diff --git a/src/bedit_core/serialization/__init__.py b/src/bedit_core/serialization/__init__.py
new file mode 100644
index 0000000..d00b159
--- /dev/null
+++ b/src/bedit_core/serialization/__init__.py
@@ -0,0 +1,57 @@
+"""Public file loading and saving API for bedit documents."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from bedit_core.models import Document
+
+from . import beb_codec, json_codec
+from .migrations import migrate_document_layout, migrate_file_layout
+from .schema import SerializationError, document_from_data, document_to_data
+
+__all__ = [
+    "BEB_FILE_FORMAT_VERSION",
+    "JSON_FILE_FORMAT_VERSION",
+    "SerializationError",
+    "load",
+    "save",
+]
+
+BEB_FILE_FORMAT_VERSION = beb_codec.FILE_FORMAT_VERSION
+JSON_FILE_FORMAT_VERSION = json_codec.FILE_FORMAT_VERSION
+
+
+def load(path: str | Path) -> Document:
+    """Load a document from ``path``.
+
+    The filename extension selects the JSON or BEB codec. File-layout
+    migrations and then document-layout migrations are applied before the
+    raw schema is converted to a :class:`~bedit_core.models.Document`.
+    """
+    file_path = Path(path)
+    codec = _codec(file_path)
+    file_version, file_data = codec.load_data(file_path)
+    document_data = migrate_file_layout(
+        file_data,
+        codec.FORMAT_NAME,
+        file_version,
+    )
+    return document_from_data(migrate_document_layout(document_data))
+
+
+def save(document: Document, path: str | Path) -> None:
+    """Save ``document`` to a JSON or BEB file selected by ``path``'s suffix."""
+    file_path = Path(path)
+    codec = _codec(file_path)
+    codec.save_data(document_to_data(document), file_path)
+
+
+def _codec(path: Path) -> Any:
+    """Return the codec module associated with a supported filename suffix."""
+    suffix = path.suffix.lower()
+    if suffix == ".json":
+        return json_codec
+    if suffix == ".beb":
+        return beb_codec
+    raise ValueError(f"unsupported file extension {path.suffix!r}; expected '.json' or '.beb'")
diff --git a/src/bedit_core/serialization/beb_codec.py b/src/bedit_core/serialization/beb_codec.py
new file mode 100644
index 0000000..7d4802d
--- /dev/null
+++ b/src/bedit_core/serialization/beb_codec.py
@@ -0,0 +1,75 @@
+"""Versioned binary BEB codec.
+
+The uncompressed header contains the magic bytes and codec version, allowing
+the correct decoder to be selected even if later versions change compression
+or serialization methods.
+"""
+
+from __future__ import annotations
+
+import zlib
+from pathlib import Path
+from typing import Any, Mapping
+
+import msgpack
+
+FORMAT_NAME = "beb"
+FILE_FORMAT_VERSION = 1
+BEB_MAGIC = b"BEB\x00"
+_VERSION_SIZE = 4
+
+
+def load_data(path: Path) -> tuple[int, Mapping[str, Any]]:
+    """Read a BEB header and decode its raw document mapping.
+
+    An unknown version is returned with an empty mapping so the migration
+    layer can produce the standard unsupported-version error without trying
+    an incompatible decoder.
+    """
+    try:
+        payload = path.read_bytes()
+        if not payload.startswith(BEB_MAGIC):
+            raise ValueError("missing BEB file header")
+        header_end = len(BEB_MAGIC) + _VERSION_SIZE
+        if len(payload) < header_end:
+            raise ValueError("truncated BEB file header")
+        version = int.from_bytes(payload[len(BEB_MAGIC) : header_end], "big")
+        encoded = payload[header_end:]
+        decoder = _DECODERS.get(version)
+        if decoder is None:
+            # The version can be inspected without trying to decompress or
+            # deserialize using the wrong algorithm.
