Public surface (guardrail_calculus)

The builder/authoring API plus domain scalars. This layer is still being shaped by Phases 2–3, so several builders are documented by signature only for now.

guardrail_calculus

Public guardrail-calculus surface.

This module is mostly public re-export and high-level API assembly. The foundational nodes, predicates, dimensional protocols, and the local proxy _ live in the import-safe :mod:guardrail_kernel; the generic dimension-algebra dispatch engine and symbolic conditional in :mod:guardrail_calculus.dimensions; concrete temporal values (duration_hours, clock_hm, ...) in :mod:guardrail_calculus.dimension_value_builtins; propositions and the typed check chain in :mod:guardrail_calculus.facts; the explanation DAG, solver utilities, and puzzle solving in the optional :mod:guardrail_solver package.

What remains defined here is the builder/authoring surface: Routes, schema-rooted refs, IR serialization, and the agent/graph/subgraph/subject builders. (The last two bespoke domain-scalar classes, Probability and Money, migrated onto the values API — probability/money_gbp/money_usd in :mod:guardrail_calculus.dimension_value_builtins.)

money_gbp module-attribute

money_gbp = ctor(GBP)

money_usd module-attribute

money_usd = ctor(USD)

routes module-attribute

routes = RoutesFactory()

ref module-attribute

ref = RefFactory()

agent module-attribute

agent = AgentFactory()

graph module-attribute

graph = GraphFactory()

subgraph module-attribute

subgraph = SubgraphFactory()

subject module-attribute

subject = SubjectFactory()

Routes dataclass

Represents a set of target destinations for graph routing.

Attributes:

Name	Type	Description
targets	frozenset[RouteTarget]	Set of successor vertex references.
is_out	bool	True if routing out of a subgraph.

Examples:

>>> from guardrail_calculus import Path, Routes, VertexRef
>>> r = Routes(frozenset([VertexRef(Path(("a",)))]))
>>> r.label
'routes.to(a)'

Source code in src/guardrail_calculus/__init__.py

@typing.final
@dataclass(frozen=True)
class Routes:
    """Represents a set of target destinations for graph routing.

    Attributes:
        targets: Set of successor vertex references.
        is_out: True if routing out of a subgraph.

    Examples:
        >>> from guardrail_calculus import Path, Routes, VertexRef
        >>> r = Routes(frozenset([VertexRef(Path(("a",)))]))
        >>> r.label
        'routes.to(a)'
    """

    targets: frozenset[RouteTarget] = frozenset()
    is_out: bool = False

    def includes(self, target: RouteTarget) -> Predicate[_Includes]:
        return Predicate(
            _node=_Includes(self, target),
            label=f"{display_value(self)} includes {display_value(target)}",
        )

    @property
    def label(self) -> str:
        if self.is_out:
            return "routes.out"

        if not self.targets:
            return "routes.none()"

        return "routes.to(" + ", ".join(display_value(t) for t in self.targets) + ")"

targets class-attribute instance-attribute

targets = frozenset()

is_out class-attribute instance-attribute

is_out = False

label property

label

includes

includes(target)

Source code in src/guardrail_calculus/__init__.py

def includes(self, target: RouteTarget) -> Predicate[_Includes]:
    return Predicate(
        _node=_Includes(self, target),
        label=f"{display_value(self)} includes {display_value(target)}",
    )

RoutesFactory

Factory to build Routes instances using helper methods.

Examples:

>>> from guardrail_calculus import Path, VertexRef, routes
>>> routes.to(VertexRef(Path(("a",)))).label
'routes.to(a)'
>>> routes.none().label
'routes.none()'

Source code in src/guardrail_calculus/__init__.py

class RoutesFactory:
    """Factory to build [Routes][guardrail_calculus.Routes] instances using helper methods.

    Examples:
        >>> from guardrail_calculus import Path, VertexRef, routes
        >>> routes.to(VertexRef(Path(("a",)))).label
        'routes.to(a)'
        >>> routes.none().label
        'routes.none()'
    """

    @property
    def out(self) -> Routes:
        return Routes(is_out=True)

    def none(self) -> Routes:
        return Routes()

    def to(self, *targets: RouteTarget) -> Routes:
        return Routes(targets=frozenset(targets))

out property

out

none

none()

Source code in src/guardrail_calculus/__init__.py

def none(self) -> Routes:
    return Routes()

to

to(*targets)

Source code in src/guardrail_calculus/__init__.py

def to(self, *targets: RouteTarget) -> Routes:
    return Routes(targets=frozenset(targets))

RefFactory

Source code in src/guardrail_calculus/__init__.py

class RefFactory:
    def __getitem__[T](self, schema_type: type[T]) -> T:
        return cast(T, Proxy(schema_type))

AgentBuilder dataclass

Builder for constructing an AgentBlock.

