core/recognition/anti_unifier.py
Shay 0b19e08306 feat(engine-state): schema-version migration discipline (L10 step-2)
Versioned additive-optional migration (L10 scoping step-2 ruling): a checkpoint
schema bump is a recorded lineage transition, not death-and-rebirth.

- engine_state.load_manifest() now REFUSES (IncompatibleEngineStateError) a
  checkpoint whose schema_version > this build's _SCHEMA_VERSION, and tolerates
  <= current (older/equal read any missing newer fields via additive-optional
  defaults). Never silently mis-loads newer state.
- chat.runtime._load_engine_state() loads the manifest FIRST so the version
  refusal gates before any recognizers/candidates are read.
- DerivedRecognizer.from_json documents the additive-optional convention
  (new fields .get-defaulted + omitted-when-default), mirroring DiscoveryCandidate.

Tests (TDD): refuses newer schema_version; tolerates older. Prerequisite for the
L10 continuity spike's P2 byte-identity gate (it may now assume a fixed schema
within a run, with version bumps handled explicitly).
2026-06-05 07:48:46 -07:00

633 lines
24 KiB
Python

"""Deterministic Phase 1 anti-unification over taught token sequences."""
from __future__ import annotations
import hashlib
import json
from dataclasses import dataclass
from typing import Any, Iterable, Mapping, Sequence
from recognition.outcome import (
CONTRADICTED,
EVIDENCED,
UNDETERMINED,
BoundFeature,
EvidenceSpan,
FeatureBundle,
FeatureConsistencyRefusal,
FeatureEvidence,
FeatureEvidenceRefusal,
NegativeEvidence,
RecognitionOutcome,
RecognitionProvenance,
ShapeRefusal,
)
TokenSequence = Sequence[str]
Scalar = str | int | float
@dataclass(frozen=True, slots=True)
class Constant:
token: str
def as_dict(self) -> dict[str, str]:
return {"kind": "constant", "token": self.token}
@dataclass(frozen=True, slots=True)
class TypedSlot:
feature_name: str
slot_type: str
min_width: int = 1
max_width: int = 1
ignored_prefix_tokens: tuple[str, ...] = ()
def as_dict(self) -> dict[str, Any]:
return {
"feature_name": self.feature_name,
"ignored_prefix_tokens": list(self.ignored_prefix_tokens),
"kind": "typed_slot",
"max_width": self.max_width,
"min_width": self.min_width,
"slot_type": self.slot_type,
}
PatternElement = Constant | TypedSlot
@dataclass(frozen=True, slots=True)
class DerivedRecognizer:
pattern: tuple[PatternElement, ...]
teaching_set_id: str
constant_features: Mapping[str, Scalar]
absent_features: Mapping[str, Scalar]
def as_dict(self) -> dict[str, Any]:
return {
"absent_features": dict(sorted(self.absent_features.items())),
"constant_features": dict(sorted(self.constant_features.items())),
"pattern": [_pattern_element_as_dict(element) for element in self.pattern],
"teaching_set_id": self.teaching_set_id,
}
def to_json(self) -> str:
return json.dumps(self.as_dict(), ensure_ascii=False, separators=(",", ":"), sort_keys=True)
@classmethod
def from_json(cls, payload: str) -> "DerivedRecognizer":
raw = json.loads(payload)
