core/tests/test_relational_reader.py
Shay 6f70e834ff
feat: one-hop sound relational entailment (inverse/symmetric) + capability-index lane (#775)
* feat(determine): one-hop sound relational entailment — inverse/converse + symmetric

Mastery-v2 Step 3 lead capability (core). DETERMINE could perceive relational
structure (16 predicates, breadth-complete) but could not derive the simplest
entailed relational facts — it read only the stored direction. This adds two
SOUND one-hop rules that read a stored edge in its other lawful direction:

  INVERSE/converse   greater_than(a,b)  <=  told less_than(b,a)
  SYMMETRIC          sibling_of(b,a)    <=  told sibling_of(a,b)

Strictly scoped (per review): OPEN-WORLD (asserts only True, never False — the
answer=False path stays unbuilt, INV-30's firewall holds), ONE hop (NO transitive
chaining), DECLARED rules only. Ontology/metadata-driven, not prose intuition:

- generate/meaning_graph/relational.py: declarative algebra tables — _INVERSE_PAIRS
  (the converse edges; the pack carries no inverse metadata) + INVERSE_OF (derived,
  an involution) + SYMMETRIC_PREDICATES (mirrors the pack's graph.edge.symmetric tag).
  load_relational_pack_symmetric() reads the pack ontology; a test pins the constant
  equal to it (no silent divergence).
- generate/determine/determine.py: _relational_one_hop() between direct entailment
  and transitive subsumption; new Determined.rule provenance field (direct/inverse/
  symmetric/subsumption) — replay-safe (render reads only basis; trace hashes surface).

wrong=0 confuser block proves the rule cannot over-fire: less_than is not
self-inverse, sibling_of does not imply parent_of, greater_than does not imply
equal_to, NO transitive chain (direct or through-inverse), and no answer=False is
ever emitted. An obsolete pin (test_symmetric_converse_is_not_faked, which asserted
the pre-rule direct-only floor) is updated to the new sound behavior and a preserved
asymmetric-converse bite.

Tests: test_determine_relational_inference.py (13, the capability contract) +
updated test_relational_reader.py. Broad relational/determine/grounding +
capability_index baseline sweep: 114 green; baseline digest UNCHANGED (this is a
determination capability; the comprehension lanes are untouched).

REMAINING for the full lead PR (next): the measurement lane — an independent-oracle
evals/relational_inference lane + a capability_index adapter + baseline re-freeze
(breadth 9->10, wrong_total still 0), so the capability registers on the yardstick.

INTEGRATION NOTE (cross-PR): this adds 2 Determined() construction sites (now 4:
direct/inverse/symmetric/subsumption, ALL answer=True). When rebased onto main with
INV-30 (PR #770), update INV-30's test_determine_construction_sites_are_visible count
2 -> 4 and confirm all four assert True. Merge order: #770 -> this.

* feat(capability-index): relational-inference lane — breadth 9->10, wrong=0

Mastery-v2 Step 3 lead, measurement half. Puts the one-hop relational-inference
capability ON the capability index with an independent-oracle lane:

- evals/relational_inference/v1/: cases.jsonl (13 positive: 8 inverse + 5 symmetric,
  authored from the relation algebra INDEPENDENTLY of determine/relational per
  INV-25/27), refusals.jsonl (8 confusers that MUST refuse), provenance.md (incl. the
  honest overlaps_event reader-surface coverage gap).
- evals/comprehension/relational_inference_runner.py: told fact(s) -> determine(query)
  vs gold; refusal = coverage miss, disagreement = wrong (structurally 0 on positives).
- evals/capability_index/adapters.py: comprehension_relational_inference_result added
  to ADAPTERS.
- baseline.json re-frozen: breadth 9 -> 10, capability_score 0.94403 -> 0.949483,
  wrong_total still 0, digest deliberately changed (35dea2b2...).
- tests/test_relational_inference_lane.py: SHA-pinned gold, positive wrong=0 +
  coverage>0, and the confuser wrong=0 BITE (every confuser must refuse).

Also reconciles INV-30 (now in main via #770): determine.py grew from 2 to 3
Determined() construction sites (direct, the shared relational one-hop constructor,
transitive subsumption) — ALL answer=True. test_determine_construction_sites_are_visible
updated 2 -> 3 (correcting the prior commit note's overcount of 4; inverse and symmetric
share the single _relational_determined constructor, so it is ONE new site).

Also fixes a .gitignore gap: ADR-0219 gen-dir checkpoints (engine_state/current,
engine_state/gen-*/) were not ignored (only the old flat engine_state/*.json patterns
were), so runtime checkpoints could be committed by accident.

