a6be96410c
_orient_result_to_anchor used np.dot (Euclidean dot product) alongside cga_inner to decide hemisphere flips. When CGA inner was positive (correct hemisphere) but Euclidean was negative, the flip negated CGA alignment — making correctly-oriented fields rank last in vault recall. Changes: - Move hemisphere check into finalize_turn so all paths (ChatRuntime, SessionContext.respond) get consistent protection. - Use CGA inner product only, removing the forbidden Euclidean metric. - Remove _orient_result_to_anchor (subsumed by finalize_turn). - Remove SessionContext.arespond (dead code, no callers). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
108 lines
3.7 KiB
Python
108 lines
3.7 KiB
Python
from __future__ import annotations
|
|
|
|
import numpy as np
|
|
|
|
from algebra.backend import cga_inner
|
|
from algebra.versor import unitize_versor
|
|
from session.context import SessionContext
|
|
from vocab.manifold import VocabManifold
|
|
|
|
|
|
def _positive_unit_reflector(seed: int) -> np.ndarray:
|
|
rng = np.random.default_rng(seed)
|
|
vec4 = rng.standard_normal(4).astype(np.float32)
|
|
norm4 = float(np.linalg.norm(vec4))
|
|
if norm4 < 1e-6:
|
|
vec4[0] = 1.0
|
|
norm4 = 1.0
|
|
|
|
vec = np.zeros(5, dtype=np.float32)
|
|
vec[:4] = vec4
|
|
vec[4] = 0.25 * norm4 * np.tanh(float(rng.standard_normal()))
|
|
|
|
mv = np.zeros(32, dtype=np.float32)
|
|
mv[1:6] = vec
|
|
return unitize_versor(mv)
|
|
|
|
|
|
def _random_rotor(seed: int) -> np.ndarray:
|
|
"""Build a random even-grade versor (rotor) for CGA inner-product comparison.
|
|
|
|
Field states are even-grade (grade 0+2+4). Reflectors are grade-1 and
|
|
orthogonal under cga_inner, so we need rotors to get nonzero scores.
|
|
"""
|
|
a = _positive_unit_reflector(seed)
|
|
b = _positive_unit_reflector(seed + 10000)
|
|
from algebra.cl41 import geometric_product as gp
|
|
rotor = gp(a, b)
|
|
return unitize_versor(rotor)
|
|
|
|
|
|
def _vocab() -> VocabManifold:
|
|
vocab = VocabManifold()
|
|
vocab.add("logos", _positive_unit_reflector(1))
|
|
vocab.add("arche", _positive_unit_reflector(2))
|
|
vocab.add("pneuma", _positive_unit_reflector(3))
|
|
vocab.add("truth", _positive_unit_reflector(4))
|
|
vocab.add("nous", _positive_unit_reflector(5))
|
|
return vocab
|
|
|
|
|
|
def _farther_unrelated(result_F: np.ndarray, prompt_F: np.ndarray, start_seed: int) -> np.ndarray:
|
|
prompt_score = cga_inner(result_F, prompt_F)
|
|
for seed in range(start_seed, start_seed + 2048):
|
|
candidate = _random_rotor(seed)
|
|
if prompt_score > cga_inner(result_F, candidate):
|
|
return candidate
|
|
raise AssertionError("Could not construct a deterministic farther unrelated versor.")
|
|
|
|
|
|
def test_finalize_turn_enforces_cga_hemisphere_consistency() -> None:
|
|
"""Vault entries must share the anchor's CGA hemisphere for recall to rank correctly."""
|
|
from generate.result import GenerationResult
|
|
from field.state import FieldState
|
|
|
|
vocab = _vocab()
|
|
session = SessionContext(vocab=vocab)
|
|
session.ingest(["logos"])
|
|
anchor = session._anchor_field.copy()
|
|
|
|
anti_F = -session.state.F
|
|
anti_state = FieldState(
|
|
F=anti_F,
|
|
node=session.state.node,
|
|
step=session.state.step,
|
|
holonomy=session.state.holonomy,
|
|
energy=session.state.energy,
|
|
valence=session.state.valence,
|
|
)
|
|
anti_result = GenerationResult(tokens=("arche",), final_state=anti_state, vault_hits=0)
|
|
assert cga_inner(anti_F, anchor) < 0.0, "precondition: anti-hemisphere field"
|
|
|
|
session.finalize_turn(anti_result, dialogue_role="assert")
|
|
|
|
assert cga_inner(session.state.F, anchor) >= 0.0, (
|
|
"finalize_turn must flip anti-hemisphere fields to keep vault recall consistent"
|
|
)
|
|
|
|
|
|
def test_repeated_prompt_accumulates_field_and_stays_prompt_coherent() -> None:
|
|
session = SessionContext(vocab=_vocab())
|
|
prompt = ["logos", "arche"]
|
|
|
|
initial = session.ingest(prompt)
|
|
first = session.respond(max_tokens=4)
|
|
|
|
second_prompt_state = session.ingest(prompt)
|
|
assert not np.array_equal(second_prompt_state.F, initial.F)
|
|
|
|
second = session.respond(max_tokens=4)
|
|
|
|
assert second.tokens != first.tokens
|
|
assert not np.array_equal(second.final_state.F, first.final_state.F)
|
|
|
|
for i, result in enumerate((first, second)):
|
|
random_unrelated = _farther_unrelated(result.final_state.F, initial.F, 11 + (i * 64))
|
|
prompt_score = cga_inner(result.final_state.F, initial.F)
|
|
random_score = cga_inner(result.final_state.F, random_unrelated)
|
|
assert prompt_score > random_score
|