diff --git a/.github/actions/changes/action.yaml b/.github/actions/changes/action.yaml
index 89ebd57552..3296311c14 100644
--- a/.github/actions/changes/action.yaml
+++ b/.github/actions/changes/action.yaml
@@ -16,6 +16,9 @@ outputs:
   fabric:
     description: "'true' if the Fabric agent-eval runtime, its tests, or its dependency extra changed"
     value: ${{ steps.filter.outputs.fabric }}
+  evaluator-sdk-closure:
+    description: "'true' if anything that can change the agent_eval import closure changed"
+    value: ${{ steps.filter.outputs.evaluator-sdk-closure }}
   e2e:
     description: "'true' if any e2e test files changed"
     value: ${{ steps.filter.outputs.e2e }}
@@ -95,6 +98,13 @@ runs:
             - 'packages/nemo_evaluator_sdk/pyproject.toml'
             - '.github/workflows/ci.yaml'
             - '.github/actions/changes/action.yaml'
+          evaluator-sdk-closure:
+            # Paths that can change what `import ...harbor_runtime` loads or how the smoke job runs.
+            - 'packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/**'  # package-wide import closure
+            - 'packages/nemo_evaluator_sdk/pyproject.toml'  # declared [project] dependencies
+            - 'uv.lock'  # locked versions installed for the package
+            - '.github/workflows/ci.yaml'  # smoke job definition
+            - '.github/actions/changes/action.yaml'  # this filter (self-coverage)
           e2e:
             - 'e2e/**'
           docs:
diff --git a/.github/workflows/ci.yaml b/.github/workflows/ci.yaml
index 4e45c75c05..7af62daf74 100644
--- a/.github/workflows/ci.yaml
+++ b/.github/workflows/ci.yaml
@@ -44,6 +44,7 @@ jobs:
       test: ${{ steps.changes.outputs.test }}
       deps: ${{ steps.changes.outputs.deps }}
       fabric: ${{ steps.changes.outputs.fabric }}
+      evaluator-sdk-closure: ${{ steps.changes.outputs.evaluator-sdk-closure }}
       e2e: ${{ steps.changes.outputs.e2e }}
       docs: ${{ steps.changes.outputs.docs }}
       web-studio: ${{ steps.changes.outputs.web-studio }}
@@ -1000,6 +1001,50 @@ jobs:
           uv run --frozen --no-sync pytest \
             packages/nemo_evaluator_sdk/tests/agent_eval/test_fabric_surface.py -v
 
+  # Installs only nemo-evaluator-sdk (no workspace/extras/dev) and imports
+  # agent_eval.runtimes.harbor_runtime. Catches undeclared deps (workspace sync hides them)
+  # and broken lazy barrels that pull in the metric/execution stack.
+  evaluator-sdk-closure-smoke:
+    name: Evaluator SDK dependency-closure smoke (Linux, py${{ matrix.python-version }})
+    needs: [changes]
+    if: >
+      !cancelled() &&
+      (needs.changes.outputs.deps == 'true' || needs.changes.outputs.evaluator-sdk-closure == 'true')
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+    strategy:
+      fail-fast: false
+      matrix:
+        # Must satisfy uv.lock's requires-python (>=3.12,<3.14), which this job syncs against.
+        python-version: ["3.12"]
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10 # v6.0.3
+        with:
+          persist-credentials: false
+      - name: Install uv
+        uses: astral-sh/setup-uv@37802adc94f370d6bfd71619e3f0bf239e1f3b78 # v7.6.0
+        with:
+          python-version: ${{ matrix.python-version }}
+          enable-cache: true
+          cache-dependency-glob: uv.lock
+      - name: Install nemo-evaluator-sdk with no extras and no dev group
+        run: uv sync --frozen --package nemo-evaluator-sdk --no-dev
+      # pytest is not in the package-only sync; install the runner only (do not add SDK deps).
+      # WHY BOTH THIS JOB AND THE UNIT TEST:
+      # - Unit-test CI syncs the whole workspace: a sibling (e.g. nemo-platform-sdk) may already
+      #   provide httpx, so an undeclared SDK dep still imports — green here, broken for consumers
+      #   who only installed nemo-evaluator-sdk.
+      # - This job: `uv sync --package nemo-evaluator-sdk --no-dev` (no extras/siblings) so that
+      #   undeclared-dep case fails, then runs test_lazy_public_api (laziness assertions; one source
+      #   of truth — no duplicated heredoc probe).
+      - name: Run lazy-import closure unit test under the stripped install
+        run: |
+          uv pip install --python .venv/bin/python pytest
+          uv run --frozen --no-sync pytest \
+            packages/nemo_evaluator_sdk/tests/test_lazy_public_api.py::test_agent_eval_import_does_not_pull_the_execution_stack -v
+
   python-integration-test:
     name: Python integration tests
     needs: [policy-wasm]
@@ -1993,6 +2038,7 @@ jobs:
       - python-unit-test-tools
       - python-unit-test
       - fabric-wheel-smoke
+      - evaluator-sdk-closure-smoke
       # Enable if you want this required
       # - python-integration-test
       - require-nvskills
diff --git a/packages/nemo_evaluator_sdk/pyproject.toml b/packages/nemo_evaluator_sdk/pyproject.toml
index 7ca2798b77..6de714d919 100644
--- a/packages/nemo_evaluator_sdk/pyproject.toml
+++ b/packages/nemo_evaluator_sdk/pyproject.toml
@@ -25,6 +25,7 @@ dependencies = [
   "pyarrow>=19.0.1",
   "pandas>=1.5.3",
   "openai>=1.61.0",
+  "httpx>=0.27.0,<1",
   "sacrebleu>=2.5.1",
   "rouge_score==0.1.2",
   "ragas==0.4.3",
diff --git a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/__init__.py b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/__init__.py
index fd959d5b40..0f251c80d3 100644
--- a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/__init__.py
+++ b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/__init__.py
@@ -6,9 +6,9 @@
 The public surface resolves lazily (PEP 562). Importing this package must not drag in the
 execution/backend or metric stack: importing any submodule runs this module first, so eager
 re-exports made ``import nemo_evaluator_sdk.agent_eval.runtimes.harbor_runtime`` — all the
-optimizer needs — cost ~1400 modules (openai, sacrebleu, zstandard, pyarrow, ...) instead
-of ~480, and turned every one of those transitive packages into an evaluation-time failure mode
-for the SDK-backed evaluator.
+optimizer needs — cost ~1400 modules (openai, sacrebleu, zstandard, ...) instead of ~485, and
+turned every one of those transitive packages into an evaluation-time failure mode for the
+SDK-backed evaluator.
 
 Add a new re-export to ``_LAZY_ATTRS``, the ``TYPE_CHECKING`` block and ``__all__`` — never as a
 module-level import. ``tests/test_lazy_public_api.py`` locks the boundary in.
@@ -16,14 +16,19 @@
 
 # ruff: noqa: I001 - the vendored SDK mirror uses different import-order settings.
 
-from importlib import import_module
-from importlib.metadata import PackageNotFoundError
+from importlib import import_module as _import_module
+from importlib.metadata import PackageNotFoundError as _PackageNotFoundError
 from importlib.metadata import version as _package_version
 from typing import TYPE_CHECKING
 
 if TYPE_CHECKING:
     # Annotations and static analysis only; these must never execute at run time. Listing the
     # names in ``__all__`` is what marks them as re-exports for ruff and the type checkers.
+    #
+    # AGENTS.md ("Python Style notes") says not to import types under TYPE_CHECKING and to use a
+    # regular import "when possible". A regular import is exactly what this module exists to
+    # remove, so the exception is deliberate: these names are re-exports, not annotations, and
+    # every one of them resolves for real through ``__getattr__`` below.
     from nemo_evaluator_sdk.agent_stream_translation import (
         AgentStreamTranslation,
         AgentStreamTranslationContext,
@@ -97,10 +102,24 @@
         SecretRef,
     )
 
-try:
-    version = _package_version("nemo-evaluator-sdk")
-except PackageNotFoundError:
-    version = "0.0.0"
+
+def _resolve_version() -> str:
+    """Report the version of whichever distribution actually shipped this code.
+
+    ``nemo-evaluator-sdk`` is not published on its own — this package is also vendored into the
+    ``nemo-platform`` wheel as ``nemo_platform.beta.evaluator``. There the SDK distribution does
+    not exist, so resolving only that name reported ``"0.0.0"`` unconditionally and any telemetry
+    or support log that read it got a useless constant.
+    """
+    for distribution in ("nemo-evaluator-sdk", "nemo-platform"):
+        try:
+            return _package_version(distribution)
+        except _PackageNotFoundError:
+            continue
+    return "0.0.0"
+
+
+version = _resolve_version()
 
 # Re-exported name -> the submodule that defines it, relative to this package. Relative on
 # purpose: the vendoring tool mirrors this file into nemo_platform.beta.evaluator by rewriting
@@ -251,21 +270,29 @@
 def __getattr__(name: str) -> object:
     """Import the submodule that defines ``name`` on first access (PEP 562).
 