+            return version, {}
+        data = decoder(encoded)
+    except (OSError, ValueError, zlib.error, msgpack.exceptions.MsgpackException) as exc:
+        raise ValueError(f"could not read BEB document {path}: {exc}") from exc
+    if not isinstance(data, Mapping):
+        raise ValueError(f"BEB document {path} must contain a map at its root")
+    return version, data
+
+
+def save_data(data: Mapping[str, Any], path: Path) -> None:
+    """Encode raw document data using the current BEB version and write it."""
+    encoded = _encode_v1(data)
+    version = FILE_FORMAT_VERSION.to_bytes(_VERSION_SIZE, "big")
+    path.write_bytes(BEB_MAGIC + version + encoded)
+
+
+def _encode_v1(data: Mapping[str, Any]) -> bytes:
+    """Encode BEB version 1 as zlib-compressed MessagePack."""
+    return zlib.compress(msgpack.packb(dict(data), use_bin_type=True))
+
+
+def _decode_v1(payload: bytes) -> Any:
+    """Decode a zlib-compressed MessagePack BEB version 1 payload."""
+    return msgpack.unpackb(
+        zlib.decompress(payload),
+        raw=False,
+        strict_map_key=False,
+    )
+
+
+# Keep old decoders when adding a new BEB encoding version.
+_DECODERS = {
+    1: _decode_v1,
+}
diff --git a/src/bedit_core/serialization/json_codec.py b/src/bedit_core/serialization/json_codec.py
new file mode 100644
index 0000000..82e8d71
--- /dev/null
+++ b/src/bedit_core/serialization/json_codec.py
@@ -0,0 +1,36 @@
+"""Versioned UTF-8 JSON codec used by the serialization API."""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from typing import Any, Mapping
+
+FORMAT_NAME = "json"
+FILE_FORMAT_VERSION = 1
+
+
+def load_data(path: Path) -> tuple[int, Mapping[str, Any]]:
+    """Read JSON and return its file-layout version and raw root object.
+
+    JSON files without ``file_format_version`` are interpreted as version 1.
+    Schema conversion and migrations are intentionally handled by higher
+    layers.
+    """
+    try:
+        with path.open("r", encoding="utf-8") as stream:
+            data = json.load(stream)
+    except (OSError, UnicodeError, json.JSONDecodeError) as exc:
+        raise ValueError(f"could not read JSON document {path}: {exc}") from exc
+    if not isinstance(data, Mapping):
+        raise ValueError(f"JSON document {path} must contain an object at its root")
+    version = data.get("file_format_version", 1)
+    return version, data
+
+
+def save_data(data: Mapping[str, Any], path: Path) -> None:
+    """Write raw document data as indented UTF-8 JSON at the current version."""
+    output = {"file_format_version": FILE_FORMAT_VERSION, **data}
+    with path.open("w", encoding="utf-8") as stream:
+        json.dump(output, stream, ensure_ascii=False, indent=2)
+        stream.write("\n")
diff --git a/src/bedit_core/serialization/migrations.py b/src/bedit_core/serialization/migrations.py
new file mode 100644
index 0000000..4712db3
--- /dev/null
+++ b/src/bedit_core/serialization/migrations.py
@@ -0,0 +1,104 @@
+"""Registries and runners for file-layout and document-layout migrations.
+
+Register a migration under version ``N`` to convert raw data from ``N`` to
+``N + 1``. JSON and BEB file migrations are independent; document migrations
+run after file migrations and are shared by both codecs.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Callable, Mapping
+from copy import deepcopy
+from typing import Any
+
+from . import beb_codec, json_codec
+from .schema import SerializationError
+
+DOCUMENT_FORMAT_VERSION = 1
+
+Migration = Callable[[dict[str, Any]], dict[str, Any]]
+
+# A migration registered under N converts version N to N + 1. File migrations
+# are separate because JSON and BEB may evolve independently.
