diff --git a/tests/workloads/aiconfig/test_aiconfigurator_sdk_contract.py b/tests/workloads/aiconfig/test_aiconfigurator_sdk_contract.py
new file mode 100644
index 000000000..de36f09fb
--- /dev/null
+++ b/tests/workloads/aiconfig/test_aiconfigurator_sdk_contract.py
@@ -0,0 +1,217 @@
+# SPDX-FileCopyrightText: NVIDIA CORPORATION & AFFILIATES
+# Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Compatibility canary between ``runtime/predictor.py`` and the pinned aiconfigurator SDK.
+
+``runtime/predictor.py`` is our adapter onto ``aiconfigurator.sdk.*``; the version
+is pinned by ``AiconfiguratorTestDefinition.python_environment`` (``aiconfigurator~=0.5.0``).
+If a pinned-version bump renames/removes a SDK symbol the adapter relies on, the
+adapter breaks at run time and DSE silently stops producing metrics -- with nothing
+to warn us, because the rest of the aiconfig tests mock or never touch the SDK.
+
+This canary closes that gap *without* the heavy (~559 MB, torch) install: it fetches
+only the pinned source with ``uv pip install --no-deps`` (seconds, no torch) and
+audits it with the ``ast`` module -- no imports, so the torch-importing SDK modules
+parse fine. It asserts the exact surface ``predictor.py`` consumes still exists.
+
+Scope: this catches input-side drift (renamed/removed fields, enum members, funcs).
+It does NOT catch a changed *return* shape (e.g. a renamed summary-df column); that
+would need a real prediction run and is out of scope for a lightweight canary.
+
+Marked ``ci_only``: it requires network/``uv`` to fetch the pinned source, so it runs
+in CI's dedicated ``-m ci_only`` step rather than the hermetic unit suite. Keep the
+contract constants below in lockstep with ``runtime/predictor.py``.
+"""
+
+from __future__ import annotations
+
+import ast
+import functools
+import shutil
+import subprocess
+import tempfile
+from pathlib import Path
+
+import pytest
+
+from cloudai.workloads.aiconfig import AiconfiguratorCmdArgs, AiconfiguratorTestDefinition
+from cloudai.workloads.aiconfig.aiconfigurator import Agg
+
+pytestmark = pytest.mark.ci_only
+
+# --- The SDK surface runtime/predictor.py depends on (keep in lockstep with it) ---
+
+# aic_config.ModelConfig(...) / aic_config.RuntimeConfig(...) keyword arguments.
+_MODEL_CONFIG_FIELDS = [
+    "tp_size",
+    "pp_size",
+    "attention_dp_size",
+    "moe_tp_size",
+    "moe_ep_size",
+    "gemm_quant_mode",
+    "moe_quant_mode",
+    "kvcache_quant_mode",
+    "fmha_quant_mode",
+    "comm_quant_mode",
+    "nextn",
+    "nextn_accept_rates",
+    "overwrite_num_layers",
+]
+_RUNTIME_CONFIG_FIELDS = ["batch_size", "isl", "osl"]
+
+# common.<Enum>[<member>] resolved by predictor._to_enum from the CLI defaults.
+_QUANT_ENUM_MEMBERS = {
+    "GEMMQuantMode": "fp8_block",
+    "MoEQuantMode": "fp8",
+    "KVCacheQuantMode": "fp8",
+    "FMHAQuantMode": "fp8",
+    "CommQuantMode": "half",
+}
+
+# Free functions predictor.py calls, keyed by the sdk-relative source file.
+_FREE_FUNCTIONS = {
+    "perf_database.py": "get_database",
+    "models.py": "get_model",
+    "backends/factory.py": "get_backend",
+}
+_INFERENCE_SESSION_METHODS = ["run_static"]
+
+
+def _pinned_requirement() -> str:
+    """The aiconfigurator requirement string the workload pins (single source of truth)."""
+    tdef = AiconfiguratorTestDefinition(
+        name="aiconfig",
+        description="contract",
+        test_template_name="Aiconfigurator",
+        cmd_args=AiconfiguratorCmdArgs(
+            model_name="LLAMA3.1_70B", system="h200_sxm", isl=1, osl=1, agg=Agg(batch_size=1, ctx_tokens=1)
+        ),
+    )
+    reqs = tdef.python_environment.requirements
+    assert len(reqs) == 1, f"expected a single aiconfigurator requirement, got {reqs}"
+    return reqs[0]
+
+
+@functools.lru_cache(maxsize=1)
+def _fetch_pinned_sdk_src() -> Path:
+    """``uv pip install --no-deps`` the pinned aiconfigurator into a temp dir; return its ``sdk`` path.
+
+    ``--no-deps`` skips the entire torch/cuda tree, so this is seconds and tiny. We
+    never import the package (its compute modules need torch); callers AST-parse it.