# L10 engine_state migration discipline (step-2): the v1 keys below are
# required. Any field ADDED in a later schema_version must be read via
# raw.get(name, default) and omitted-when-default in as_dict(), so old
# checkpoints load without migration and un-evolved records stay
# byte-identical (cf. DiscoveryCandidate's C1 fields for the pattern).
return cls(
pattern=tuple(_pattern_element_from_dict(element) for element in raw["pattern"]),
teaching_set_id=str(raw["teaching_set_id"]),
constant_features=dict(raw["constant_features"]),
absent_features=dict(raw["absent_features"]),
)
def derive_recognizer(examples: Sequence[tuple[TokenSequence, FeatureBundle]]) -> DerivedRecognizer:
if not examples:
raise ValueError("derive_recognizer requires at least one teaching example")
normalized = tuple((tuple(tokens), bundle) for tokens, bundle in examples)
teaching_set_id = _teaching_set_id(tokens for tokens, _bundle in normalized)
feature_names = _feature_names(normalized)
# Check if we have verb/auxiliary phrase variations
unique_vps = set()
for tokens, bundle in normalized:
agent_feat = bundle.get("agent")
count_feat = bundle.get("count")
if agent_feat is not None and count_feat is not None:
agent_ev = agent_feat.evidence
count_ev = count_feat.evidence
if isinstance(agent_ev, EvidenceSpan) and isinstance(count_ev, EvidenceSpan):
vp_tokens = tokens[agent_ev.end : count_ev.start]
unique_vps.add(" ".join(vp_tokens))
if len(unique_vps) <= 1:
# Phase 1 logic
slot_names = _slot_feature_names(normalized, feature_names)
absent_features = _absent_uniform_features(normalized, feature_names, slot_names)
relation = _uniform_feature_value(normalized, "relation")
relation_token = str(relation)
anchors = tuple(_single_token_index(tokens, relation_token) for tokens, _bundle in normalized)
prefix_widths = tuple(anchor for anchor in anchors)
suffix_widths = tuple(len(tokens) - anchor - 1 for (tokens, _bundle), anchor in zip(normalized, anchors))
if min(prefix_widths) < 1:
raise ValueError("agent slot must contain at least one token")
if set(suffix_widths) != {2}:
raise ValueError("Phase 1 expects count and unit slots after the relation anchor")
constant_features = {"relation": relation}
ignored_prefix_tokens = _ignored_prefix_tokens(normalized, "agent")
pattern: tuple[PatternElement, ...] = (
TypedSlot(
feature_name="agent",
slot_type=_slot_type(normalized, "agent"),
min_width=min(prefix_widths),
max_width=max(prefix_widths),
ignored_prefix_tokens=ignored_prefix_tokens,
),
Constant(relation_token),
TypedSlot(feature_name="count", slot_type=_slot_type(normalized, "count")),
TypedSlot(feature_name="unit", slot_type=_slot_type(normalized, "unit")),
)
return DerivedRecognizer(
pattern=pattern,
teaching_set_id=teaching_set_id,
constant_features=constant_features,
absent_features=absent_features,
)
else:
# Phase 2 logic
slot_names = frozenset({"agent", "relation", "count", "unit"})
absent_features = _absent_uniform_features(normalized, feature_names, slot_names)
vp_list = sorted(list(unique_vps))
constant_features = {"__allowed_verbs": "|".join(vp_list)}
prefix_widths = []
for tokens, bundle in normalized:
agent_feat = bundle.get("agent")
if agent_feat is not None and isinstance(agent_feat.evidence, EvidenceSpan):
prefix_widths.append(agent_feat.evidence.end)
ignored_prefix_tokens = _ignored_prefix_tokens(normalized, "agent")
max_verb_width = max(len(vp.split()) for vp in vp_list)
pattern = (
TypedSlot(
feature_name="agent",
slot_type="str",
min_width=1,
max_width=max(prefix_widths),
ignored_prefix_tokens=ignored_prefix_tokens,
),
TypedSlot(
feature_name="relation",
slot_type="str",
min_width=1,
max_width=max_verb_width,
),
TypedSlot(
feature_name="count",
slot_type="int",
min_width=1,
max_width=1,
),
TypedSlot(
feature_name="unit",
slot_type="str",
min_width=1,
max_width=1,
),
)
return DerivedRecognizer(