+FILE_FORMAT_VERSIONS = {
+    json_codec.FORMAT_NAME: json_codec.FILE_FORMAT_VERSION,
+    beb_codec.FORMAT_NAME: beb_codec.FILE_FORMAT_VERSION,
+}
+FILE_MIGRATIONS: dict[str, dict[int, Migration]] = {
+    json_codec.FORMAT_NAME: {},
+    beb_codec.FORMAT_NAME: {},
+}
+DOCUMENT_MIGRATIONS: dict[int, Migration] = {}
+
+
+def migrate_file_layout(
+    data: Mapping[str, Any],
+    file_format: str,
+    version: int,
+) -> dict[str, Any]:
+    """Migrate codec-specific raw data to that codec's current file layout.
+
+    The input is copied, so migrations cannot mutate the codec's decoded
+    object. ``file_format_version`` is removed before document migration.
+    """
+    migrated = deepcopy(dict(data))
+    version = _version(version, f"{file_format} file format version")
+    try:
+        target = FILE_FORMAT_VERSIONS[file_format]
+        migrations = FILE_MIGRATIONS[file_format]
+    except KeyError:
+        raise SerializationError(f"unknown file format {file_format!r}") from None
+    migrated = _apply(
+        migrated,
+        version,
+        target,
+        migrations,
+        f"{file_format} file",
+    )
+    migrated.pop("file_format_version", None)
+    return migrated
+
+
+def migrate_document_layout(data: Mapping[str, Any]) -> dict[str, Any]:
+    """Migrate raw document fields to ``DOCUMENT_FORMAT_VERSION``."""
+    migrated = deepcopy(dict(data))
+    version = _version(migrated.get("format_version"), "format_version")
+    migrated = _apply(
+        migrated,
+        version,
+        DOCUMENT_FORMAT_VERSION,
+        DOCUMENT_MIGRATIONS,
+        "document",
+    )
+    migrated["format_version"] = DOCUMENT_FORMAT_VERSION
+    return migrated
+
+
+def _apply(
+    data: dict[str, Any],
+    version: int,
+    target: int,
+    migrations: Mapping[int, Migration],
+    kind: str,
+) -> dict[str, Any]:
+    """Apply consecutive migrations from ``version`` up to ``target``."""
+    if version > target:
+        raise SerializationError(
+            f"unsupported {kind} format version {version}; newest supported version is {target}"
+        )
+    while version < target:
+        migration = migrations.get(version)
+        if migration is None:
+            raise SerializationError(
+                f"no {kind} migration is registered from version {version} to {version + 1}"
+            )
+        data = migration(data)
+        version += 1
+    return data
+
+
+def _version(value: Any, field: str) -> int:
+    """Validate and return a positive integer version value."""
+    if not isinstance(value, int) or isinstance(value, bool) or value < 1:
+        raise SerializationError(f"{field}: expected a positive integer")
+    return value
diff --git a/src/bedit_core/serialization/schema.py b/src/bedit_core/serialization/schema.py
new file mode 100644
index 0000000..a3edca7
--- /dev/null
+++ b/src/bedit_core/serialization/schema.py
@@ -0,0 +1,381 @@
+"""Conversion between raw serialization values and model dataclasses."""
+
+from __future__ import annotations
+
+from collections.abc import Mapping
+from enum import Enum
+from typing import Any
+
+from bedit_core.models import (
+    BondCausality,
+    BondConnection,
+    BondPort,
+    Component,
+    ComponentID,
+    Connection,
+    ConnectionID,
+    Document,
+    EquationImplementation,
+    Graph,
+    GraphImplementation,
+    ID,
+    Interface,
+    Parameter,
+    ParameterID,
+    Port,
+    PortCausality,
+    PortID,
+    SignalConnection,
+    SignalDirection,
+    SignalPort,
+)
+
+class SerializationError(ValueError):
+    """Raised when serialized data does not match the bedit schema."""
+
+
+def document_to_data(document: Document) -> dict[str, Any]:
+    """Convert a document tree to JSON/MessagePack-compatible values."""
