tensor. The product of shape must equal values.size().
-message DenseFloatTensor {
-  repeated uint64 shape = 1;
-  repeated float values = 2;
-}
-
-// Token accounting for one complete unary task request.
-message TaskUsage {
-  uint64 input_tokens = 1;
-  optional uint64 cached_input_tokens = 2;
-}
-
-enum TaskOutputGranularity {
-  TASK_OUTPUT_GRANULARITY_UNSPECIFIED = 0;
-  TASK_OUTPUT_GRANULARITY_SEQUENCE = 1;
-  TASK_OUTPUT_GRANULARITY_TOKEN = 2;
-}
-
-enum TaskValueSemantics {
-  TASK_VALUE_SEMANTICS_UNSPECIFIED = 0;
-  TASK_VALUE_SEMANTICS_LOGITS = 1;
-  TASK_VALUE_SEMANTICS_PROBABILITIES = 2;
-  TASK_VALUE_SEMANTICS_LOG_PROBABILITIES = 3;
-  TASK_VALUE_SEMANTICS_SIMILARITY = 4;
-  TASK_VALUE_SEMANTICS_RELEVANCE = 5;
-  TASK_VALUE_SEMANTICS_REWARD = 6;
-  TASK_VALUE_SEMANTICS_MODEL_DEFINED = 7;
-}