Verification: new lane 13/0/0 coverage 1.0; capability_baseline digest matches the
re-freeze; INV-30 + architectural invariants green; smoke 99/99.

* docs+test: reconcile stale relational-inference docstrings; assert rule provenance (review #775)

Addresses two review-requested cleanups on #775:

1. Stale docstrings. generate/meaning_graph/relational.py said the symmetric converse
   is "a sound-but-incomplete refusal at DETERMINE" and there is "NO transitive/
   symmetric/rule inference"; generate/determine/determine.py said "symmetric-converse
   questions are Undetermined", "no transitive/symmetric/rule inference", and "asserts
   only answer=True on a direct hit". All false after the one-hop relational algebra
   landed. Updated: the reader does direct-reading only; DETERMINE applies declared
   one-hop inverse/converse + pack-declared symmetric (plus the existing member/subset
   subsumption), still open-world / never answer=False; transitive relational closure,
   negation, and closed-world falsehood remain out of scope.

2. Rule-provenance gold now meaningful. The relational_inference runner compared only
   (answer, predicate, subject, object), ignoring res.rule though the gold carries
   "rule": "inverse"/"symmetric". Added res.rule to the got/gold tuple so a "right
   answer via the wrong rule path" can no longer pass silently. Lane still 13/0/0;
   baseline counts (and digest) unchanged.

Verification: runner 13/0/0; lane + capability_baseline + unit + INV-30 green (22).
2026-06-15 11:39:41 -07:00