-    ``AttributeError`` is the required failure mode, not ``ImportError``: ``from pkg import sub``
-    only falls back to importing a submodule when attribute lookup raises ``AttributeError``, and
-    ``hasattr`` checks against this package depend on it too.
+    An *unknown* name raises ``AttributeError``, which is required: ``from pkg import sub`` only
+    falls back to importing a submodule when attribute lookup raises ``AttributeError``.
+
+    A *known* name whose submodule fails to import propagates that ``ImportError`` unchanged, and
+    that is deliberate — ``ModuleNotFoundError: No module named 'sacrebleu'`` is far more useful
+    than an ``AttributeError`` claiming ``BLEUMetric`` does not exist. The consequence is that
+    ``hasattr(nemo_evaluator_sdk, name)`` raises rather than returning ``False`` when a name's
+    dependencies are not installed, since ``hasattr`` only swallows ``AttributeError``. To probe
+    for an optional part of the surface, catch ``ImportError`` around the access instead of using
+    ``hasattr``; to probe only for name membership, test against ``__all__``.
     """
     submodule = _LAZY_ATTRS.get(name)
     if submodule is None:
         raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
-    value = getattr(import_module(submodule, __name__), name)
+    value = getattr(_import_module(submodule, __name__), name)
     globals()[name] = value  # cache, so later lookups skip __getattr__ entirely
     return value
 
 
 def __dir__() -> list[str]:
-    # The declared surface plus any submodule the caller has already imported. ``import_module``
-    # and ``TYPE_CHECKING`` are machinery for __getattr__, not API, so keep them out of
-    # autocomplete and inspect.getmembers.
-    public = {name for name in globals() if not name.startswith("_")} - {"TYPE_CHECKING", "import_module"}
+    # The declared surface plus any submodule the caller has already imported. Everything this
+    # module needs for its own machinery is imported under a leading underscore so the filter
+    # below keeps it out of autocomplete and inspect.getmembers without a name-by-name denylist;
+    # ``TYPE_CHECKING`` is the one exception, kept unaliased so type checkers still recognise it.
+    public = {name for name in globals() if not name.startswith("_")} - {"TYPE_CHECKING"}
     return sorted(set(__all__) | public)
diff --git a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/agent_eval/runtimes/docker_sandbox.py b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/agent_eval/runtimes/docker_sandbox.py
index 298ef1335f..2dac3c9480 100644
--- a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/agent_eval/runtimes/docker_sandbox.py
+++ b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/agent_eval/runtimes/docker_sandbox.py
@@ -53,9 +53,8 @@ class SandboxSDK:
 
 def _load_agents_sdk() -> SandboxSDK:
     try:
-        # The OpenAI Agents SDK ships under the `nemo-evaluator-sdk[agent-runtimes]` extra and is
-        # imported only when this Docker runtime is actually used, so it is absent from the default
-        # type-checking environment.
+        # The OpenAI Agents SDK is imported only when this Docker runtime is actually used, so it
+        # is absent from the default type-checking environment.
         from agents.run import RunConfig  # ty: ignore[unresolved-import]
         from agents.sandbox import Manifest, SandboxAgent, SandboxRunConfig  # ty: ignore[unresolved-import]
         from agents.sandbox.config import DEFAULT_PYTHON_SANDBOX_IMAGE  # ty: ignore[unresolved-import]
@@ -68,7 +67,14 @@ def _load_agents_sdk() -> SandboxSDK:
         from agents import Runner  # ty: ignore[unresolved-import]
         from docker import from_env as docker_from_env
     except ImportError as exc:
-        raise RuntimeError("DockerSandboxAgentRuntime requires `nemo-evaluator-sdk[agent-runtimes]`") from exc
+        # Audience split is in the error text: SDK extras are not propagated into the
+        # vendored nemo_platform.beta.evaluator mirror.
+        raise RuntimeError(
+            "DockerSandboxAgentRuntime requires the openai-agents[docker] Python packages. "
+            "Standalone SDK: pip install 'nemo-evaluator-sdk[agent-runtimes]'. "
+            "Vendored nemo-platform.beta.evaluator (no SDK extras): "
+            "pip install 'openai-agents[docker]'"
+        ) from exc
 
     return SandboxSDK(
         Runner=Runner,
diff --git a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/agent_inference.py b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/agent_inference.py
index fcde7d8d62..a9b95197f2 100644
--- a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/agent_inference.py
+++ b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/agent_inference.py
@@ -27,7 +27,6 @@
 
 import httpx
 from httpx import Timeout
-from jsonpath_ng import parse as jsonpath_parse
 from pydantic import BaseModel, ConfigDict, Field
 
 from nemo_evaluator_sdk.agent_stream_translation import (
@@ -788,6 +787,10 @@ def _extract_jsonpath(
     required: bool = True,
 ) -> Any:
     """Extract a value from data using a JSONPath expression."""
+    # Imported here rather than at module scope: this is jsonpath_ng's only use in the module, and
+    # the module sits on the agent_eval run-time path via agent_eval/evaluator.py.
+    from jsonpath_ng import parse as jsonpath_parse
+
     expr = jsonpath_parse(path)
     matches = expr.find(data)
     if not matches:
diff --git a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/metrics/aggregation.py b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/metrics/aggregation.py
index 4a626e9c1a..def34d4e6b 100644
--- a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/metrics/aggregation.py
+++ b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/metrics/aggregation.py
@@ -3,10 +3,12 @@
 
 """Aggregation data structures and computations for metric results."""
 
+from __future__ import annotations
+
 import math
 from collections import OrderedDict, defaultdict
 from collections.abc import Mapping, Sequence
-from typing import Protocol, cast, runtime_checkable
+from typing import TYPE_CHECKING, Protocol, cast, runtime_checkable
 
 from nemo_evaluator_sdk.metrics.protocol import (
     BooleanValue,
@@ -28,7 +30,11 @@
     RubricScoreStat,
     ScoreStats,
 )
-from nemo_evaluator_sdk.values.scores import RubricScore, Score
+
+if TYPE_CHECKING:
+    # Importing the score configuration module loads jsonschema. Aggregation's lightweight helpers
+    # (notably compute_percentiles) do not need it, so keep this typing-only edge deferred.
+    from nemo_evaluator_sdk.values.scores import Score
 
 
 def is_aggregateable_output_spec(output_spec: MetricOutputSpec) -> bool:
@@ -445,6 +451,8 @@ def aggregate_metrics(
 
 def rubric_definitions_from_scores(scores: Sequence[Score]) -> dict[str, list[RubricScoreStat]]:
     """Return declared rubric buckets keyed by score name."""
+    from nemo_evaluator_sdk.values.scores import RubricScore
+
     definitions: dict[str, list[RubricScoreStat]] = {}
     for score in scores:
         if not isinstance(score, RubricScore):
diff --git a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/values/__init__.py b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/values/__init__.py
index a92357e26f..37f80764a5 100644
--- a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/values/__init__.py
+++ b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/values/__init__.py
@@ -1,121 +1,241 @@
 # SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
-"""Public value types for evaluator SDK runtime."""
+"""Public value types for evaluator SDK runtime.
+
+The public interface resolves lazily (PEP 562), for the same reason the package root does:
+every ``from nemo_evaluator_sdk.values.X import ...`` runs this barrel first, so
+eagerly re-exporting all 97 names dragged ``.datasets``/``.results`` (pyarrow, numpy) and
+``.metrics``/``.scores`` (jsonschema, jinja2) into ``agent_eval``, which uses none of them.
+Measured: 485 modules and +57 MB RSS for ``import agent_eval.runtimes.harbor_runtime`` before,
+300 modules and pydantic alone after.
+
+Add a new re-export to ``_LAZY_ATTRS``, the ``TYPE_CHECKING`` block and ``__all__`` — never as a
+module-level import. ``tests/test_lazy_public_api.py`` locks the boundary in.
+"""
 
