core/generate/meaning_graph/model.py
Shay 733d4df366 feat(comprehend): MeaningGraph — neutral general-meaning interlingua (Phase 2a)
The refusal-first, provenance-carrying structure the field-decode produces and
the domain reasoners project from. Sibling of the binding-graph (ADR-0132) but
carries GENERAL meaning (entities + n-ary named relations), neutral to the
engine substrate (no algebra/field/numpy import), and imposes NO acyclicity
(relation cycles are well-formed, unlike the equation DAG).

Refusal-first construction: non-identifier ids, empty predicates, zero-arity
relations, duplicate entity ids, and relations referencing unknown entities all
refuse at construction. Deterministic to_canonical_string for replay/hashing.
Polarity (negated) is first-class. kind/predicate carry no closed vocab yet
(defer-substrate-vocab).

Phase 2a foundation under Path alpha (field standing-hand + refusal floor); the
field-decode -> refusal-floor -> MeaningGraph reader is the next increment.

18 new tests (each bites under its named violation); architectural invariants +
capability index green.
2026-06-05 15:45:59 -07:00

229 lines
8.8 KiB
Python

"""Frozen data model for the MeaningGraph — the general-meaning interlingua.
This module is the typed boundary between comprehended natural language and the
domain reasoners. Like ``generate.binding_graph.model`` it holds *only data* —
no parser, no solver, no algebra — and every dataclass is ``frozen=True,
slots=True`` with immutable ``tuple`` collections.
Refusal-first: invalid construction raises ``MeaningGraphError`` rather than
silently coercing. Neutral by design: imports nothing from ``algebra`` /
``field`` / ``numpy`` / the engine, so the structure is a fair meeting point for
two independent decodings (INV-26-style neutrality).
Distinct from the binding-graph in two deliberate ways:
- it carries GENERAL meaning (entities + n-ary named relations), not
quantities/equations;
- it imposes **no acyclicity** constraint. A cycle in general relations
("A loves B, B loves A") is well-formed, not the circular *reasoning* the
binding-graph's equation DAG forbids.
"""
from __future__ import annotations
from dataclasses import dataclass, field
class MeaningGraphError(ValueError):
"""Raised on invalid MeaningGraph construction; never silently coerces."""
def _require_non_empty_str(value: object, field_name: str) -> None:
if not isinstance(value, str) or value == "":
raise MeaningGraphError(f"{field_name} must be a non-empty str; got {value!r}")
def _require_identifier(value: object, field_name: str) -> None:
_require_non_empty_str(value, field_name)
assert isinstance(value, str)
if not value.isidentifier():
raise MeaningGraphError(
f"{field_name} must be a Python identifier; got {value!r}"
)
# --------------------------------------------------------------------------- #
# MeaningSpan — provenance
# --------------------------------------------------------------------------- #
@dataclass(frozen=True, slots=True)
class MeaningSpan:
"""An immutable pointer back to a ``[start, end)`` slice of the NL source.
``text`` is retained verbatim so downstream tooling can audit the span
without re-reading the source document.
"""
source_id: str
start: int
end: int
text: str
def __post_init__(self) -> None:
_require_non_empty_str(self.source_id, "MeaningSpan.source_id")
if not isinstance(self.start, int) or isinstance(self.start, bool):
raise MeaningGraphError(f"MeaningSpan.start must be int; got {self.start!r}")
if not isinstance(self.end, int) or isinstance(self.end, bool):
raise MeaningGraphError(f"MeaningSpan.end must be int; got {self.end!r}")
if self.start < 0:
raise MeaningGraphError(f"MeaningSpan.start must be >= 0; got {self.start}")
if self.end <= self.start:
raise MeaningGraphError(
f"MeaningSpan.end must be > start; got start={self.start}, end={self.end}"
)
_require_non_empty_str(self.text, "MeaningSpan.text")
def to_canonical_string(self) -> str:
return f"{self.source_id}[{self.start}:{self.end}]"
# --------------------------------------------------------------------------- #
# Entity
# --------------------------------------------------------------------------- #
@dataclass(frozen=True, slots=True)
class Entity:
"""A referent lifted from language: a stable id + surface name + provenance.
``entity_id`` is a Python identifier so relations can key it safely.