Attributes:

Name	Type	Description
output_type	type[OutputsT] | None	Optional class representing outputs.
rules	tuple[AgentRule, ...]	Tuple of accumulated rules.
pending_given	Prop[Any] | None	The pending given predicate.

Examples:

>>> from guardrail_calculus import agent
>>> builder = agent().given(True).then(True)
>>> len(builder.rules)
1

Source code in src/guardrail_calculus/__init__.py

@dataclass(frozen=True)
class AgentBuilder[OutputsT]:
    """Builder for constructing an [AgentBlock][guardrail_calculus.AgentBlock].

    Attributes:
        output_type: Optional class representing outputs.
        rules: Tuple of accumulated rules.
        pending_given: The pending given predicate.

    Examples:
        >>> from guardrail_calculus import agent
        >>> builder = agent().given(True).then(True)
        >>> len(builder.rules)
        1
    """

    output_type: type[OutputsT] | None = None
    rules: tuple[AgentRule, ...] = ()
    pending_given: Prop[Any] | None = None

    def given(self, value: Boolish_Bound | Prop[Any]) -> AgentBuilder[OutputsT]:
        return dataclasses.replace(self, pending_given=prop(value))

    def then(
        self,
        *values: Boolish_Bound | Prop[Any],
    ) -> AgentBuilder[OutputsT]:
        rule = AgentRule(
            given=self.pending_given,
            consequences=tuple(prop(value) for value in values),
        )
        return AgentBuilder(
            output_type=self.output_type,
            rules=(*self.rules, rule),
            pending_given=None,
        )

    def block(self) -> AgentBlock[OutputsT]:
        return AgentBlock(output_type=self.output_type, rules=self.rules)

output_type class-attribute instance-attribute

output_type = None

rules class-attribute instance-attribute

rules = ()

pending_given class-attribute instance-attribute

pending_given = None

given

given(value)

Source code in src/guardrail_calculus/__init__.py

def given(self, value: Boolish_Bound | Prop[Any]) -> AgentBuilder[OutputsT]:
    return dataclasses.replace(self, pending_given=prop(value))

then

then(*values)

Source code in src/guardrail_calculus/__init__.py

def then(
    self,
    *values: Boolish_Bound | Prop[Any],
) -> AgentBuilder[OutputsT]:
    rule = AgentRule(
        given=self.pending_given,
        consequences=tuple(prop(value) for value in values),
    )
    return AgentBuilder(
        output_type=self.output_type,
        rules=(*self.rules, rule),
        pending_given=None,
    )

block

block()

Source code in src/guardrail_calculus/__init__.py

def block(self) -> AgentBlock[OutputsT]:
    return AgentBlock(output_type=self.output_type, rules=self.rules)

AgentBlock dataclass

A constructed agent block with a set of rules.

Attributes:

Name	Type	Description
output_type	type[OutputsT] | None	Optional class representing outputs.
rules	tuple[AgentRule, ...]	Tuple of rules inside the agent.

Examples:

>>> from guardrail_calculus import AgentBlock
>>> block = AgentBlock(output_type=None, rules=())
>>> len(block.rules)
0

Source code in src/guardrail_calculus/__init__.py

@dataclass(frozen=True)
class AgentBlock[OutputsT]:
    """A constructed agent block with a set of rules.

    Attributes:
        output_type: Optional class representing outputs.
        rules: Tuple of rules inside the agent.

    Examples:
        >>> from guardrail_calculus import AgentBlock
        >>> block = AgentBlock(output_type=None, rules=())
        >>> len(block.rules)
        0
    """

    output_type: type[OutputsT] | None
    rules: tuple[AgentRule, ...]

output_type instance-attribute

output_type

rules instance-attribute

rules

AgentRule dataclass

A rule inside an agent containing a premise and a set of consequences.

Attributes:

Name	Type	Description
given	Prop[Any] | None	Optional premise proposition.
consequences	tuple[Prop[Any], ...]	Sequence of consequences to apply.

Examples:

>>> from guardrail_calculus import AgentRule
>>> rule = AgentRule(given=None, consequences=())
>>> rule.given is None
True

Source code in src/guardrail_calculus/__init__.py

@dataclass(frozen=True)
class AgentRule:
    """A rule inside an agent containing a premise and a set of consequences.

    Attributes:
        given: Optional premise proposition.
        consequences: Sequence of consequences to apply.