 # ruff: noqa: I001 - the vendored SDK mirror uses different import-order settings.
 
-from nemo_evaluator_sdk.values.agents import (
-    Agent,
-    AgentBase,
-    GenericAgent,
-    NatAgentConfig,
-    NemoAgentToolkitAgent,
-)
-from nemo_evaluator_sdk.values.atif import (
-    FinalMetrics,
-    Metrics,
-    Observation,
-    ObservationResult,
-    Step,
-    ToolCall,
-    Trajectory,
-)
-from nemo_evaluator_sdk.values.common import SecretRef, SupportedJobTypes
-from nemo_evaluator_sdk.values.dataset_schemas import (
-    FieldMapping,
-    InputSchema,
-)
-from nemo_evaluator_sdk.values.datasets import DatasetInput, DatasetRows
-from nemo_evaluator_sdk.values.evidence import (
-    CandidateEvidence,
-    CommandResult,
-    EvidenceDescriptor,
-    FilesystemDiff,
-    FilesystemEntry,
-    LocalFilesystemEvidence,
-    LogHandle,
-    TraceHandle,
-    WellKnownEvidenceKey,
-    parse_atif,
-)
-from nemo_evaluator_sdk.values.metrics import (
-    BLEU,
-    F1,
-    ROUGE,
-    AgentGoalAccuracy,
-    AnswerAccuracy,
-    ContextEntityRecall,
-    ContextPrecision,
-    ContextRecall,
-    ContextRelevance,
-    ExactMatch,
-    Faithfulness,
-    LLMJudge,
-    MetricBase,
-    NemoAgentToolkitRemote,
-    NoiseSensitivity,
-    NumberCheck,
-    Remote,
-    ResponseGroundedness,
-    ResponseRelevancy,
-    StringCheck,
-    ToolCallAccuracy,
-    ToolCalling,
-    TopicAdherence,
-    TunableRagEvaluator,
-)
-from nemo_evaluator_sdk.values.models import Model, ModelRef, ReasoningParams
-from nemo_evaluator_sdk.values.params import (
-    InferenceParams,
-    RunConfig,
-    RunConfigOnline,
-    RunConfigOnlineModel,
-)
-from nemo_evaluator_sdk.values.protocol import (
-    BooleanValue,
-    CandidateOutput,
-    ContinuousScore,
-    DatasetRow,
-    DiscreteScore,
-    Label,
-    MetricDescriptor,
-    MetricDiagnostic,
-    MetricInput,
-    MetricOutput,
-    MetricOutputSpec,
-    MetricResult,
-    MetricTypeName,
-)
-from nemo_evaluator_sdk.values.results import (
-    AggregatedMetricResult,
-    AggregateFieldName,
-    AggregateRangeScore,
-    AggregateRubricScore,
-    AggregateScore,
-    AggregateScoreBase,
-    DefaultAggregateFieldName,
-    EvaluationResult,
-    Histogram,
-    HistogramBin,
-    MetricScore,
-    Percentiles,
-    RowScore,
-    RubricScoreStat,
-    RubricScoreValue,
-    SampleResult,
-    ScoreStats,
-)
-from nemo_evaluator_sdk.values.scores import (
-    JSONScoreParser,
-    RangeScore,
-    RegexScoreParser,
-    RemoteScore,
-    Rubric,
-    RubricScore,
-    Score,
-    score_discriminator,
-)
+from importlib import import_module as _import_module
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from nemo_evaluator_sdk.values.agents import (
+        Agent,
+        AgentBase,
+        GenericAgent,
+        NatAgentConfig,
+        NemoAgentToolkitAgent,
+    )
+    from nemo_evaluator_sdk.values.atif import (
+        FinalMetrics,
+        Metrics,
+        Observation,
+        ObservationResult,
+        Step,
+        ToolCall,
+        Trajectory,
+    )
+    from nemo_evaluator_sdk.values.common import SecretRef, SupportedJobTypes
+    from nemo_evaluator_sdk.values.dataset_schemas import (
+        FieldMapping,
+        InputSchema,
+    )
+    from nemo_evaluator_sdk.values.datasets import DatasetInput, DatasetRows
+    from nemo_evaluator_sdk.values.evidence import (
+        CandidateEvidence,
+        CommandResult,
+        EvidenceDescriptor,
+        FilesystemDiff,
+        FilesystemEntry,
+        LocalFilesystemEvidence,
+        LogHandle,
+        TraceHandle,
+        WellKnownEvidenceKey,
+        parse_atif,
+    )
+    from nemo_evaluator_sdk.values.metrics import (
+        BLEU,
+        F1,
+        ROUGE,
+        AgentGoalAccuracy,
+        AnswerAccuracy,
+        ContextEntityRecall,
+        ContextPrecision,
+        ContextRecall,
+        ContextRelevance,
+        ExactMatch,
+        Faithfulness,
+        LLMJudge,
+        MetricBase,
+        NemoAgentToolkitRemote,
+        NoiseSensitivity,
+        NumberCheck,
+        Remote,
+        ResponseGroundedness,
+        ResponseRelevancy,
+        StringCheck,
+        ToolCallAccuracy,
+        ToolCalling,
+        TopicAdherence,
+        TunableRagEvaluator,
+    )
+    from nemo_evaluator_sdk.values.models import Model, ModelRef, ReasoningParams
+    from nemo_evaluator_sdk.values.params import (
+        InferenceParams,
+        RunConfig,
+        RunConfigOnline,
+        RunConfigOnlineModel,
+    )
+    from nemo_evaluator_sdk.values.protocol import (
+        BooleanValue,
+        CandidateOutput,
+        ContinuousScore,
+        DatasetRow,
+        DiscreteScore,
+        Label,
+        MetricDescriptor,
+        MetricDiagnostic,
+        MetricInput,
+        MetricOutput,
+        MetricOutputSpec,
+        MetricResult,
+        MetricTypeName,
+    )
+    from nemo_evaluator_sdk.values.results import (
+        AggregatedMetricResult,
+        AggregateFieldName,
+        AggregateRangeScore,
+        AggregateRubricScore,
+        AggregateScore,
+        AggregateScoreBase,
+        DefaultAggregateFieldName,
+        EvaluationResult,
+        Histogram,
+        HistogramBin,
+        MetricScore,
+        Percentiles,
+        RowScore,
+        RubricScoreStat,
+        RubricScoreValue,
+        SampleResult,
+        ScoreStats,
+    )
+    from nemo_evaluator_sdk.values.scores import (
+        JSONScoreParser,
+        RangeScore,
+        RegexScoreParser,
+        RemoteScore,
+        Rubric,
+        RubricScore,
+        Score,
+        score_discriminator,
+    )
+
+
+# Re-exported name -> the submodule that defines it, relative to this package. Relative on
+# purpose: the vendoring tool mirrors this file into nemo_platform.beta.evaluator by rewriting
+# module paths, and a relative name has nothing to rewrite, so the mirror is correct by
+# construction. Mirrors the TYPE_CHECKING block above, in the same order.
+_LAZY_ATTRS: dict[str, str] = {
+    "Agent": ".agents",
+    "AgentBase": ".agents",
+    "GenericAgent": ".agents",
+    "NatAgentConfig": ".agents",
+    "NemoAgentToolkitAgent": ".agents",
+    "FinalMetrics": ".atif",
+    "Metrics": ".atif",
+    "Observation": ".atif",
+    "ObservationResult": ".atif",
+    "Step": ".atif",
+    "ToolCall": ".atif",
+    "Trajectory": ".atif",
+    "SecretRef": ".common",
+    "SupportedJobTypes": ".common",
+    "FieldMapping": ".dataset_schemas",
+    "InputSchema": ".dataset_schemas",
+    "DatasetInput": ".datasets",
+    "DatasetRows": ".datasets",
+    "CandidateEvidence": ".evidence",
+    "CommandResult": ".evidence",
+    "EvidenceDescriptor": ".evidence",
+    "FilesystemDiff": ".evidence",
+    "FilesystemEntry": ".evidence",
+    "LocalFilesystemEvidence": ".evidence",
+    "LogHandle": ".evidence",
+    "TraceHandle": ".evidence",
+    "WellKnownEvidenceKey": ".evidence",
+    "parse_atif": ".evidence",
+    "BLEU": ".metrics",
+    "F1": ".metrics",
+    "ROUGE": ".metrics",
+    "AgentGoalAccuracy": ".metrics",
+    "AnswerAccuracy": ".metrics",
+    "ContextEntityRecall": ".metrics",
+    "ContextPrecision": ".metrics",
+    "ContextRecall": ".metrics",
+    "ContextRelevance": ".metrics",
+    "ExactMatch": ".metrics",
+    "Faithfulness": ".metrics",
+    "LLMJudge": ".metrics",
+    "MetricBase": ".metrics",
+    "NemoAgentToolkitRemote": ".metrics",
+    "NoiseSensitivity": ".metrics",
+    "NumberCheck": ".metrics",
+    "Remote": ".metrics",
+    "ResponseGroundedness": ".metrics",
+    "ResponseRelevancy": ".metrics",
+    "StringCheck": ".metrics",
+    "ToolCallAccuracy": ".metrics",
+    "ToolCalling": ".metrics",
+    "TopicAdherence": ".metrics",
+    "TunableRagEvaluator": ".metrics",
+    "Model": ".models",
+    "ModelRef": ".models",
+    "ReasoningParams": ".models",
+    "InferenceParams": ".params",
+    "RunConfig": ".params",
+    "RunConfigOnline": ".params",
+    "RunConfigOnlineModel": ".params",
+    "BooleanValue": ".protocol",
+    "CandidateOutput": ".protocol",
+    "ContinuousScore": ".protocol",
+    "DatasetRow": ".protocol",
+    "DiscreteScore": ".protocol",
+    "Label": ".protocol",
+    "MetricDescriptor": ".protocol",
+    "MetricDiagnostic": ".protocol",
+    "MetricInput": ".protocol",
+    "MetricOutput": ".protocol",
+    "MetricOutputSpec": ".protocol",
+    "MetricResult": ".protocol",
+    "MetricTypeName": ".protocol",
+    "AggregatedMetricResult": ".results",
+    "AggregateFieldName": ".results",
+    "AggregateRangeScore": ".results",
+    "AggregateRubricScore": ".results",
+    "AggregateScore": ".results",
+    "AggregateScoreBase": ".results",
+    "DefaultAggregateFieldName": ".results",
+    "EvaluationResult": ".results",
+    "Histogram": ".results",
+    "HistogramBin": ".results",
+    "MetricScore": ".results",
+    "Percentiles": ".results",
+    "RowScore": ".results",
+    "RubricScoreStat": ".results",
+    "RubricScoreValue": ".results",
+    "SampleResult": ".results",
+    "ScoreStats": ".results",
+    "JSONScoreParser": ".scores",
+    "RangeScore": ".scores",
+    "RegexScoreParser": ".scores",
+    "RemoteScore": ".scores",
+    "Rubric": ".scores",
+    "RubricScore": ".scores",
+    "Score": ".scores",
+    "score_discriminator": ".scores",
+}
 