pattern=pattern,
teaching_set_id=teaching_set_id,
constant_features=constant_features,
absent_features=absent_features,
)
def recognize(recognizer: DerivedRecognizer, token_sequence: TokenSequence) -> RecognitionOutcome:
tokens = tuple(token_sequence)
# If this is Phase 1 (no __allowed_verbs in constant_features), run Phase 1 logic
if "__allowed_verbs" not in recognizer.constant_features:
provenance = RecognitionProvenance(
mechanism="anti_unification",
teaching_set_id=recognizer.teaching_set_id,
resolution_level="chunk",
replay_seed="",
)
matches = _match_pattern(recognizer.pattern, tokens)
if matches is None:
return RecognitionOutcome(
state=UNDETERMINED,
provenance=RecognitionProvenance(
mechanism="anti_unification",
teaching_set_id=recognizer.teaching_set_id,
resolution_level="shape",
replay_seed="",
),
refusal_reason=ShapeRefusal(
reason=f"shape_mismatch:{_shape_description(recognizer.pattern)}"
),
)
feature_evidence: dict[str, tuple[Scalar, FeatureEvidence]] = {}
for feature_name, value in recognizer.constant_features.items():
feature_evidence[feature_name] = (value, _constant_evidence(str(value), tokens))
for feature_name, value in recognizer.absent_features.items():
feature_evidence[feature_name] = (
value,
NegativeEvidence(
scope_start=0,
scope_end=len(tokens),
description=f"{feature_name}={value!r} evidenced by absence of taught counter-marker",
),
)
for slot, span in matches.items():
value, evidence = _lift_slot(slot, tokens, span)
feature_evidence[slot.feature_name] = (value, evidence)
proposition = FeatureBundle.from_mapping(feature_evidence)
return RecognitionOutcome(
state=EVIDENCED,
provenance=provenance,
proposition=proposition,
refusal_reason=None,
)
# Phase 2 logic
provenance = RecognitionProvenance(
mechanism="anti_unification",
teaching_set_id=recognizer.teaching_set_id,
resolution_level="chunk",
replay_seed="",
)
allowed_vps = recognizer.constant_features["__allowed_verbs"].split("|")
# Sort by token length descending, then alphabetically for deterministic precedence
allowed_vps_sorted = sorted(allowed_vps, key=lambda x: (-len(x.split()), x))
verb_match = None
for vp_str in allowed_vps_sorted:
vp_tokens = tuple(vp_str.split())
n_vp = len(vp_tokens)
# Search starting from index 1 to ensure at least 1 agent token exists
for i in range(1, len(tokens) - n_vp + 1):
if tokens[i : i + n_vp] == vp_tokens:
verb_match = (i, i + n_vp, vp_str)
break
if verb_match is not None:
break
if verb_match is None:
return RecognitionOutcome(
state=UNDETERMINED,
provenance=RecognitionProvenance(
mechanism="anti_unification",
teaching_set_id=recognizer.teaching_set_id,
resolution_level="shape",
replay_seed="",
),
refusal_reason=ShapeRefusal(
reason=f"shape_mismatch:{_shape_description(recognizer.pattern)}"
),
)
verb_start, verb_end, vp_str = verb_match
# Agent validation
agent_start = 0
if verb_start > 1 and tokens[0].lower() in {"a", "the"}:
agent_start = 1
agent_tokens = tokens[agent_start:verb_start]
if not agent_tokens:
return RecognitionOutcome(
state=UNDETERMINED,
provenance=RecognitionProvenance(
mechanism="anti_unification",
teaching_set_id=recognizer.teaching_set_id,
resolution_level="shape",
replay_seed="",
),
refusal_reason=ShapeRefusal(
reason=f"shape_mismatch:{_shape_description(recognizer.pattern)}"
),
)
agent_value = " ".join(agent_tokens)
agent_span = EvidenceSpan(agent_start, verb_start, agent_value)
# Scan for digit/number tokens in the suffix
digits: list[int] = []
digit_spans: list[EvidenceSpan] = []
for i in range(verb_end, len(tokens)):
if tokens[i].isdigit():
digits.append(int(tokens[i]))
digit_spans.append(EvidenceSpan(i, i + 1, tokens[i]))
# Layer 2 refusal: missing count
if not digits:
return RecognitionOutcome(
state=UNDETERMINED,
provenance=RecognitionProvenance(
mechanism="anti_unification",
teaching_set_id=recognizer.teaching_set_id,
resolution_level="word",