+    if not isinstance(document, Document):
+        raise TypeError("document must be a Document")
+    return {
+        "format_version": document.format_version,
+        "id": str(document.id),
+        "name": document.name,
+        "root": {
+            str(component_id): _component_to_data(component)
+            for component_id, component in document.root.items()
+        },
+        "metadata": _plain_value(document.metadata),
+    }
+
+
+def document_from_data(data: Mapping[str, Any]) -> Document:
+    """Validate raw values and construct a complete document dataclass tree."""
+    obj = _mapping(data, "document")
+    return Document(
+        format_version=_integer(_required(obj, "format_version", "document"), "document.format_version"),
+        id=ID(_string(_required(obj, "id", "document"), "document.id")),
+        name=_string(_required(obj, "name", "document"), "document.name"),
+        root={
+            ComponentID(_string(key, "document.root key")): _component_from_data(value, f"document.root[{key!r}]")
+            for key, value in _mapping(_required(obj, "root", "document"), "document.root").items()
+        },
+        metadata=_metadata(obj.get("metadata")),
+    )
+
+
+def _component_to_data(component: Component) -> dict[str, Any]:
+    """Convert a component and its nested model objects to raw values."""
+    return {
+        "name": component.name,
+        "interface": {
+            "ports": {
+                str(port_id): _port_to_data(port)
+                for port_id, port in component.interface.ports.items()
+            }
+        },
+        "parameters": {
+            str(parameter_id): _parameter_to_data(parameter)
+            for parameter_id, parameter in component.parameters.items()
+        },
+        "implementation": _implementation_to_data(component.implementation),
+    }
+
+
+def _component_from_data(value: Any, where: str) -> Component:
+    """Construct a component from raw values, using ``where`` in errors."""
+    obj = _mapping(value, where)
+    interface = _mapping(_required(obj, "interface", where), f"{where}.interface")
+    ports = _mapping(_required(interface, "ports", f"{where}.interface"), f"{where}.interface.ports")
+    parameters = _mapping(_required(obj, "parameters", where), f"{where}.parameters")
+    return Component(
+        name=_string(_required(obj, "name", where), f"{where}.name"),
+        interface=Interface(
+            ports={
+                PortID(_string(key, f"{where}.interface.ports key")): _port_from_data(port, f"{where}.interface.ports[{key!r}]")
+                for key, port in ports.items()
+            }
+        ),
+        parameters={
+            ParameterID(_string(key, f"{where}.parameters key")): _parameter_from_data(parameter, f"{where}.parameters[{key!r}]")
+            for key, parameter in parameters.items()
+        },
+        implementation=_implementation_from_data(_required(obj, "implementation", where), f"{where}.implementation"),
+    )
+
+
+def _port_to_data(port: Port) -> dict[str, Any]:
+    """Convert a derived port and add its serialized type discriminator."""
+    common = {
+        "name": port.name,
+        "direction": port.direction.value,
+        "multiplicity": port.multiplicity,
+        "matrix_size": list(port.matrix_size),
+        "description": port.description,
+    }
+    if isinstance(port, SignalPort):
+        return {
+            "port_type": "signal",
+            **common,
+            "value_type": port.value_type,
+            "quantity": port.quantity,
+            "unit": port.unit,
+        }
+    if isinstance(port, BondPort):
+        return {
+            "port_type": "bond",
+            **common,
+            "domain": port.domain,
+            "causality_preference": port.causality_preference.value,
+        }
+    raise TypeError(f"unsupported port class: {type(port).__name__}")
+
+
+def _port_from_data(value: Any, where: str) -> Port:
+    """Construct the port subclass selected by ``port_type``."""