 __all__ = [
     "Agent",
@@ -217,3 +337,30 @@
     "TopicAdherence",
     "TunableRagEvaluator",
 ]
+
+
+def __getattr__(name: str) -> object:
+    """Import the submodule that defines ``name`` on first access (PEP 562).
+
+    An *unknown* name raises ``AttributeError``, which is required: ``from pkg import sub`` only
+    falls back to importing a submodule when attribute lookup raises ``AttributeError``.
+
+    A *known* name whose submodule fails to import propagates that ``ImportError`` unchanged, so
+    the real cause is not hidden behind a bogus "no attribute". The consequence is that
+    ``hasattr`` raises rather than returning ``False`` when a name's dependencies are missing;
+    catch ``ImportError`` around the access, or test membership against ``__all__``.
+    """
+    submodule = _LAZY_ATTRS.get(name)
+    if submodule is None:
+        raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
+    value = getattr(_import_module(submodule, __name__), name)
+    globals()[name] = value  # cache, so later lookups skip __getattr__ entirely
+    return value
+
+
+def __dir__() -> list[str]:
+    # The declared surface plus any submodule the caller has already imported. Machinery is
+    # imported under a leading underscore so the filter keeps it out of autocomplete without a
+    # denylist; ``TYPE_CHECKING`` is the one exception, unaliased so type checkers recognise it.
+    public = {name for name in globals() if not name.startswith("_")} - {"TYPE_CHECKING"}
+    return sorted(set(__all__) | public)
diff --git a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/values/results.py b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/values/results.py
index 02eef50977..9c06fefd76 100644
--- a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/values/results.py
+++ b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/values/results.py
@@ -9,11 +9,13 @@
 import math
 from collections.abc import Mapping
 from difflib import get_close_matches
-from typing import Any, Literal, Self
+from typing import TYPE_CHECKING, Any, Literal, Self
 
-import pyarrow as pa
 from pydantic import BaseModel, ConfigDict, Field, field_serializer, field_validator, model_serializer
 
+if TYPE_CHECKING:
+    import pyarrow as pa
+
 from nemo_evaluator_sdk.values.protocol import MetricDiagnostic, MetricOutput, MetricResult
 
 ResultView = Literal["rows", "aggregate"]
@@ -792,6 +794,12 @@ def to_table(self, view: ResultView = "rows") -> pa.Table:
         Returns:
             Table built from ``to_records(view=view)``.
         """
+        # Imported here rather than at module scope: pyarrow (plus its numpy tail) costs ~31 MB
+        # RSS and 223 modules, this is its only runtime use in the module, and the module is on
+        # the agent_eval import path, which never calls this method. The return annotation is a
+        # string already (`from __future__ import annotations`), so it needs no import.
+        import pyarrow as pa
+
         return pa.Table.from_pylist(self.to_records(view=view))
 
     def to_pandas(self, view: ResultView = "rows"):
diff --git a/packages/nemo_evaluator_sdk/tests/agent_eval/test_docker_sandbox_runtime.py b/packages/nemo_evaluator_sdk/tests/agent_eval/test_docker_sandbox_runtime.py
index c33faae163..9d300b967e 100644
--- a/packages/nemo_evaluator_sdk/tests/agent_eval/test_docker_sandbox_runtime.py
+++ b/packages/nemo_evaluator_sdk/tests/agent_eval/test_docker_sandbox_runtime.py
@@ -196,9 +196,13 @@ def fake_import(
 
     monkeypatch.setattr(builtins, "__import__", fake_import)
 
-    with pytest.raises(RuntimeError, match=r"nemo-evaluator-sdk\[agent-runtimes\]"):
+    with pytest.raises(RuntimeError, match=r"nemo-evaluator-sdk\[agent-runtimes\]") as exc_info:
         docker_sandbox._load_agents_sdk()
 
+    message = str(exc_info.value)
+    assert "pip install 'openai-agents[docker]'" in message
+    assert "openai-agents[docker]>=" not in message
+
 
 def test_manifest_uses_instruction_and_never_leaks_intent() -> None:
     # The instruction is surfaced verbatim; `task.intent` is eval-side metadata and must never leak to
diff --git a/packages/nemo_evaluator_sdk/tests/test_lazy_public_api.py b/packages/nemo_evaluator_sdk/tests/test_lazy_public_api.py
index ae24309a5d..771251e1c9 100644
--- a/packages/nemo_evaluator_sdk/tests/test_lazy_public_api.py
+++ b/packages/nemo_evaluator_sdk/tests/test_lazy_public_api.py
@@ -1,29 +1,77 @@
 # SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
-"""Guards the lazy (PEP 562) public surface of ``nemo_evaluator_sdk/__init__.py``.
+"""Guards the lazy (PEP 562) public surfaces of the evaluator SDK.
 
-Importing any submodule runs the package ``__init__`` first, so a single convenience
-``from nemo_evaluator_sdk.… import …`` added at module scope silently re-drags the whole
-backend/benchmark/metric stack into every consumer that only wanted ``agent_eval`` — the
-regression these tests exist to catch (AALGO-429).
+Two barrels re-export lazily, and importing any submodule runs both in turn, so a single
+convenience ``from … import …`` added at either module scope silently re-drags the whole
+backend/benchmark/metric stack into every consumer that only wanted ``agent_eval``:
+
+* ``nemo_evaluator_sdk/__init__.py`` (AALGO-429) — the execution/backend and metric stack.
+* ``nemo_evaluator_sdk/values/__init__.py`` (AALGO-311) — pyarrow, numpy, jinja2 and jsonschema,
+  together with the deferred pyarrow import in ``values/results.py``.
+
+Both are covered here in their source form and in the ``nemo_platform.beta.evaluator`` mirror the
+vendoring tool generates.
+
+Every *assertion* runs out-of-process. Resolving a whole public surface imports openai, sacrebleu,
+ragas and the execution stack; doing that in-process would leave them in ``sys.modules`` for every
+test that runs after it in the session, so a future in-process "module X must not be imported"
+check — the natural way someone would extend this file — would depend on collection order. The one
+in-process call is the ``find_spec`` availability probe below, which is deliberately restricted to
+a top-level name so that it imports nothing.
 """
 
+import importlib.util
 import json
 import subprocess
 import sys
 
 import pytest
 
+_VENDORED_MIRROR = "nemo_platform.beta.evaluator"
+
 # Imported out-of-process on purpose: by the time this module runs under pytest, sibling suites
 # have already pulled the execution stack into sys.modules, so an in-process check proves nothing.
-_PROBE = """
+_IMPORT_SURFACE_PROBE = """
 import json, sys
 from nemo_evaluator_sdk.agent_eval.runtimes.harbor_runtime import HarborAgentTaskRunner
 assert HarborAgentTaskRunner is not None
 print(json.dumps(sorted(sys.modules)))
 """
 
+# Resolves every declared name for real. This is also the only thing left that validates each
+# re-exported submodule still imports at all: before the package went lazy, `import
+# nemo_evaluator_sdk` executed all of them, so a syntax error or a broken third-party import
+# failed immediately everywhere. Narrowing this loop, skipping this test, or dropping a name from
+# `__all__` silently gives that guarantee up.
+_PUBLIC_SURFACE_PROBE = """
+import json, sys
+
+name, submodule_name = sys.argv[1], sys.argv[2]
+module = __import__(name, fromlist=["__all__"])
+
+for attribute in module.__all__:      # AttributeError/ImportError names the offender in stderr
+    getattr(module, attribute)
+
+try:
+    module.NoSuchName
+except AttributeError:
+    pass
+else:
+    raise AssertionError("unknown attribute did not raise AttributeError")
+
+# The `from pkg import submodule` fallback, which only fires when __getattr__ raises
+# AttributeError rather than KeyError or ImportError.
+submodule = getattr(__import__(name, fromlist=[submodule_name]), submodule_name)
+assert submodule.__name__ == f"{name}.{submodule_name}", submodule.__name__
+
+print(json.dumps({
+    "resolved": len(module.__all__),
+    "missing_from_dir": sorted(set(module.__all__) - set(dir(module))),
+}))
+"""
+
 
 def test_agent_eval_import_does_not_pull_the_execution_stack() -> None:
     """The optimizer imports only ``agent_eval``; it must not pay for backends and benchmarks.
@@ -32,7 +80,7 @@ def test_agent_eval_import_does_not_pull_the_execution_stack() -> None:
     SDK-path failure at evaluation time. Keeping the boundary tight is what lets the consumer's
     deferred SDK import actually contain a broken install.
     """
-    proc = subprocess.run([sys.executable, "-c", _PROBE], capture_output=True, text=True, timeout=120)
+    proc = subprocess.run([sys.executable, "-c", _IMPORT_SURFACE_PROBE], capture_output=True, text=True, timeout=120)
     assert proc.returncode == 0, proc.stderr
 
