core/vocab/manifold.py
Shay ab4c7cb0c3
feat(epistemic): Phase 3 state tagging spine (#220)
* feat(epistemic): add first-class state enums

* feat(epistemic): tag TurnEvent with state axes

* feat(epistemic): serialize turn state axes

* feat(packs): tag curated and inferred unit entries

* feat(epistemic): expose word-level state on manifold

* feat(epistemic): expose vault status mapping

* feat(epistemic): preserve pack entry states through compiler

* test(epistemic): cover phase 3 state tagging spine

* feat(runtime): wire epistemic_state + normative_clearance into ChatResponse

Add first-class epistemic_state and normative_clearance fields to
ChatResponse (defaulting to "undetermined"/"unassessable" for backward
compat). Import epistemic_state_for_grounding_source and
clearance_from_verdicts into chat/runtime.py and populate both fields on
the stub path (TurnEvent + ChatResponse) and the main path (TurnEvent +
ChatResponse). Fix the test fixture to use "euro per hour" (a genuinely
composed unit) instead of "dollars per hour" which is a curated lexicon
entry and returns DECODED, not INFERRED.

* test(cognition): update term_capture_rate baseline from 0.9167 to 1.0

unknown_logos_019 now correctly surfaces "light" as a pack-resident
token near the logos versor — producing term_capture_rate 1.0 on both
main and Phase 3. The 0.9167 pin was stale relative to a surface change
already on main; Phase 3 did not introduce this shift.
2026-05-24 11:26:06 -07:00