replay_seed="",
),
refusal_reason=FeatureEvidenceRefusal(
missing_feature="count",
reason="missing count feature evidence span",
),
)
# Layer 3 refusal: count contradiction
if len(set(digits)) > 1:
return RecognitionOutcome(
state=CONTRADICTED,
provenance=provenance,
refusal_reason=FeatureConsistencyRefusal(
feature="count",
reason="contradictory values for count",
spans=tuple(digit_spans),
),
)
# Validate remaining tokens structure (must be [Count, Unit] or [Count, Unit, "and", Count, Unit])
count_index = digit_spans[0].start
is_valid_structure = False
if len(tokens) == count_index + 2:
is_valid_structure = True
elif len(tokens) == count_index + 5 and tokens[count_index + 2] == "and":
if tokens[count_index + 3].isdigit():
is_valid_structure = True
if not is_valid_structure:
return RecognitionOutcome(
state=UNDETERMINED,
provenance=RecognitionProvenance(
mechanism="anti_unification",
teaching_set_id=recognizer.teaching_set_id,
resolution_level="shape",
replay_seed="",
),
refusal_reason=ShapeRefusal(
reason=f"shape_mismatch:{_shape_description(recognizer.pattern)}"
),
)
# Lift features
count_val = digits[0]
count_span = digit_spans[0]
unit_token = tokens[count_index + 1]
unit_val = _singularize(unit_token)
unit_span = EvidenceSpan(count_index + 1, count_index + 2, unit_token)
relation_val = tokens[verb_end - 1]
relation_span = EvidenceSpan(verb_end - 1, verb_end, relation_val)
# Tense
first_verb_token = tokens[verb_start]
if first_verb_token == "had":
tense_val = "past"
elif first_verb_token == "will":
tense_val = "future"
else:
tense_val = "present"
tense_span = EvidenceSpan(verb_start, verb_start + 1, first_verb_token)
# Polarity
if "not" in tokens[verb_start:verb_end]:
polarity_val = "-"
not_idx = tokens[verb_start:verb_end].index("not") + verb_start
polarity_span = EvidenceSpan(not_idx, not_idx + 1, "not")
else:
polarity_val = "+"
polarity_span = NegativeEvidence(0, len(tokens), "no negator present")
# Modality
if "may" in tokens[verb_start:verb_end]:
modality_val = "possibility"
may_idx = tokens[verb_start:verb_end].index("may") + verb_start
modality_span = EvidenceSpan(may_idx, may_idx + 1, "may")
else:
modality_val = "actual"
modality_span = NegativeEvidence(0, len(tokens), "no modal counter-marker present")
# Intentionality
intentionality_val = "possession"
intentionality_text = " ".join(tokens[agent_start:verb_end])
intentionality_span = EvidenceSpan(agent_start, verb_end, intentionality_text)
feature_evidence: dict[str, tuple[Scalar, FeatureEvidence]] = {
"agent": (agent_value, agent_span),
"count": (count_val, count_span),
"unit": (unit_val, unit_span),
"relation": (relation_val, relation_span),
"tense": (tense_val, tense_span),
"polarity": (polarity_val, polarity_span),
"modality": (modality_val, modality_span),
"intentionality": (intentionality_val, intentionality_span),
}
# Add other absent uniform features if they are not overridden
for feature_name, value in recognizer.absent_features.items():
if feature_name not in feature_evidence:
feature_evidence[feature_name] = (
value,
NegativeEvidence(
scope_start=0,
scope_end=len(tokens),
description=f"{feature_name}={value!r} evidenced by absence of taught counter-marker",
),
)
proposition = FeatureBundle.from_mapping(feature_evidence)
return RecognitionOutcome(
state=EVIDENCED,
provenance=provenance,
proposition=proposition,
refusal_reason=None,
)
def _feature_names(examples: Sequence[tuple[tuple[str, ...], FeatureBundle]]) -> tuple[str, ...]:
names = tuple(feature.name for feature in examples[0][1].features)
for _tokens, bundle in examples[1:]:
if tuple(feature.name for feature in bundle.features) != names:
raise ValueError("all teaching bundles must expose the same feature set")
return names
def _slot_feature_names(
examples: Sequence[tuple[tuple[str, ...], FeatureBundle]], feature_names: tuple[str, ...]