     # Last line only: the child inherits the environment, so a sitecustomize banner or an
@@ -42,40 +90,62 @@ def test_agent_eval_import_does_not_pull_the_execution_stack() -> None:
     leaked = sorted(name for name in modules if name.startswith("nemo_evaluator_sdk.execution"))
     assert leaked == [], f"the package __init__ re-drags the execution stack: {leaked}"
 
-    # Exactly the packages the eager __init__ used to drag in; each one goes red if it returns.
+    # Every one of these was on this path before the two barrels went lazy, so each goes red if it
+    # returns: openai/sacrebleu/zstandard came from the root __init__, and pyarrow/numpy/jinja2/
+    # jsonschema from the values/ barrel plus values/results.py's module-scope pyarrow import.
     # rouge_score is deliberately absent: metrics/rouge.py already defers it into a
     # cached_property, so it was never on this path and asserting it would prove nothing.
-    heavy = {"openai", "sacrebleu", "zstandard"} & modules
-    assert heavy == set(), f"heavy metric-stack dependencies pulled in: {sorted(heavy)}"
-
-    # A canary, not a spec: measured at 483 modules when this landed, down from 1416. Raise the
-    # bound if a genuine agent_eval dependency lands; a jump of >100 means something re-drags a
-    # barrel module and should be investigated rather than accommodated.
-    assert len(modules) < 700, f"agent_eval import surface grew to {len(modules)} modules"
-
-
-def test_every_public_name_resolves() -> None:
-    """``__all__`` and ``_LAZY_ATTRS`` must not drift apart.
+    heavy = {"openai", "sacrebleu", "zstandard", "pyarrow", "numpy", "jinja2", "jsonschema"} & modules
+    assert heavy == set(), f"heavy dependencies pulled into the agent_eval path: {sorted(heavy)}"
+
+    # A canary, not a spec: measured at 300 modules once both barrels went lazy, down from 1416.
+    # The bound is deliberately close — the assertions above enumerate known offenders, so only
+    # this catches a re-drag through some other route. Raise it only for a dependency agent_eval
+    # genuinely needs, and say which in the commit message.
+    assert len(modules) < 380, f"agent_eval import surface grew to {len(modules)} modules"
+
+
+@pytest.mark.parametrize(
+    ("module_name", "submodule_name"),
+    [
+        ("nemo_evaluator_sdk", "values"),
+        ("nemo_evaluator_sdk.values", "models"),
+        (_VENDORED_MIRROR, "values"),
+        (f"{_VENDORED_MIRROR}.values", "models"),
+    ],
+)
+def test_every_public_name_resolves(module_name: str, submodule_name: str) -> None:
+    """``__all__`` and ``_LAZY_ATTRS`` must not drift apart, in the source or the vendored mirror.
 
     A typo in the lazy table is invisible until a consumer hits that one attribute, so resolve
     the whole surface in one pass.
-    """
-    import nemo_evaluator_sdk
-
-    for name in nemo_evaluator_sdk.__all__:
-        assert getattr(nemo_evaluator_sdk, name) is not None, name
 
-    assert set(nemo_evaluator_sdk.__all__) <= set(dir(nemo_evaluator_sdk))
-
-
-def test_unknown_attribute_raises_attribute_error() -> None:
-    """``from pkg import submodule`` and ``hasattr`` both rely on ``AttributeError`` here."""
-    import nemo_evaluator_sdk
-
-    with pytest.raises(AttributeError):
-        nemo_evaluator_sdk.NoSuchName  # noqa: B018
+    The mirror leg is what keeps the relative module paths in ``_LAZY_ATTRS`` honest. They are
+    relative precisely so the vendoring tool has nothing to rewrite; if someone switches them to
+    an f-string (which the rewriter does not touch) or to absolute literals under a changed
+    rewriter, the mirror resolves to the wrong package. Existing tests import *through* the mirror
+    but never resolve its surface, so nothing else would notice.
+    """
+    # Probe the top-level package only. `find_spec` on a dotted name RAISES ModuleNotFoundError
+    # when a parent is missing rather than returning None, so probing the full path would error
+    # in exactly the case this guard exists for (no nemo-platform installed). It also imports
+    # every parent in-process, which would defeat this module's isolation.
+    root_package = module_name.partition(".")[0]
+    if importlib.util.find_spec(root_package) is None:
+        pytest.skip(f"{root_package} is not installed in this environment")
+
+    proc = subprocess.run(
+        [sys.executable, "-c", _PUBLIC_SURFACE_PROBE, module_name, submodule_name],
+        capture_output=True,
+        text=True,
+        timeout=120,
+    )
+    assert proc.returncode == 0, proc.stderr
 
-    # The submodule fallback that AttributeError enables.
-    from nemo_evaluator_sdk import values
+    result = json.loads(proc.stdout.strip().splitlines()[-1])
+    assert result["missing_from_dir"] == [], f"__all__ names absent from dir(): {result['missing_from_dir']}"
+    assert result["resolved"] > 0
 
-    assert values.__name__ == "nemo_evaluator_sdk.values"
+    # `version` is deliberately not asserted here: this workspace installs nemo-evaluator-sdk
+    # itself, and its declared version is literally "0.0.0", so the distribution-fallback in
+    # _resolve_version() only changes behaviour in a built wheel, where the SDK is absent.
diff --git a/packages/nemo_platform/pyproject.toml b/packages/nemo_platform/pyproject.toml
index 6a7aa8af0b..e2acf75f19 100644
--- a/packages/nemo_platform/pyproject.toml
+++ b/packages/nemo_platform/pyproject.toml
@@ -292,6 +292,7 @@ nemo-evaluator-sdk = [
   "pyarrow>=19.0.1",
   "pandas>=1.5.3",
   "openai>=1.61.0",
+  "httpx>=0.27.0,<1",
   "sacrebleu>=2.5.1",
   "rouge_score==0.1.2",
   "ragas==0.4.3",
diff --git a/sdk/python/nemo-platform/pyproject.toml b/sdk/python/nemo-platform/pyproject.toml
index 3cdbb33916..d45b878903 100644
--- a/sdk/python/nemo-platform/pyproject.toml
+++ b/sdk/python/nemo-platform/pyproject.toml
@@ -62,6 +62,7 @@ nemo-evaluator-sdk = [
   "langchain-nvidia-ai-endpoints>=1.4.3,<2.0.0",
   "nemo-relay>=0.6.0,<0.7",
   "nemo-fabric>=0.1.1,<0.3.0",
+  "httpx>=0.27.0,<1",
 ]
 
 [project.entry-points."nemo.skills"]
diff --git a/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/__init__.py b/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/__init__.py
index 54b5891b76..48cda897cc 100644
--- a/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/__init__.py
+++ b/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/__init__.py
@@ -6,9 +6,9 @@
 The public surface resolves lazily (PEP 562). Importing this package must not drag in the
 execution/backend or metric stack: importing any submodule runs this module first, so eager
 re-exports made ``import nemo_platform.beta.evaluator.agent_eval.runtimes.harbor_runtime`` — all the
-optimizer needs — cost ~1400 modules (openai, sacrebleu, zstandard, pyarrow, ...) instead
-of ~480, and turned every one of those transitive packages into an evaluation-time failure mode
-for the SDK-backed evaluator.
+optimizer needs — cost ~1400 modules (openai, sacrebleu, zstandard, ...) instead of ~485, and
+turned every one of those transitive packages into an evaluation-time failure mode for the
+SDK-backed evaluator.
 