    Examples:
        >>> from guardrail_calculus import AgentRule
        >>> rule = AgentRule(given=None, consequences=())
        >>> rule.given is None
        True
    """

    given: Prop[Any] | None
    consequences: tuple[Prop[Any], ...]

given instance-attribute

given

consequences instance-attribute

consequences

GraphBlock dataclass

A constructed graph block.

Attributes:

Name	Type	Description
vertices	Mapping[str, object]	Mapping from vertex names to blocks.

Examples:

>>> from guardrail_calculus import graph
>>> block = graph(a="vertex_a")
>>> block.vertices["a"]
'vertex_a'

Source code in src/guardrail_calculus/__init__.py

@dataclass(frozen=True)
class GraphBlock:
    """A constructed graph block.

    Attributes:
        vertices: Mapping from vertex names to blocks.

    Examples:
        >>> from guardrail_calculus import graph
        >>> block = graph(a="vertex_a")
        >>> block.vertices["a"]
        'vertex_a'
    """

    vertices: Mapping[str, object]

vertices instance-attribute

vertices

SubgraphBlock dataclass

A constructed subgraph block.

Attributes:

Name	Type	Description
output_type	type[OutputsT] | None	Optional class representing outputs.
vertices	Mapping[str, object]	Mapping from vertex names to sibling blocks.

Examples:

>>> from guardrail_calculus import subgraph
>>> block = subgraph(a="vertex_a")
>>> block.vertices["a"]
'vertex_a'

Source code in src/guardrail_calculus/__init__.py

@dataclass(frozen=True)
class SubgraphBlock[OutputsT]:
    """A constructed subgraph block.

    Attributes:
        output_type: Optional class representing outputs.
        vertices: Mapping from vertex names to sibling blocks.

    Examples:
        >>> from guardrail_calculus import subgraph
        >>> block = subgraph(a="vertex_a")
        >>> block.vertices["a"]
        'vertex_a'
    """

    output_type: type[OutputsT] | None
    vertices: Mapping[str, object]

output_type instance-attribute

output_type

vertices instance-attribute

vertices

probability

probability(value)

Construct a probability literal in [0, 1].

A thin authoring name only: Probability is comparison-only with no registered constructors, so this direct builder is its one authoring path — but it restates no bounds. checked_real gates finiteness (the REAL carrier's wire precondition), and DimensionValue itself evaluates the dimension's declared registry invariant (0 <= v <= 1) at construction — the same node tree the solver asserts on every Probability slot.

Parameters:

Name	Type	Description	Default
value	float	The probability, between 0 and 1 inclusive.	required

Returns:

Type	Description
DisplacementValue[Literal['Probability'], float]	The constructed probability value.

Source code in src/guardrail_calculus/dimension_value_builtins.py

def probability(value: float) -> DisplacementValue[Literal["Probability"], float]:
    """Construct a probability literal in [0, 1].

    A thin authoring name only: Probability is comparison-only with no
    registered constructors, so this direct builder is its one authoring
    path — but it restates *no* bounds. ``checked_real`` gates finiteness
    (the REAL carrier's wire precondition), and ``DimensionValue`` itself
    evaluates the dimension's declared registry invariant (``0 <= v <= 1``)
    at construction — the same node tree the solver asserts on every
    Probability slot.

    Args:
        value: The probability, between 0 and 1 inclusive.

    Returns:
        The constructed probability value.
    """
    # Probability is ``Vector``-structured, hence a displacement, and this is
    # a constructor, so the concrete class is both what it builds and what it
    # declares -- no cast, and no union for a caller's arithmetic to widen on.
    return DisplacementValue[Literal["Probability"], float](
        _node=checked_real(value), unit=None, _dimension=PROBABILITY
    )

validate_routes_out

validate_routes_out(block, *, in_subgraph=False)

Validate that routes.out is only used inside subgraphs.

Parameters:

Name	Type	Description	Default
block	AgentBlock[Any]	The agent block to validate.	required
in_subgraph	bool	True if nested inside a subgraph.	False

Raises:

Type	Description
ValueError	If routes.out is used outside a subgraph.

Source code in src/guardrail_calculus/__init__.py

def validate_routes_out(
    block: AgentBlock[Any],
    *,
    in_subgraph: bool = False,
) -> None:
    """Validate that `routes.out` is only used inside subgraphs.

    Args:
        block: The agent block to validate.
        in_subgraph: True if nested inside a subgraph.

    Raises:
        ValueError: If `routes.out` is used outside a subgraph.
    """
    for rule in block.rules:
        for consequence in rule.consequences:
            if _is_routes_out_assignment(consequence) and not in_subgraph:
                raise ValueError("routes.out is only valid inside a subgraph.")