301 lines
12 KiB
Python

"""
VocabManifold — the geometric vocabulary.
Each word is a versor in Cl(4,1). nearest(F) finds the closest word
by CGA inner product — no cosine similarity, no ANN index.
Invariant: every stored versor must satisfy the full Cl(4,1) unit-versor
condition V * reverse(V) ≈ ±1. This rejects non-scalar construction residue,
not merely scalar grade-norm drift, and is enforced at insertion time in
add() and at replacement time in update().
Normalization doctrine for this module:
- Raw coordinate vectors (e.g. from external embeddings) must be
lifted via unitize_versor() (algebra/versor.py) BEFORE calling add().
- This module does not call any normalization function internally.
- Rotor construction between word-versors is NOT a vocabulary concern.
Use algebra.rotor.word_transition_rotor(A, B) when a transition
operator is needed in field or generation logic.
Indexed access:
get_versor_at(idx) — returns a copy of the stored versor by integer index.
get_word_at(idx) — returns the word string by integer index.
index_of(word) — returns the integer index for a stored word.
These are the primitives generation uses; VocabManifold does not build
operators. Algebra builds operators. Vocab stores points.
Hot path: nearest() routes cga_inner through algebra.backend, which
dispatches to the Rust extension when available.
"""
import numpy as np
from algebra.backend import cga_inner
from algebra.cl41 import geometric_product, reverse
from algebra.versor import versor_unit_residual
from core.epistemic_state import EpistemicState, coerce_epistemic_state
from core.physics.energy import EnergyProfile
from core.physics.valence import ValenceBundle
from language_packs.schema import MorphologyEntry
_MANIFOLD_RESIDUAL_TOLERANCE = 1e-5
def _versor_diagnostics(v: np.ndarray) -> tuple[float, float, float]:
product = geometric_product(v, reverse(v))
scalar = float(product[0])
residue = product.copy()
residue[0] = 0.0
residue_norm = float(np.linalg.norm(residue))
residual = versor_unit_residual(v, allow_negative=True)
return residual, scalar, residue_norm
def _assert_manifold_versor(word: str, versor: np.ndarray, *, replacement: bool = False) -> None:
residual, scalar, residue_norm = _versor_diagnostics(versor)
if residual > _MANIFOLD_RESIDUAL_TOLERANCE:
noun = "replacement versor" if replacement else "versor"
action = "Call algebra.versor.unitize_versor() before update()." if replacement else (
"If lifting from a raw array, call algebra.versor.unitize_versor() first."
)
raise ValueError(
f"Word '{word}': {noun} residual {residual:.4e} exceeds "
f"{_MANIFOLD_RESIDUAL_TOLERANCE:.1e}; scalar={scalar:.4f}, "
f"non_scalar_residue={residue_norm:.4e}. Pass a clean Cl(4,1) "
f"unit versor satisfying V*reverse(V)≈±1. {action}"
)
class VocabManifold:
def __init__(self):
self._words: list[str] = []
self._versors: list[np.ndarray] = [] # each shape (32,), unit-versor ±1
self._morphology_by_word: dict[str, MorphologyEntry] = {}
self._language_by_word: dict[str, str] = {}
self._energy_by_word: dict[str, EnergyProfile] = {}
self._valence_by_word: dict[str, ValenceBundle] = {}
self._epistemic_state_by_word: dict[str, str] = {}
self._transient_words: set[str] = set()
self._unknown_token_log: list[dict[str, object]] = []
def add(
self,
word: str,
versor: np.ndarray,
morphology: MorphologyEntry | None = None,
language: str | None = None,
energy: EnergyProfile | None = None,
valence: ValenceBundle | None = None,
epistemic_state: str | EpistemicState | None = None,
) -> None:
"""
Register a word-versor pair.
Enforces the Cl(4,1) manifold invariant: V * reverse(V) must be
approximately +1 or -1 as a full multivector residual, not merely
in its scalar component. This rejects raw coordinates, external
embeddings, and dirty construction products.
If your source is a raw float array, call
algebra.versor.unitize_versor() first — that is the construction-time
algebra primitive. Do not call normalize_to_versor() directly;
that function is reserved for the injection gate.
``epistemic_state`` defaults to DECODED for compiled pack entries.
The compiler can override this when a lexical row's review status
maps to another ratified state.
Raises:
ValueError: if the full unit-versor residual is not satisfied.
"""
v = np.asarray(versor, dtype=np.float32).copy()
_assert_manifold_versor(word, v)
self._words.append(word)
self._versors.append(v)
self._epistemic_state_by_word[word] = coerce_epistemic_state(
epistemic_state,
default=EpistemicState.DECODED,
).value
resolved_language = language or (morphology.language if morphology is not None else None)
if resolved_language:
self._language_by_word[word] = resolved_language
if morphology is not None:
self._morphology_by_word[word] = morphology
if energy is not None:
self._energy_by_word[word] = energy
if valence is not None:
self._valence_by_word[word] = valence
def insert_transient(self, word: str, versor: np.ndarray) -> None:
"""
Register a session-local ad hoc word-versor pair.
Transient entries live only in this manifold instance. They use the
same storage as compiled entries so nearest() and get_versor() remain
exact manifold operations, but no language pack persistence path ever
sees them.
"""
if word in self._words and word not in self._transient_words:
raise ValueError(f"Word '{word}' already exists as a compiled manifold entry.")