290 lines
12 KiB
Python

"""Relational reader — binary-relation NL → pack-named structure, realized & determined.
The reader is the first consumer of ``en_core_relational_predicates_v1``: it maps
``<A> is [the] <connective> <B>`` onto the pack's closed predicate vocabulary, FAIL-CLOSED
(a non-template surface, a negated form, or a lemma absent from the loaded pack REFUSES).
Realize/determine consume it unchanged.
The pinned tests BITE (wrong=0):
- the reader REFUSES a non-template surface and a negated form (no fabricated read);
- the reader REFUSES when the mapped lemma is absent from the passed pack (fail-closed);
- DETERMINE asserts ONLY on direct entailment — a present-but-non-entailing question
and a symmetric CONVERSE both REFUSE (no faked symmetric/transitive inference);
- DETERMINE keeps categorical predicates EXCLUDED (admitting them would be unsound).
"""
from __future__ import annotations
import pytest
from chat.runtime import ChatRuntime
from generate.determine import Determined, Undetermined, determine
from generate.meaning_graph.reader import Comprehension, Refusal, comprehend
from generate.meaning_graph.relational import (
RELATIONAL_PREDICATES,
comprehend_relational,
load_relational_pack_lemmas,
)
from generate.realize import Realized, realize_comprehension
from session.context import SessionContext
_HIGH_INTERVAL = 10**9
_PACK_LEMMAS = load_relational_pack_lemmas()
@pytest.fixture(scope="module")
def vocab_persona():
rt = ChatRuntime(no_load_state=True)
return rt._context.vocab, rt._context.persona
def _ctx(vocab_persona) -> SessionContext:
vocab, persona = vocab_persona
return SessionContext(vocab=vocab, persona=persona, vault_reproject_interval=_HIGH_INTERVAL)
def _read(text: str):
return comprehend_relational(text, _PACK_LEMMAS)
def _tell(text: str, ctx: SessionContext):
return realize_comprehension(_read(text), ctx)
def _ask(text: str, ctx: SessionContext):
return determine(_read(text), ctx)
def _only_relation(text: str):
comp = _read(text)
assert isinstance(comp, Comprehension), comp
assert len(comp.meaning_graph.relations) == 1 and not comp.queries
return comp.meaning_graph.relations[0]
# --------------------------------------------------------------------------- #
# Reader: each relation family reads onto the right pack lemma
# --------------------------------------------------------------------------- #
def test_pack_lemmas_match_grammar() -> None:
# The static grammar and the shipped pack agree — no grammar entry outruns the pack.
assert RELATIONAL_PREDICATES == _PACK_LEMMAS
@pytest.mark.parametrize(
"text,predicate,args",
[
("Alice is the parent of Bob.", "parent_of", ("alice", "bob")),
("Bob is the child of Alice.", "child_of", ("bob", "alice")),
("Three is less than five.", "less_than", ("three", "five")),
("Nine is greater than two.", "greater_than", ("nine", "two")),
("Five is equal to five.", "equal_to", ("five", "five")),
("The box is left of the cup.", "left_of", ("box", "cup")),
("North station is adjacent to south station.", "adjacent_to",
("north_station", "south_station")),
("Monday is before Friday.", "before_event", ("monday", "friday")),
("Lunch is after breakfast.", "after_event", ("lunch", "breakfast")),
],
)
def test_relation_reads_onto_pack_lemma(text, predicate, args) -> None:
rel = _only_relation(text)
assert rel.predicate == predicate
assert rel.arguments == args
assert rel.negated is False
def test_oov_arguments_are_read() -> None:
# Only the PREDICATE is closed-vocabulary; the arguments may be arbitrary (OOV).
rel = _only_relation("Zorptak is the parent of Quxley.")
assert rel.predicate == "parent_of"
assert rel.arguments == ("zorptak", "quxley")
def test_question_form_reads_a_query_not_a_fact() -> None:
comp = _read("Is Alice the parent of Bob?")
assert isinstance(comp, Comprehension)
assert not comp.meaning_graph.relations
assert len(comp.queries) == 1
q = comp.queries[0]
assert q.predicate == "parent_of" and q.arguments == ("alice", "bob") and not q.negated
# --------------------------------------------------------------------------- #
# Reader: fail-closed — never guess (wrong=0 at the comprehension layer)
# --------------------------------------------------------------------------- #
def test_no_connective_refuses() -> None:
res = _read("Alice greets Bob.")
assert isinstance(res, Refusal) and res.reason == "no_relational_template"
def test_no_copula_refuses() -> None:
res = _read("Alice parent of Bob.")
assert isinstance(res, Refusal) and res.reason == "no_relational_template"
def test_negated_form_refuses() -> None:
# "is not the parent of" breaks the structural copula (the copula is not adjacent
# to the connective) and "not" is reserved — either way it REFUSES; negation never
# reads as a positive relation.
res = _read("Alice is not the parent of Bob.")
assert isinstance(res, Refusal)
assert res.reason in {"no_relational_template", "reserved_word_in_np", "incomplete_relation"}
def test_incomplete_relation_refuses() -> None:
res = _read("Alice is the parent of.")
assert isinstance(res, Refusal) and res.reason == "incomplete_relation"
@pytest.mark.parametrize(
"text",
[
"Carol is the sibling of Dan during school.", # trailing temporal qualifier
"Alice is the parent of Bob during the trip.", # (the asymmetric twin — now same verdict)
"Five is equal to ten during testing.",
"X is left of Y inside Z.", # two relational structures in one clause
"Noon is during overlaps dusk.",
],
)
def test_trailing_relational_structure_refuses(text) -> None:
# A connective token leaking into an argument slot means unparsed relational
# structure -> REFUSE, never fabricate a compound entity ("dan_during_school").
# Without the guard the refusal net is accidental (fires only on reader._RESERVED),
# so "… during the trip" refuses while "… during school" fabricates — that
# comprehension-layer wrong=0 hole is what this bites.
res = _read(text)
assert isinstance(res, Refusal)
assert res.reason in {"extra_relational_structure", "reserved_word_in_np"}