 Add a new re-export to ``_LAZY_ATTRS``, the ``TYPE_CHECKING`` block and ``__all__`` — never as a
 module-level import. ``tests/test_lazy_public_api.py`` locks the boundary in.
@@ -16,14 +16,19 @@
 
 # ruff: noqa: I001 - the vendored SDK mirror uses different import-order settings.
 
-from importlib import import_module
-from importlib.metadata import PackageNotFoundError
+from importlib import import_module as _import_module
+from importlib.metadata import PackageNotFoundError as _PackageNotFoundError
 from importlib.metadata import version as _package_version
 from typing import TYPE_CHECKING
 
 if TYPE_CHECKING:
     # Annotations and static analysis only; these must never execute at run time. Listing the
     # names in ``__all__`` is what marks them as re-exports for ruff and the type checkers.
+    #
+    # AGENTS.md ("Python Style notes") says not to import types under TYPE_CHECKING and to use a
+    # regular import "when possible". A regular import is exactly what this module exists to
+    # remove, so the exception is deliberate: these names are re-exports, not annotations, and
+    # every one of them resolves for real through ``__getattr__`` below.
     from nemo_platform.beta.evaluator.agent_stream_translation import (
         AgentStreamTranslation,
         AgentStreamTranslationContext,
@@ -97,10 +102,24 @@
         SecretRef,
     )
 
-try:
-    version = _package_version("nemo-evaluator-sdk")
-except PackageNotFoundError:
-    version = "0.0.0"
+
+def _resolve_version() -> str:
+    """Report the version of whichever distribution actually shipped this code.
+
+    ``nemo-evaluator-sdk`` is not published on its own — this package is also vendored into the
+    ``nemo-platform`` wheel as ``nemo_platform.beta.evaluator``. There the SDK distribution does
+    not exist, so resolving only that name reported ``"0.0.0"`` unconditionally and any telemetry
+    or support log that read it got a useless constant.
+    """
+    for distribution in ("nemo-evaluator-sdk", "nemo-platform"):
+        try:
+            return _package_version(distribution)
+        except _PackageNotFoundError:
+            continue
+    return "0.0.0"
+
+
+version = _resolve_version()
 
 # Re-exported name -> the submodule that defines it, relative to this package. Relative on
 # purpose: the vendoring tool mirrors this file into nemo_platform.beta.evaluator by rewriting
@@ -251,21 +270,29 @@
 def __getattr__(name: str) -> object:
     """Import the submodule that defines ``name`` on first access (PEP 562).
 
-    ``AttributeError`` is the required failure mode, not ``ImportError``: ``from pkg import sub``
-    only falls back to importing a submodule when attribute lookup raises ``AttributeError``, and
-    ``hasattr`` checks against this package depend on it too.
+    An *unknown* name raises ``AttributeError``, which is required: ``from pkg import sub`` only
+    falls back to importing a submodule when attribute lookup raises ``AttributeError``.
+
+    A *known* name whose submodule fails to import propagates that ``ImportError`` unchanged, and
+    that is deliberate — ``ModuleNotFoundError: No module named 'sacrebleu'`` is far more useful
+    than an ``AttributeError`` claiming ``BLEUMetric`` does not exist. The consequence is that
+    ``hasattr(nemo_evaluator_sdk, name)`` raises rather than returning ``False`` when a name's
+    dependencies are not installed, since ``hasattr`` only swallows ``AttributeError``. To probe
+    for an optional part of the surface, catch ``ImportError`` around the access instead of using
+    ``hasattr``; to probe only for name membership, test against ``__all__``.
     """
     submodule = _LAZY_ATTRS.get(name)
     if submodule is None:
         raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
-    value = getattr(import_module(submodule, __name__), name)
+    value = getattr(_import_module(submodule, __name__), name)
     globals()[name] = value  # cache, so later lookups skip __getattr__ entirely
     return value
 
 
 def __dir__() -> list[str]:
-    # The declared surface plus any submodule the caller has already imported. ``import_module``
-    # and ``TYPE_CHECKING`` are machinery for __getattr__, not API, so keep them out of
-    # autocomplete and inspect.getmembers.
-    public = {name for name in globals() if not name.startswith("_")} - {"TYPE_CHECKING", "import_module"}
+    # The declared surface plus any submodule the caller has already imported. Everything this
+    # module needs for its own machinery is imported under a leading underscore so the filter
+    # below keeps it out of autocomplete and inspect.getmembers without a name-by-name denylist;
+    # ``TYPE_CHECKING`` is the one exception, kept unaliased so type checkers still recognise it.
+    public = {name for name in globals() if not name.startswith("_")} - {"TYPE_CHECKING"}
     return sorted(set(__all__) | public)
diff --git a/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/agent_eval/runtimes/docker_sandbox.py b/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/agent_eval/runtimes/docker_sandbox.py
index ac7f2544d5..bdf1cc979e 100644
--- a/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/agent_eval/runtimes/docker_sandbox.py
+++ b/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/agent_eval/runtimes/docker_sandbox.py
@@ -53,9 +53,8 @@ class SandboxSDK:
 
 def _load_agents_sdk() -> SandboxSDK:
     try:
-        # The OpenAI Agents SDK ships under the `nemo-evaluator-sdk[agent-runtimes]` extra and is
-        # imported only when this Docker runtime is actually used, so it is absent from the default
-        # type-checking environment.
+        # The OpenAI Agents SDK is imported only when this Docker runtime is actually used, so it
+        # is absent from the default type-checking environment.
         from agents.run import RunConfig  # ty: ignore[unresolved-import]
         from agents.sandbox import Manifest, SandboxAgent, SandboxRunConfig  # ty: ignore[unresolved-import]
         from agents.sandbox.config import DEFAULT_PYTHON_SANDBOX_IMAGE  # ty: ignore[unresolved-import]
@@ -68,7 +67,14 @@ def _load_agents_sdk() -> SandboxSDK:
         from agents import Runner  # ty: ignore[unresolved-import]
         from docker import from_env as docker_from_env
     except ImportError as exc:
-        raise RuntimeError("DockerSandboxAgentRuntime requires `nemo-evaluator-sdk[agent-runtimes]`") from exc
+        # Audience split is in the error text: SDK extras are not propagated into the
+        # vendored nemo_platform.beta.evaluator mirror.
+        raise RuntimeError(
+            "DockerSandboxAgentRuntime requires the openai-agents[docker] Python packages. "
+            "Standalone SDK: pip install 'nemo-evaluator-sdk[agent-runtimes]'. "
+            "Vendored nemo-platform.beta.evaluator (no SDK extras): "
+            "pip install 'openai-agents[docker]'"
+        ) from exc
 
     return SandboxSDK(
         Runner=Runner,
diff --git a/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/agent_inference.py b/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/agent_inference.py
index 2e759fe3e2..181b29bbdf 100644
--- a/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/agent_inference.py
+++ b/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/agent_inference.py
@@ -27,7 +27,6 @@
 
 import httpx
 from httpx import Timeout
-from jsonpath_ng import parse as jsonpath_parse
 from pydantic import BaseModel, ConfigDict, Field
 
 from nemo_platform.beta.evaluator.agent_stream_translation import (
@@ -788,6 +787,10 @@ def _extract_jsonpath(
     required: bool = True,
 ) -> Any:
     """Extract a value from data using a JSONPath expression."""
+    # Imported here rather than at module scope: this is jsonpath_ng's only use in the module, and
+    # the module sits on the agent_eval run-time path via agent_eval/evaluator.py.
+    from jsonpath_ng import parse as jsonpath_parse
+
     expr = jsonpath_parse(path)
     matches = expr.find(data)
     if not matches:
diff --git a/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/metrics/aggregation.py b/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/metrics/aggregation.py
index 3df0f7a395..22698fab8f 100644
--- a/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/metrics/aggregation.py
+++ b/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/metrics/aggregation.py
@@ -3,10 +3,12 @@
 
 """Aggregation data structures and computations for metric results."""
 
+from __future__ import annotations
+
 import math
 from collections import OrderedDict, defaultdict
 from collections.abc import Mapping, Sequence
-from typing import Protocol, cast, runtime_checkable
+from typing import TYPE_CHECKING, Protocol, cast, runtime_checkable
 
 from nemo_platform.beta.evaluator.metrics.protocol import (
     BooleanValue,
@@ -28,7 +30,11 @@
     RubricScoreStat,
     ScoreStats,
 )
-from nemo_platform.beta.evaluator.values.scores import RubricScore, Score
+
+if TYPE_CHECKING:
+    # Importing the score configuration module loads jsonschema. Aggregation's lightweight helpers
+    # (notably compute_percentiles) do not need it, so keep this typing-only edge deferred.
+    from nemo_platform.beta.evaluator.values.scores import Score
 
 
 def is_aggregateable_output_spec(output_spec: MetricOutputSpec) -> bool:
@@ -445,6 +451,8 @@ def aggregate_metrics(
 
 def rubric_definitions_from_scores(scores: Sequence[Score]) -> dict[str, list[RubricScoreStat]]:
     """Return declared rubric buckets keyed by score name."""
+    from nemo_platform.beta.evaluator.values.scores import RubricScore
+
     definitions: dict[str, list[RubricScoreStat]] = {}
     for score in scores:
         if not isinstance(score, RubricScore):
diff --git a/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/values/__init__.py b/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/values/__init__.py
index db93c9807b..18efea4ac9 100644
--- a/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/values/__init__.py
+++ b/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/values/__init__.py
@@ -1,121 +1,241 @@
 # SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
-"""Public value types for evaluator SDK runtime."""
+"""Public value types for evaluator SDK runtime.
+
+The public interface resolves lazily (PEP 562), for the same reason the package root does:
+every ``from nemo_platform.beta.evaluator.values.X import ...`` runs this barrel first, so
+eagerly re-exporting all 97 names dragged ``.datasets``/``.results`` (pyarrow, numpy) and
+``.metrics``/``.scores`` (jsonschema, jinja2) into ``agent_eval``, which uses none of them.
+Measured: 485 modules and +57 MB RSS for ``import agent_eval.runtimes.harbor_runtime`` before,
+300 modules and pydantic alone after.
+
+Add a new re-export to ``_LAZY_ATTRS``, the ``TYPE_CHECKING`` block and ``__all__`` — never as a
+module-level import. ``tests/test_lazy_public_api.py`` locks the boundary in.
+"""
 