if word in self._transient_words:
self.update(word, versor)
return
self.add(word, versor, epistemic_state=EpistemicState.UNVERIFIED_NOVEL)
self._transient_words.add(word)
def is_transient(self, word: str) -> bool:
"""Return True when word was inserted as a session-local transient."""
return word in self._transient_words
def morphology_entries(self) -> tuple[MorphologyEntry, ...]:
"""Return morphology entries carried by compiled manifold surfaces."""
return tuple(self._morphology_by_word.values())
def record_unknown_token(
self,
token: str,
root_used: str,
operators_applied: tuple[str, ...],
versor_condition_score: float,
) -> None:
"""Append an audit record for gate-constructed unknown-token grounding."""
self._unknown_token_log.append(
{
"token": token,
"root_used": root_used,
"operators_applied": operators_applied,
"versor_condition_score": versor_condition_score,
}
)
@property
def unknown_token_log(self) -> tuple[dict[str, object], ...]:
"""Session-local audit trail of ad hoc unknown-token constructions."""
return tuple(dict(entry) for entry in self._unknown_token_log)
def update(self, word: str, versor: np.ndarray) -> None:
"""
Replace the versor for an existing word in-place.
Used by the alignment correction pass after compilation to nudge
cross-language aligned pairs toward each other without rebuilding
the full manifold. The full unit-versor residual is enforced
identically to add().
Raises:
KeyError: if the word is not already in the manifold.
ValueError: if the full unit-versor residual is not satisfied.
"""
try:
idx = self._words.index(word)
except ValueError:
raise KeyError(f"Word '{word}' not in vocabulary; use add() for new entries.")
v = np.asarray(versor, dtype=np.float32).copy()
_assert_manifold_versor(word, v, replacement=True)
self._versors[idx] = v
def get_versor(self, word: str) -> np.ndarray:
"""Look up a word's versor by string. Raises KeyError if not found."""
try:
idx = self._words.index(word)
return self._versors[idx].copy()
except ValueError:
raise KeyError(f"Word '{word}' not in vocabulary.")
def get_versor_at(self, idx: int) -> np.ndarray:
"""
Return a copy of the stored versor at integer index.
This is the indexed access primitive for generation — algebra
uses these points to construct transition operators.
"""
return self._versors[idx].copy()
def get_word_at(self, idx: int) -> str:
"""Return the word string at integer index."""
return self._words[idx]
def index_of(self, word: str) -> int:
"""Return the integer index for a stored word. Raises KeyError if missing."""
try:
return self._words.index(word)
except ValueError:
raise KeyError(f"Word '{word}' not in vocabulary.")
def morphology_for_word(self, word: str) -> MorphologyEntry | None:
"""Return structured morphology for a stored surface, if the pack provided it."""
return self._morphology_by_word.get(word)
def energy_for_word(self, word: str) -> EnergyProfile | None:
"""Return ADR-0006 energy profile for a stored surface, when available."""
return self._energy_by_word.get(word)
def valence_for_word(self, word: str) -> ValenceBundle | None:
"""Return ADR-0007 valence bundle for a stored surface, when available."""
return self._valence_by_word.get(word)
def epistemic_state_for_word(self, word: str) -> str:
"""Return the ratified epistemic state for a stored surface.
Missing state metadata defaults to UNDETERMINED rather than
silently treating the entry as decoded.
"""
if word not in self._words:
raise KeyError(f"Word '{word}' not in vocabulary.")
return self._epistemic_state_by_word.get(word, EpistemicState.UNDETERMINED.value)
def language_for_word(self, word: str) -> str | None:
"""Return the language code for a stored surface, if known."""
morphology = self._morphology_by_word.get(word)
if morphology is not None:
return morphology.language
return self._language_by_word.get(word)
def indices_for_language(self, lang: str) -> np.ndarray:
"""Return manifold indices whose language matches lang."""
matches = [
idx
for idx, word in enumerate(self._words)
if self.language_for_word(word) == lang
]
return np.asarray(matches, dtype=np.int64)
def nearest(
self,
F: np.ndarray,
exclude_idx: int = -1,
exclude_indices: set[int] | frozenset[int] | None = None,
candidate_indices: np.ndarray | list[int] | tuple[int, ...] | None = None,
) -> tuple[str, int]:
"""
Find the word whose versor is closest to F by CGA inner product.
Returns (word, index). O(|vocab|), exact, no approximation.
cga_inner(X, Y) = -d^2 / 2 for null vectors: maximizing = minimizing distance.
Hot path: cga_inner routes through algebra.backend.
"""
blocked = set(exclude_indices or ())
if exclude_idx >= 0:
blocked.add(exclude_idx)
if candidate_indices is None:
candidates = range(len(self._versors))
else:
candidates = [int(idx) for idx in candidate_indices]
best_score = -np.inf
best_idx = -1
# Strict `>` is load-bearing: on ties, the first candidate in iteration
# order wins. ADR-0024 inner-loop re-selection relies on this for
# deterministic rejected_attempts ordering across runs.
for i in candidates:
if i in blocked:
continue
score = cga_inner(F, self._versors[i])
if score > best_score:
best_score = score
best_idx = i
if best_idx < 0:
raise ValueError("No candidate word available after exclusions.")
return self._words[best_idx], best_idx
def __len__(self) -> int:
return len(self._words)