@pytest.mark.parametrize(
"text",
[
"Monday before Friday is.", # dangling copula — not a fact
"Is Monday before Friday.", # question-shaped statement (period, not '?')
],
)
def test_dangling_copula_refuses(text) -> None:
# The copula must sit adjacent to the connective; a stray/dangling "is" must not
# let an ungrammatical surface read as a stored fact.
res = _read(text)
assert isinstance(res, Refusal) and res.reason == "no_relational_template"
def test_no_fabricated_entity_enters_realized_memory(vocab_persona) -> None:
# The damage the misread caused was a fabricated entity persisted as realized
# knowledge. Confirm the hazardous input realizes NOTHING.
ctx = _ctx(vocab_persona)
told = realize_comprehension(_read("Alice is the parent of Bob during school."), ctx)
from generate.realize import NotRealized
assert isinstance(told, NotRealized) and told.reason == "refusal"
# and the fabricated entity is not recallable
assert isinstance(_ask("Is Alice the parent of bob_during_school?", ctx), Undetermined)
def test_fail_closed_on_pack_membership_bites() -> None:
# The fail-closed gate: a grammar lemma absent from the PASSED pack must refuse,
# even though the static grammar maps it. Break the pack -> the read must refuse.
starved_pack = _PACK_LEMMAS - {"parent_of"}
res = comprehend_relational("Alice is the parent of Bob.", starved_pack)
assert isinstance(res, Refusal) and res.reason == "relational_lemma_not_in_pack"
# a different relation still reads through the same starved pack (only parent_of gone)
ok = comprehend_relational("Three is less than five.", starved_pack)
assert isinstance(ok, Comprehension)
# --------------------------------------------------------------------------- #
# End-to-end: realize a relational fact, then determine it (as-told, never verified)
# --------------------------------------------------------------------------- #
def test_realize_then_determine_relation(vocab_persona) -> None:
ctx = _ctx(vocab_persona)
told = _tell("Alice is the parent of Bob.", ctx)
assert isinstance(told, Realized) and told.created
res = _ask("Is Alice the parent of Bob?", ctx)
assert isinstance(res, Determined)
assert res.answer is True
assert res.basis == "as_told" # SPECULATIVE grounds — never "verified"
assert res.predicate == "parent_of"
assert res.subject == "alice" and res.object == "bob"
def test_distinct_relations_about_one_subject_stay_distinct(vocab_persona) -> None:
# R1 relation-space recall: two facts about "alice" collide on the field versor but
# are keyed apart structurally; each determines independently.
ctx = _ctx(vocab_persona)
_tell("Alice is the parent of Bob.", ctx)
_tell("Alice is the sibling of Carol.", ctx)
assert isinstance(_ask("Is Alice the parent of Bob?", ctx), Determined)
assert isinstance(_ask("Is Alice the sibling of Carol?", ctx), Determined)
# --------------------------------------------------------------------------- #
# wrong=0 BITE — direct + DECLARED one-hop relational rules only; no fabrication
# --------------------------------------------------------------------------- #
def test_present_but_non_entailing_refuses(vocab_persona) -> None:
# A record about "alice" exists, but not the asked pair -> REFUSE, never assert.
ctx = _ctx(vocab_persona)
_tell("Alice is the parent of Bob.", ctx)
res = _ask("Is Alice the parent of Carol?", ctx)
assert isinstance(res, Undetermined) and res.reason == "not_entailed"
def test_symmetric_converse_now_soundly_determines(vocab_persona) -> None:
# sibling_of IS symmetric (pack-declared ``graph.edge.symmetric``): the converse now
# SOUNDLY determines via the one-hop symmetric rule (mastery-v2 Step 3) — a DECLARED
# sound rule, not a fake. The stored direction still determines directly. (Was
# test_symmetric_converse_is_not_faked, which pinned the pre-rule direct-only floor.)
ctx = _ctx(vocab_persona)
_tell("Alice is the sibling of Bob.", ctx)
direct = _ask("Is Alice the sibling of Bob?", ctx)
assert isinstance(direct, Determined) and direct.rule == "direct" # stored dir
converse = _ask("Is Bob the sibling of Alice?", ctx)
assert isinstance(converse, Determined) and converse.answer is True
assert converse.rule == "symmetric"
def test_asymmetric_converse_is_not_faked(vocab_persona) -> None:
# The wrong=0 bite, preserved: an ASYMMETRIC predicate's same-predicate converse is
# NEVER faked. parent_of is not symmetric — parent_of(alice, bob) does NOT entail
# parent_of(bob, alice). (Its sound converse is the DIFFERENT predicate child_of.)
ctx = _ctx(vocab_persona)
_tell("Alice is the parent of Bob.", ctx)
converse = _ask("Is Bob the parent of Alice?", ctx)
assert isinstance(converse, Undetermined)
def test_ungrounded_relation_refuses(vocab_persona) -> None:
ctx = _ctx(vocab_persona)
res = _ask("Is Alice the parent of Bob?", ctx)
assert isinstance(res, Undetermined) and res.reason == "ungrounded"
def test_subset_supported_other_categoricals_excluded(vocab_persona) -> None:
# C admits `subset` (direct + transitive subsumption is sound); the OTHER
# categoricals (disjoint/intersects/some_not) stay EXCLUDED — their truth is not a
# stored-pair lookup and there is no sound is-a chain for them.
from generate.meaning_graph.model import Entity, MeaningGraph, MeaningSpan
from generate.meaning_graph.reader import Query
ctx = _ctx(vocab_persona)
subset_q = comprehend("Are all men mortals?")
assert isinstance(subset_q, Comprehension) and subset_q.queries[0].predicate == "subset"
# subset is supported now: nothing realized → ungrounded (NOT unsupported_query)
assert determine(subset_q, ctx).reason == "ungrounded"
# a disjoint query is still refused as out-of-supported-set
span = MeaningSpan("input", 0, 3, "x y")
graph = MeaningGraph(
entities=(Entity("a", "a", span, "class"), Entity("b", "b", span, "class")),
relations=(),
)
disjoint_q = Comprehension(meaning_graph=graph, queries=(Query("disjoint", ("a", "b"), span),))
assert determine(disjoint_q, ctx).reason == "unsupported_query"