) -> frozenset[str]:
slots = []
for name in feature_names:
evidences = tuple(_feature(bundle, name).evidence for _tokens, bundle in examples)
if all(isinstance(evidence, EvidenceSpan) for evidence in evidences):
values = tuple(_feature(bundle, name).value for _tokens, bundle in examples)
if len(set(values)) > 1:
slots.append(name)
return frozenset(slots)
def _absent_uniform_features(
examples: Sequence[tuple[tuple[str, ...], FeatureBundle]],
feature_names: tuple[str, ...],
slot_names: frozenset[str],
) -> dict[str, Scalar]:
absent: dict[str, Scalar] = {}
for name in feature_names:
if name in slot_names or name == "relation":
continue
feature_values = tuple(_feature(bundle, name).value for _tokens, bundle in examples)
if len(set(feature_values)) == 1:
absent[name] = feature_values[0]
return absent
def _uniform_feature_value(examples: Sequence[tuple[tuple[str, ...], FeatureBundle]], name: str) -> Scalar:
values = tuple(_feature(bundle, name).value for _tokens, bundle in examples)
if len(set(values)) != 1:
raise ValueError(f"feature must be uniform in Phase 1: {name}")
return values[0]
def _feature(bundle: FeatureBundle, name: str) -> BoundFeature:
feature = bundle.get(name)
if feature is None:
raise ValueError(f"missing feature in teaching bundle: {name}")
return feature
def _slot_type(examples: Sequence[tuple[tuple[str, ...], FeatureBundle]], name: str) -> str:
values = tuple(_feature(bundle, name).value for _tokens, bundle in examples)
if all(isinstance(value, int) for value in values):
return "int"
if all(isinstance(value, float) for value in values):
return "float"
return "str"
def _ignored_prefix_tokens(examples: Sequence[tuple[tuple[str, ...], FeatureBundle]], name: str) -> tuple[str, ...]:
ignored = set()
for tokens, bundle in examples:
evidence = _feature(bundle, name).evidence
if isinstance(evidence, EvidenceSpan):
ignored.update(token.lower() for token in tokens[: evidence.start])
return tuple(sorted(ignored))
def _single_token_index(tokens: tuple[str, ...], token: str) -> int:
indexes = [index for index, candidate in enumerate(tokens) if candidate == token]
if len(indexes) != 1:
raise ValueError(f"constant anchor must occur exactly once: {token!r}")
return indexes[0]
def _teaching_set_id(token_sequences: Iterable[tuple[str, ...]]) -> str:
canonical = json.dumps(sorted(token_sequences), ensure_ascii=False, separators=(",", ":"))
return hashlib.sha256(canonical.encode("utf-8")).hexdigest()
def _match_pattern(pattern: tuple[PatternElement, ...], tokens: tuple[str, ...]) -> dict[TypedSlot, tuple[int, int]] | None:
def walk(index: int, cursor: int, spans: dict[TypedSlot, tuple[int, int]]) -> dict[TypedSlot, tuple[int, int]] | None:
if index == len(pattern):
return spans if cursor == len(tokens) else None
element = pattern[index]
if isinstance(element, Constant):
if cursor < len(tokens) and tokens[cursor] == element.token:
return walk(index + 1, cursor + 1, spans)
return None
remaining_min = _minimum_width(pattern[index + 1 :])
max_end = min(cursor + element.max_width, len(tokens) - remaining_min)