+    """
+    uv = shutil.which("uv")
+    if uv is None:
+        raise RuntimeError("uv is not available to fetch the pinned aiconfigurator source")
+
+    target = Path(tempfile.mkdtemp(prefix="aiconfigurator_sdk_"))
+    proc = subprocess.run(
+        [uv, "pip", "install", "--no-deps", "--target", str(target), _pinned_requirement()],
+        capture_output=True,
+        text=True,
+    )
+    if proc.returncode != 0:
+        raise RuntimeError(f"failed to fetch aiconfigurator --no-deps: {proc.stderr.strip()[:400]}")
+
+    sdk = target / "aiconfigurator" / "sdk"
+    if not sdk.is_dir():
+        raise RuntimeError(f"aiconfigurator.sdk not found under fetched package at {sdk}")
+    return sdk
+
+
+def _sdk_src() -> Path:
+    try:
+        return _fetch_pinned_sdk_src()
+    except RuntimeError as exc:
+        pytest.skip(str(exc))
+
+
+def _parse(sdk: Path, rel: str) -> ast.Module:
+    return ast.parse((sdk / rel).read_text(encoding="utf-8"))
+
+
+def _find_class(tree: ast.Module, name: str) -> ast.ClassDef | None:
+    return next((n for n in ast.walk(tree) if isinstance(n, ast.ClassDef) and n.name == name), None)
+
+
+def _defines_function(tree: ast.Module, name: str) -> bool:
+    return any(isinstance(n, (ast.FunctionDef, ast.AsyncFunctionDef)) and n.name == name for n in ast.walk(tree))
+
+
+def _accepted_kwargs(cls: ast.ClassDef) -> tuple[set[str], bool]:
+    """Constructor-accepted kwargs: explicit ``__init__`` params, else annotated (dataclass/pydantic) fields."""
+    init = next(
+        (b for b in cls.body if isinstance(b, (ast.FunctionDef, ast.AsyncFunctionDef)) and b.name == "__init__"),
+        None,
+    )
+    if init is not None:
+        args = init.args
+        names = {p.arg for p in (args.posonlyargs + args.args + args.kwonlyargs) if p.arg != "self"}
+        return names, args.kwarg is not None
+    fields = {b.target.id for b in cls.body if isinstance(b, ast.AnnAssign) and isinstance(b.target, ast.Name)}
+    return fields, False
+
+
+def _enum_members(cls: ast.ClassDef) -> set[str]:
+    members: set[str] = set()
+    for stmt in cls.body:
+        if isinstance(stmt, ast.Assign):
+            members.update(t.id for t in stmt.targets if isinstance(t, ast.Name))
+        elif isinstance(stmt, ast.AnnAssign) and isinstance(stmt.target, ast.Name):
+            members.add(stmt.target.id)
+    return members
+
+
+@pytest.mark.parametrize(
+    "cls_name, expected, rel",
+    [
+        ("ModelConfig", _MODEL_CONFIG_FIELDS, "config.py"),
+        ("RuntimeConfig", _RUNTIME_CONFIG_FIELDS, "config.py"),
+    ],
+)
+def test_config_classes_accept_required_kwargs(cls_name: str, expected: list[str], rel: str) -> None:
+    cls = _find_class(_parse(_sdk_src(), rel), cls_name)
+    assert cls is not None, f"aiconfigurator.sdk.config.{cls_name} no longer exists"
+
+    accepted, has_var_kwargs = _accepted_kwargs(cls)
+    if has_var_kwargs:
+        return
+    missing = sorted(set(expected) - accepted)
+    assert not missing, f"{cls_name} no longer accepts {missing}; predictor.py is incompatible with the pinned SDK"
+
+
+def test_quant_enums_expose_used_members() -> None:
+    tree = _parse(_sdk_src(), "common.py")
+    for enum_name, member in _QUANT_ENUM_MEMBERS.items():
+        cls = _find_class(tree, enum_name)
+        assert cls is not None, f"aiconfigurator.sdk.common.{enum_name} no longer exists"
+        members = _enum_members(cls)
+        assert member in members, f"{enum_name} no longer defines member {member!r} that predictor.py relies on"
+
+
+@pytest.mark.parametrize("rel, func", sorted(_FREE_FUNCTIONS.items()))
+def test_sdk_free_functions_exist(rel: str, func: str) -> None:
+    assert _defines_function(_parse(_sdk_src(), rel), func), (
+        f"aiconfigurator.sdk.{rel.replace('/', '.').removesuffix('.py')}.{func} no longer exists"
+    )
+
+
+def test_inference_session_exposes_used_methods() -> None:
+    cls = _find_class(_parse(_sdk_src(), "inference_session.py"), "InferenceSession")
+    assert cls is not None, "aiconfigurator.sdk.inference_session.InferenceSession no longer exists"
+
+    methods = {b.name for b in cls.body if isinstance(b, (ast.FunctionDef, ast.AsyncFunctionDef))}
+    missing = sorted(set(_INFERENCE_SESSION_METHODS) - methods)
+    assert not missing, f"InferenceSession no longer defines {missing} that predictor.py calls"