+    obj = _mapping(value, where)
+    port_type = _string(_required(obj, "port_type", where), f"{where}.port_type")
+    common = {
+        "name": _string(_required(obj, "name", where), f"{where}.name"),
+        "direction": _enum(SignalDirection, _required(obj, "direction", where), f"{where}.direction"),
+        "multiplicity": _boolean(obj.get("multiplicity", False), f"{where}.multiplicity"),
+        "matrix_size": _matrix_size(obj.get("matrix_size", [1, 1]), f"{where}.matrix_size"),
+        "description": _optional_string(obj.get("description"), f"{where}.description"),
+    }
+    if port_type == "signal":
+        return SignalPort(
+            **common,
+            value_type=_string(obj.get("value_type", "Real"), f"{where}.value_type"),
+            quantity=_optional_string(obj.get("quantity"), f"{where}.quantity"),
+            unit=_optional_string(obj.get("unit"), f"{where}.unit"),
+        )
+    if port_type == "bond":
+        return BondPort(
+            **common,
+            domain=_string(obj.get("domain", ""), f"{where}.domain"),
+            causality_preference=_enum(
+                PortCausality,
+                obj.get("causality_preference", PortCausality.INDIFFERENT.value),
+                f"{where}.causality_preference",
+            ),
+        )
+    raise SerializationError(f"{where}.port_type: unsupported value {port_type!r}")
+
+
+def _parameter_to_data(parameter: Parameter) -> dict[str, Any]:
+    """Convert a parameter to raw values."""
+    return {
+        "name": parameter.name,
+        "value": _plain_value(parameter.value),
+        "value_type": parameter.value_type,
+        "quantity": parameter.quantity,
+        "unit": parameter.unit,
+        "description": parameter.description,
+    }
+
+
+def _parameter_from_data(value: Any, where: str) -> Parameter:
+    """Construct a parameter from validated raw values."""
+    obj = _mapping(value, where)
+    return Parameter(
+        name=_string(_required(obj, "name", where), f"{where}.name"),
+        value=obj.get("value", 1.0),
+        value_type=_string(obj.get("value_type", "Real"), f"{where}.value_type"),
+        quantity=_optional_string(obj.get("quantity"), f"{where}.quantity"),
+        unit=_optional_string(obj.get("unit"), f"{where}.unit"),
+        description=_optional_string(obj.get("description"), f"{where}.description"),
+    )
+
+
+def _implementation_to_data(implementation: GraphImplementation | EquationImplementation) -> dict[str, Any]:
+    """Convert an implementation and add its serialized discriminator."""
+    if isinstance(implementation, GraphImplementation):
+        return {
+            "implementation_type": "graph",
+            "graph": _graph_to_data(implementation.graph),
+        }
+    if isinstance(implementation, EquationImplementation):
+        return {
+            "implementation_type": "equation",
+            "declarations": list(implementation.declarations),
+            "initial_equations": list(implementation.initial_equations),
+            "equations": list(implementation.equations),
+        }
+    raise TypeError(f"unsupported implementation class: {type(implementation).__name__}")
+
+
+def _implementation_from_data(value: Any, where: str) -> GraphImplementation | EquationImplementation:
+    """Construct the implementation selected by ``implementation_type``."""
+    obj = _mapping(value, where)
+    implementation_type = _string(
+        _required(obj, "implementation_type", where),
+        f"{where}.implementation_type",
+    )
+    if implementation_type == "graph":
+        return GraphImplementation(_graph_from_data(_required(obj, "graph", where), f"{where}.graph"))
+    if implementation_type == "equation":
+        return EquationImplementation(
+            declarations=_string_list(obj.get("declarations", []), f"{where}.declarations"),
+            initial_equations=_string_list(obj.get("initial_equations", []), f"{where}.initial_equations"),
+            equations=_string_list(obj.get("equations", []), f"{where}.equations"),
+        )
+    raise SerializationError(
+        f"{where}.implementation_type: unsupported value {implementation_type!r}"
+    )
+
+
+def _graph_to_data(graph: Graph) -> dict[str, Any]:
+    """Convert a graph, including its components and connections."""
+    return {
+        "components": {
+            str(component_id): _component_to_data(component)
+            for component_id, component in graph.components.items()
+        },
+        "connections": {
+            str(connection_id): _connection_to_data(connection)
+            for connection_id, connection in graph.connections.items()
+        },
+    }
+
+
+def _graph_from_data(value: Any, where: str) -> Graph:
+    """Construct a graph and all nested dataclasses from raw values."""