 # ruff: noqa: I001 - the vendored SDK mirror uses different import-order settings.
 
-from nemo_platform.beta.evaluator.values.agents import (
-    Agent,
-    AgentBase,
-    GenericAgent,
-    NatAgentConfig,
-    NemoAgentToolkitAgent,
-)
-from nemo_platform.beta.evaluator.values.atif import (
-    FinalMetrics,
-    Metrics,
-    Observation,
-    ObservationResult,
-    Step,
-    ToolCall,
-    Trajectory,
-)
-from nemo_platform.beta.evaluator.values.common import SecretRef, SupportedJobTypes
-from nemo_platform.beta.evaluator.values.dataset_schemas import (
-    FieldMapping,
-    InputSchema,
-)
-from nemo_platform.beta.evaluator.values.datasets import DatasetInput, DatasetRows
-from nemo_platform.beta.evaluator.values.evidence import (
-    CandidateEvidence,
-    CommandResult,
-    EvidenceDescriptor,
-    FilesystemDiff,
-    FilesystemEntry,
-    LocalFilesystemEvidence,
-    LogHandle,
-    TraceHandle,
-    WellKnownEvidenceKey,
-    parse_atif,
-)
-from nemo_platform.beta.evaluator.values.metrics import (
-    BLEU,
-    F1,
-    ROUGE,
-    AgentGoalAccuracy,
-    AnswerAccuracy,
-    ContextEntityRecall,
-    ContextPrecision,
-    ContextRecall,
-    ContextRelevance,
-    ExactMatch,
-    Faithfulness,
-    LLMJudge,
-    MetricBase,
-    NemoAgentToolkitRemote,
-    NoiseSensitivity,
-    NumberCheck,
-    Remote,
-    ResponseGroundedness,
-    ResponseRelevancy,
-    StringCheck,
-    ToolCallAccuracy,
-    ToolCalling,
-    TopicAdherence,
-    TunableRagEvaluator,
-)
-from nemo_platform.beta.evaluator.values.models import Model, ModelRef, ReasoningParams
-from nemo_platform.beta.evaluator.values.params import (
-    InferenceParams,
-    RunConfig,
-    RunConfigOnline,
-    RunConfigOnlineModel,
-)
-from nemo_platform.beta.evaluator.values.protocol import (
-    BooleanValue,
-    CandidateOutput,
-    ContinuousScore,
-    DatasetRow,
-    DiscreteScore,
-    Label,
-    MetricDescriptor,
-    MetricDiagnostic,
-    MetricInput,
-    MetricOutput,
-    MetricOutputSpec,
-    MetricResult,
-    MetricTypeName,
-)
-from nemo_platform.beta.evaluator.values.results import (
-    AggregatedMetricResult,
-    AggregateFieldName,
-    AggregateRangeScore,
-    AggregateRubricScore,
-    AggregateScore,
-    AggregateScoreBase,
-    DefaultAggregateFieldName,
-    EvaluationResult,
-    Histogram,
-    HistogramBin,
-    MetricScore,
-    Percentiles,
-    RowScore,
-    RubricScoreStat,
-    RubricScoreValue,
-    SampleResult,
-    ScoreStats,
-)
-from nemo_platform.beta.evaluator.values.scores import (
-    JSONScoreParser,
-    RangeScore,
-    RegexScoreParser,
-    RemoteScore,
-    Rubric,
-    RubricScore,
-    Score,
-    score_discriminator,
-)
+from importlib import import_module as _import_module
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from nemo_platform.beta.evaluator.values.agents import (
+        Agent,
+        AgentBase,
+        GenericAgent,
+        NatAgentConfig,
+        NemoAgentToolkitAgent,
+    )
+    from nemo_platform.beta.evaluator.values.atif import (
+        FinalMetrics,
+        Metrics,
+        Observation,
+        ObservationResult,
+        Step,
+        ToolCall,
+        Trajectory,
+    )
+    from nemo_platform.beta.evaluator.values.common import SecretRef, SupportedJobTypes
+    from nemo_platform.beta.evaluator.values.dataset_schemas import (
+        FieldMapping,
+        InputSchema,
+    )
+    from nemo_platform.beta.evaluator.values.datasets import DatasetInput, DatasetRows
+    from nemo_platform.beta.evaluator.values.evidence import (
+        CandidateEvidence,
+        CommandResult,
+        EvidenceDescriptor,
+        FilesystemDiff,
+        FilesystemEntry,
+        LocalFilesystemEvidence,
+        LogHandle,
+        TraceHandle,
+        WellKnownEvidenceKey,
+        parse_atif,
+    )
+    from nemo_platform.beta.evaluator.values.metrics import (
+        BLEU,
+        F1,
+        ROUGE,
+        AgentGoalAccuracy,
+        AnswerAccuracy,
+        ContextEntityRecall,
+        ContextPrecision,
+        ContextRecall,
+        ContextRelevance,
+        ExactMatch,
+        Faithfulness,
+        LLMJudge,
+        MetricBase,
+        NemoAgentToolkitRemote,
+        NoiseSensitivity,
+        NumberCheck,
+        Remote,
+        ResponseGroundedness,
+        ResponseRelevancy,
+        StringCheck,
+        ToolCallAccuracy,
+        ToolCalling,
+        TopicAdherence,
+        TunableRagEvaluator,
+    )
+    from nemo_platform.beta.evaluator.values.models import Model, ModelRef, ReasoningParams
+    from nemo_platform.beta.evaluator.values.params import (
+        InferenceParams,
+        RunConfig,
+        RunConfigOnline,
+        RunConfigOnlineModel,
+    )
+    from nemo_platform.beta.evaluator.values.protocol import (
+        BooleanValue,
+        CandidateOutput,
+        ContinuousScore,
+        DatasetRow,
+        DiscreteScore,
+        Label,
+        MetricDescriptor,
+        MetricDiagnostic,
+        MetricInput,
+        MetricOutput,
+        MetricOutputSpec,
+        MetricResult,
+        MetricTypeName,
+    )
+    from nemo_platform.beta.evaluator.values.results import (
+        AggregatedMetricResult,
+        AggregateFieldName,
+        AggregateRangeScore,
+        AggregateRubricScore,
+        AggregateScore,
+        AggregateScoreBase,
+        DefaultAggregateFieldName,
+        EvaluationResult,
+        Histogram,
+        HistogramBin,
+        MetricScore,
+        Percentiles,
+        RowScore,
+        RubricScoreStat,
+        RubricScoreValue,
+        SampleResult,
+        ScoreStats,
+    )
+    from nemo_platform.beta.evaluator.values.scores import (
+        JSONScoreParser,
+        RangeScore,
+        RegexScoreParser,
+        RemoteScore,
+        Rubric,
+        RubricScore,
+        Score,
+        score_discriminator,
+    )
+
+
+# Re-exported name -> the submodule that defines it, relative to this package. Relative on
+# purpose: the vendoring tool mirrors this file into nemo_platform.beta.evaluator by rewriting
+# module paths, and a relative name has nothing to rewrite, so the mirror is correct by
+# construction. Mirrors the TYPE_CHECKING block above, in the same order.
+_LAZY_ATTRS: dict[str, str] = {
+    "Agent": ".agents",
+    "AgentBase": ".agents",
+    "GenericAgent": ".agents",
+    "NatAgentConfig": ".agents",
+    "NemoAgentToolkitAgent": ".agents",
+    "FinalMetrics": ".atif",
+    "Metrics": ".atif",
+    "Observation": ".atif",
+    "ObservationResult": ".atif",
+    "Step": ".atif",
+    "ToolCall": ".atif",
+    "Trajectory": ".atif",
+    "SecretRef": ".common",
+    "SupportedJobTypes": ".common",
+    "FieldMapping": ".dataset_schemas",
+    "InputSchema": ".dataset_schemas",
+    "DatasetInput": ".datasets",
+    "DatasetRows": ".datasets",
+    "CandidateEvidence": ".evidence",
+    "CommandResult": ".evidence",
+    "EvidenceDescriptor": ".evidence",
+    "FilesystemDiff": ".evidence",
+    "FilesystemEntry": ".evidence",
+    "LocalFilesystemEvidence": ".evidence",
+    "LogHandle": ".evidence",
+    "TraceHandle": ".evidence",
+    "WellKnownEvidenceKey": ".evidence",
+    "parse_atif": ".evidence",
+    "BLEU": ".metrics",
+    "F1": ".metrics",
+    "ROUGE": ".metrics",
+    "AgentGoalAccuracy": ".metrics",
+    "AnswerAccuracy": ".metrics",
+    "ContextEntityRecall": ".metrics",
+    "ContextPrecision": ".metrics",
+    "ContextRecall": ".metrics",
+    "ContextRelevance": ".metrics",
+    "ExactMatch": ".metrics",
+    "Faithfulness": ".metrics",
+    "LLMJudge": ".metrics",
+    "MetricBase": ".metrics",
+    "NemoAgentToolkitRemote": ".metrics",
+    "NoiseSensitivity": ".metrics",
+    "NumberCheck": ".metrics",
+    "Remote": ".metrics",
+    "ResponseGroundedness": ".metrics",
+    "ResponseRelevancy": ".metrics",
+    "StringCheck": ".metrics",
+    "ToolCallAccuracy": ".metrics",
+    "ToolCalling": ".metrics",
+    "TopicAdherence": ".metrics",
+    "TunableRagEvaluator": ".metrics",
+    "Model": ".models",
+    "ModelRef": ".models",
+    "ReasoningParams": ".models",
+    "InferenceParams": ".params",
+    "RunConfig": ".params",
+    "RunConfigOnline": ".params",
+    "RunConfigOnlineModel": ".params",
+    "BooleanValue": ".protocol",
+    "CandidateOutput": ".protocol",
+    "ContinuousScore": ".protocol",
+    "DatasetRow": ".protocol",
+    "DiscreteScore": ".protocol",
+    "Label": ".protocol",
+    "MetricDescriptor": ".protocol",
+    "MetricDiagnostic": ".protocol",
+    "MetricInput": ".protocol",
+    "MetricOutput": ".protocol",
+    "MetricOutputSpec": ".protocol",
+    "MetricResult": ".protocol",
+    "MetricTypeName": ".protocol",
+    "AggregatedMetricResult": ".results",
+    "AggregateFieldName": ".results",
+    "AggregateRangeScore": ".results",
+    "AggregateRubricScore": ".results",
+    "AggregateScore": ".results",
+    "AggregateScoreBase": ".results",
+    "DefaultAggregateFieldName": ".results",
+    "EvaluationResult": ".results",
+    "Histogram": ".results",
+    "HistogramBin": ".results",
+    "MetricScore": ".results",
+    "Percentiles": ".results",
+    "RowScore": ".results",
+    "RubricScoreStat": ".results",
+    "RubricScoreValue": ".results",
+    "SampleResult": ".results",
+    "ScoreStats": ".results",
+    "JSONScoreParser": ".scores",
+    "RangeScore": ".scores",
+    "RegexScoreParser": ".scores",
+    "RemoteScore": ".scores",
+    "Rubric": ".scores",
+    "RubricScore": ".scores",
+    "Score": ".scores",
+    "score_discriminator": ".scores",
+}
 