``kind`` is an optional, open free-text class hint (e.g. "person",
"number"); it carries NO closed vocabulary yet (defer-substrate-vocab —
a closed taxonomy is a deliberate later extension driven by a real use case).
"""
entity_id: str
name: str
span: MeaningSpan
kind: str | None = None
def __post_init__(self) -> None:
_require_identifier(self.entity_id, "Entity.entity_id")
_require_non_empty_str(self.name, "Entity.name")
if not isinstance(self.span, MeaningSpan):
raise MeaningGraphError(
f"Entity.span must be a MeaningSpan; got {type(self.span).__name__}"
)
if self.kind is not None and (not isinstance(self.kind, str) or self.kind == ""):
raise MeaningGraphError(
f"Entity.kind must be None or a non-empty str; got {self.kind!r}"
)
# --------------------------------------------------------------------------- #
# Relation
# --------------------------------------------------------------------------- #
@dataclass(frozen=True, slots=True)
class Relation:
"""An n-ary named predicate over entity ids, with provenance and polarity.
``predicate`` is a free-text relation name (e.g. ``"mother_of"``); like
``Entity.kind`` it carries no closed vocabulary yet. ``arguments`` is the
*ordered* tuple of entity ids the predicate relates (arity >= 1). ``negated``
captures polarity ("A is NOT the mother of B") as first-class structure.
"""
predicate: str
arguments: tuple[str, ...]
span: MeaningSpan
negated: bool = False
def __post_init__(self) -> None:
_require_non_empty_str(self.predicate, "Relation.predicate")
if not isinstance(self.arguments, tuple):
raise MeaningGraphError(
f"Relation.arguments must be a tuple; got {type(self.arguments).__name__}"
)
if len(self.arguments) == 0:
raise MeaningGraphError("Relation.arguments must be non-empty (arity >= 1)")
for arg in self.arguments:
_require_identifier(arg, "Relation.arguments entry")
if not isinstance(self.span, MeaningSpan):
raise MeaningGraphError(
f"Relation.span must be a MeaningSpan; got {type(self.span).__name__}"
)
if not isinstance(self.negated, bool):
raise MeaningGraphError(
f"Relation.negated must be a bool; got {self.negated!r}"
)
@property
def arity(self) -> int:
return len(self.arguments)
# --------------------------------------------------------------------------- #
# MeaningGraph
# --------------------------------------------------------------------------- #
@dataclass(frozen=True, slots=True)
class MeaningGraph:
"""Top-level immutable container of comprehended meaning.
Cross-collection invariants enforced at construction:
- ``entities`` carries unique ``entity_id`` values;
- every ``Relation`` argument references a known entity.
No acyclicity constraint (see module docstring). Collections are emitted in
*given* order; the graph is identity-preserving by design.
"""
entities: tuple[Entity, ...] = field(default_factory=tuple)
relations: tuple[Relation, ...] = field(default_factory=tuple)
provenance: tuple[MeaningSpan, ...] = field(default_factory=tuple)
def __post_init__(self) -> None:
for name, value, item_type in (
("entities", self.entities, Entity),
("relations", self.relations, Relation),
("provenance", self.provenance, MeaningSpan),
):
if not isinstance(value, tuple):
raise MeaningGraphError(
f"MeaningGraph.{name} must be a tuple; got {type(value).__name__}"
)
for item in value:
if not isinstance(item, item_type):
raise MeaningGraphError(
f"MeaningGraph.{name} entries must be {item_type.__name__}; "
f"got {type(item).__name__}"
)
known_ids: set[str] = set()
for ent in self.entities:
if ent.entity_id in known_ids:
raise MeaningGraphError(
f"Duplicate Entity.entity_id: {ent.entity_id!r}"
)
known_ids.add(ent.entity_id)
for rel in self.relations:
for arg in rel.arguments:
if arg not in known_ids:
raise MeaningGraphError(
f"Relation {rel.predicate!r} references unknown entity_id {arg!r}"
)
def to_canonical_string(self) -> str:
"""Deterministic string serialization for stable hashing / replay."""
lines: list[str] = []
for ent in self.entities:
lines.append(
f"E {ent.entity_id} {ent.name} kind={ent.kind} "
f"span={ent.span.to_canonical_string()}"
)
for rel in self.relations:
args = ",".join(rel.arguments)
polarity = "not " if rel.negated else ""
lines.append(
f"R {polarity}{rel.predicate}({args}) "
f"span={rel.span.to_canonical_string()}"
)
for span in self.provenance:
lines.append(f"P {span.to_canonical_string()} text={span.text}")
return "\n".join(lines)