+    obj = _mapping(value, where)
+    components = _mapping(obj.get("components", {}), f"{where}.components")
+    connections = _mapping(obj.get("connections", {}), f"{where}.connections")
+    return Graph(
+        components={
+            ComponentID(_string(key, f"{where}.components key")): _component_from_data(component, f"{where}.components[{key!r}]")
+            for key, component in components.items()
+        },
+        connections={
+            ConnectionID(_string(key, f"{where}.connections key")): _connection_from_data(connection, f"{where}.connections[{key!r}]")
+            for key, connection in connections.items()
+        },
+    )
+
+
+def _connection_to_data(connection: Connection) -> dict[str, Any]:
+    """Convert a derived connection and add its serialized discriminator."""
+    common = {"source": str(connection.source), "target": str(connection.target)}
+    if isinstance(connection, BondConnection):
+        return {
+            "connection_type": "bond",
+            **common,
+            "causality": connection.causality.value,
+            "undesired": connection.undesired,
+        }
+    if isinstance(connection, SignalConnection):
+        return {"connection_type": "signal", **common}
+    raise TypeError(f"unsupported connection class: {type(connection).__name__}")
+
+
+def _connection_from_data(value: Any, where: str) -> Connection:
+    """Construct the connection subclass selected by ``connection_type``."""
+    obj = _mapping(value, where)
+    # Files produced before the serializer existed had no connection discriminator.
+    connection_type = obj.get("connection_type")
+    if connection_type is None:
+        connection_type = "bond" if "causality" in obj or "undesired" in obj else "signal"
+    connection_type = _string(connection_type, f"{where}.connection_type")
+    common = {
+        "source": PortID(_string(_required(obj, "source", where), f"{where}.source")),
+        "target": PortID(_string(_required(obj, "target", where), f"{where}.target")),
+    }
+    if connection_type == "signal":
+        return SignalConnection(**common)
+    if connection_type == "bond":
+        return BondConnection(
+            **common,
+            causality=_enum(BondCausality, obj.get("causality", "none"), f"{where}.causality"),
+            undesired=_boolean(obj.get("undesired", False), f"{where}.undesired"),
+        )
+    raise SerializationError(f"{where}.connection_type: unsupported value {connection_type!r}")
+
+
+def _plain_value(value: Any) -> Any:
+    """Recursively convert metadata or parameter values to codec-safe values."""
+    if isinstance(value, Enum):
+        return value.value
+    if value is None or isinstance(value, (str, int, float, bool)):
+        return value
+    if isinstance(value, Mapping):
+        return {str(key): _plain_value(item) for key, item in value.items()}
+    if isinstance(value, (list, tuple)):
+        return [_plain_value(item) for item in value]
+    raise TypeError(f"value of type {type(value).__name__} is not serializable")
+
+
+def _required(obj: Mapping[str, Any], key: str, where: str) -> Any:
+    """Return a required mapping value or raise a contextual schema error."""
+    if key not in obj:
+        raise SerializationError(f"{where}: missing required field {key!r}")
+    return obj[key]
+
+
+def _mapping(value: Any, where: str) -> Mapping[str, Any]:
+    """Validate that ``value`` is a mapping."""
+    if not isinstance(value, Mapping):
+        raise SerializationError(f"{where}: expected an object")
+    return value
+
+
+def _string(value: Any, where: str) -> str:
+    """Validate that ``value`` is a string."""
+    if not isinstance(value, str):
+        raise SerializationError(f"{where}: expected a string")
+    return value
+
+
+def _integer(value: Any, where: str) -> int:
+    """Validate that ``value`` is an integer but not a boolean."""
+    if not isinstance(value, int) or isinstance(value, bool):
+        raise SerializationError(f"{where}: expected an integer")
+    return value
+
+
+def _boolean(value: Any, where: str) -> bool:
+    """Validate that ``value`` is a boolean."""