 __all__ = [
     "Agent",
@@ -217,3 +337,30 @@
     "TopicAdherence",
     "TunableRagEvaluator",
 ]
+
+
+def __getattr__(name: str) -> object:
+    """Import the submodule that defines ``name`` on first access (PEP 562).
+
+    An *unknown* name raises ``AttributeError``, which is required: ``from pkg import sub`` only
+    falls back to importing a submodule when attribute lookup raises ``AttributeError``.
+
+    A *known* name whose submodule fails to import propagates that ``ImportError`` unchanged, so
+    the real cause is not hidden behind a bogus "no attribute". The consequence is that
+    ``hasattr`` raises rather than returning ``False`` when a name's dependencies are missing;
+    catch ``ImportError`` around the access, or test membership against ``__all__``.
+    """
+    submodule = _LAZY_ATTRS.get(name)
+    if submodule is None:
+        raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
+    value = getattr(_import_module(submodule, __name__), name)
+    globals()[name] = value  # cache, so later lookups skip __getattr__ entirely
+    return value
+
+
+def __dir__() -> list[str]:
+    # The declared surface plus any submodule the caller has already imported. Machinery is
+    # imported under a leading underscore so the filter keeps it out of autocomplete without a
+    # denylist; ``TYPE_CHECKING`` is the one exception, unaliased so type checkers recognise it.
+    public = {name for name in globals() if not name.startswith("_")} - {"TYPE_CHECKING"}
+    return sorted(set(__all__) | public)
diff --git a/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/values/results.py b/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/values/results.py
index 52551f88c8..05220d061d 100644
--- a/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/values/results.py
+++ b/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/values/results.py
@@ -9,11 +9,13 @@
 import math
 from collections.abc import Mapping
 from difflib import get_close_matches
-from typing import Any, Literal, Self
+from typing import TYPE_CHECKING, Any, Literal, Self
 
-import pyarrow as pa
 from pydantic import BaseModel, ConfigDict, Field, field_serializer, field_validator, model_serializer
 
+if TYPE_CHECKING:
+    import pyarrow as pa
+
 from nemo_platform.beta.evaluator.values.protocol import MetricDiagnostic, MetricOutput, MetricResult
 
 ResultView = Literal["rows", "aggregate"]
@@ -792,6 +794,12 @@ def to_table(self, view: ResultView = "rows") -> pa.Table:
         Returns:
             Table built from ``to_records(view=view)``.
         """
+        # Imported here rather than at module scope: pyarrow (plus its numpy tail) costs ~31 MB
+        # RSS and 223 modules, this is its only runtime use in the module, and the module is on
+        # the agent_eval import path, which never calls this method. The return annotation is a
+        # string already (`from __future__ import annotations`), so it needs no import.
+        import pyarrow as pa
+
         return pa.Table.from_pylist(self.to_records(view=view))
 
     def to_pandas(self, view: ResultView = "rows"):
diff --git a/uv.lock b/uv.lock
index 055cd5b1ad..83f66a183a 100644
--- a/uv.lock
+++ b/uv.lock
@@ -4534,6 +4534,7 @@ name = "nemo-evaluator-sdk"
 version = "0.0.0"
 source = { editable = "packages/nemo_evaluator_sdk" }
 dependencies = [
+    { name = "httpx", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "jinja2", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "jsonpath-ng", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "jsonschema", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
@@ -4574,6 +4575,7 @@ dev = [
 [package.metadata]
 requires-dist = [
     { name = "harbor", marker = "python_full_version >= '3.12' and extra == 'harbor'", specifier = ">=0.16.1" },
+    { name = "httpx", specifier = ">=0.27.0,<1" },
     { name = "jinja2", specifier = ">=3.1.6" },
     { name = "jsonpath-ng", specifier = ">=1.7.0" },
     { name = "jsonschema", specifier = ">=4.23.0" },
@@ -5192,6 +5194,7 @@ nemo-evaluator-plugin = [
     { name = "typer", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
 ]
 nemo-evaluator-sdk = [
+    { name = "httpx", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "jinja2", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "jsonpath-ng", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "jsonschema", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
@@ -5621,6 +5624,7 @@ requires-dist = [
     { name = "httpx", marker = "extra == 'nemo-anonymizer-plugin'", specifier = ">=0.27" },
     { name = "httpx", marker = "extra == 'nemo-auditor-plugin'", specifier = ">=0.27" },
     { name = "httpx", marker = "extra == 'nemo-data-designer-plugin'", specifier = ">=0.27" },
+    { name = "httpx", marker = "extra == 'nemo-evaluator-sdk'", specifier = ">=0.27.0,<1" },
     { name = "httpx", marker = "extra == 'nemo-platform-sdk'", specifier = ">=0.23.0,<1" },
     { name = "httpx", marker = "extra == 'nemo-safe-synthesizer-plugin'", specifier = ">=0.27.2" },
     { name = "httpx", marker = "extra == 'plugins'", specifier = ">=0.27" },
@@ -6206,6 +6210,7 @@ aiohttp = [
     { name = "httpx-aiohttp", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
 ]
 nemo-evaluator-sdk = [
+    { name = "httpx", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "jinja2", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "jsonpath-ng", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "jsonschema", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
@@ -6248,6 +6253,7 @@ requires-dist = [
     { name = "docker", specifier = ">=7.0.0" },
     { name = "fsspec", specifier = ">=2023.1.0" },
     { name = "httpx", specifier = ">=0.23.0,<1" },
+    { name = "httpx", marker = "extra == 'nemo-evaluator-sdk'", specifier = ">=0.27.0,<1" },
     { name = "httpx-aiohttp", marker = "extra == 'aiohttp'", specifier = ">=0.1.9" },
     { name = "jinja2", marker = "extra == 'nemo-evaluator-sdk'", specifier = ">=3.1.6" },
     { name = "jsonpath-ng", marker = "extra == 'nemo-evaluator-sdk'", specifier = ">=1.7.0" },