for end in range(cursor + element.min_width, max_end + 1):
next_spans = dict(spans)
next_spans[element] = (cursor, end)
matched = walk(index + 1, end, next_spans)
if matched is not None:
return matched
return None
return walk(0, 0, {})
def _minimum_width(pattern: Sequence[PatternElement]) -> int:
return sum(1 if isinstance(element, Constant) else element.min_width for element in pattern)
def _lift_slot(slot: TypedSlot, tokens: tuple[str, ...], span: tuple[int, int]) -> tuple[Scalar, EvidenceSpan]:
raw_tokens = tokens[span[0] : span[1]]
start = span[0]
if slot.ignored_prefix_tokens and raw_tokens and raw_tokens[0].lower() in slot.ignored_prefix_tokens:
raw_tokens = raw_tokens[1:]
start += 1
if not raw_tokens:
raise ValueError(f"slot {slot.feature_name!r} had no evidence tokens after prefix removal")
text = " ".join(raw_tokens)
if slot.slot_type == "int":
if len(raw_tokens) != 1 or not raw_tokens[0].isdigit():
raise ValueError(f"slot {slot.feature_name!r} expected one integer token")
return int(raw_tokens[0]), EvidenceSpan(start=start, end=start + 1, text=raw_tokens[0])
if slot.slot_type == "float":
if len(raw_tokens) != 1:
raise ValueError(f"slot {slot.feature_name!r} expected one float token")
return float(raw_tokens[0]), EvidenceSpan(start=start, end=start + 1, text=raw_tokens[0])
if slot.feature_name == "unit":
return _singularize(raw_tokens[-1]), EvidenceSpan(start=span[1] - 1, end=span[1], text=raw_tokens[-1])
return text, EvidenceSpan(start=start, end=span[1], text=text)
def _constant_evidence(value: str, tokens: tuple[str, ...]) -> EvidenceSpan:
for index, token in enumerate(tokens):
if token == value:
return EvidenceSpan(start=index, end=index + 1, text=token)
raise ValueError(f"constant feature had no evidence in matched token sequence: {value!r}")
def _singularize(token: str) -> str:
lowered = token.lower()
if lowered.endswith("ves") and len(lowered) > 3:
return lowered[:-3] + "f"
if lowered.endswith("ies") and len(lowered) > 3:
return lowered[:-3] + "y"
if lowered.endswith("s") and len(lowered) > 1:
return lowered[:-1]
return lowered
def _shape_description(pattern: tuple[PatternElement, ...]) -> str:
pieces = []
for element in pattern:
if isinstance(element, Constant):
pieces.append(repr(element.token))
else:
pieces.append(f"<{element.feature_name}:{element.slot_type}[{element.min_width},{element.max_width}]>")
return " ".join(pieces)
def _pattern_element_as_dict(element: PatternElement) -> dict[str, Any]:
return element.as_dict()
def _pattern_element_from_dict(raw: Mapping[str, Any]) -> PatternElement:
if raw["kind"] == "constant":
return Constant(token=str(raw["token"]))
if raw["kind"] == "typed_slot":
return TypedSlot(
feature_name=str(raw["feature_name"]),
slot_type=str(raw["slot_type"]),
min_width=int(raw["min_width"]),
max_width=int(raw["max_width"]),
ignored_prefix_tokens=tuple(str(token) for token in raw.get("ignored_prefix_tokens", ())),
)
raise ValueError(f"unknown pattern element kind: {raw['kind']!r}")
__all__ = [
"Constant",
"DerivedRecognizer",
"TypedSlot",
"derive_recognizer",
"recognize",
]