+    if not isinstance(value, bool):
+        raise SerializationError(f"{where}: expected a boolean")
+    return value
+
+
+def _optional_string(value: Any, where: str) -> str | None:
+    """Validate that ``value`` is either a string or ``None``."""
+    return None if value is None else _string(value, where)
+
+
+def _string_list(value: Any, where: str) -> list[str]:
+    """Validate and copy a list of strings."""
+    if not isinstance(value, list):
+        raise SerializationError(f"{where}: expected a list")
+    return [_string(item, f"{where}[{index}]") for index, item in enumerate(value)]
+
+
+def _matrix_size(value: Any, where: str) -> list[int]:
+    """Validate and copy a two-integer matrix size."""
+    if (
+        not isinstance(value, list)
+        or len(value) != 2
+        or any(not isinstance(item, int) or isinstance(item, bool) for item in value)
+    ):
+        raise SerializationError(f"{where}: expected a list of two integers")
+    return list(value)
+
+
+def _metadata(value: Any) -> dict[str, Any] | None:
+    """Validate and copy optional document metadata."""
+    if value is None:
+        return None
+    return dict(_mapping(value, "document.metadata"))
+
+
+def _enum(enum_type: type[Enum], value: Any, where: str) -> Any:
+    """Convert a serialized enum value or raise a contextual schema error."""
+    try:
+        return enum_type(value)
+    except (TypeError, ValueError):
+        choices = ", ".join(repr(member.value) for member in enum_type)
+        raise SerializationError(f"{where}: expected one of {choices}") from None
diff --git a/tests/conftest.py b/tests/conftest.py
new file mode 100644
index 0000000..45fdee5
--- /dev/null
+++ b/tests/conftest.py
@@ -0,0 +1,37 @@
+"""Shared pytest fixtures for bedit tests.
+
+Qt-specific fixtures can be added here later when GUI testing starts. Keeping
+the model fixtures independent from Qt lets the core suite stay lightweight.
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from bedit_core.models import (
+    Component,
+    ComponentID,
+    Document,
+    Graph,
+    GraphImplementation,
+    ID,
+    Interface,
+)
+
+
+@pytest.fixture
+def minimal_document() -> Document:
+    """Return the smallest useful graph document for core tests."""
+    root_id = ComponentID("root")
+    root = Component(
+        name="Root",
+        interface=Interface(),
+        parameters={},
+        implementation=GraphImplementation(Graph()),
+    )
+    return Document(
+        format_version=1,
+        id=ID("document"),
+        name="Test document",
+        root={root_id: root},
+    )
diff --git a/tests/unit/test_models.py b/tests/unit/test_models.py
new file mode 100644
index 0000000..f4db6c4
--- /dev/null
+++ b/tests/unit/test_models.py
@@ -0,0 +1,14 @@
+"""Basic smoke tests for the core model."""
+
+from __future__ import annotations
+
+import pytest
+
+from bedit_core.models import Document, GraphImplementation
+
+
+@pytest.mark.unit
+def test_minimal_document_has_graph_root(minimal_document: Document) -> None:
+    root = next(iter(minimal_document.root.values()))
+
+    assert isinstance(root.implementation, GraphImplementation)
diff --git a/tests/unit/test_serialization.py b/tests/unit/test_serialization.py
new file mode 100644
index 0000000..8d91cea
--- /dev/null
+++ b/tests/unit/test_serialization.py
@@ -0,0 +1,17 @@
+"""Basic smoke tests for document serialization."""
+
+from __future__ import annotations
+
+import pytest
+
+from bedit_core.models import Document
+from bedit_core.serialization import load, save
+
+
+@pytest.mark.unit
+def test_json_round_trip(tmp_path, minimal_document: Document) -> None:
+    path = tmp_path / "document.json"
+
+    save(minimal_document, path)
+
+    assert load(path) == minimal_document