core/generate/realize/realize.py
Shay 8edafd04ac feat(determine): idle deductive consolidation — the loop learns from determined facts (Step D)
CLOSE the autonomous-loop spine: when idle, the engine consolidates each
soundly-derived determination back into the held self, so the next determine()
reaches it directly and can chain one hop further. The directly-answerable set
climbs monotonically across idle ticks to the deductive-closure fixed point.

Mechanism — generate/determine/consolidate.py::consolidate_once runs ONE
semi-naive layer of the member/subset deductive closure (member∘subset → member,
subset∘subset → subset; NEVER member∘member — instance-of is not transitive).
Each one-hop conclusion not yet realized is VERIFIED by the sound+complete
proof_chain ROBDD (reusing C's single verifier _verify_subsumption) and written
back via generate/realize::realize_derived as a SPECULATIVE realized record
carrying derived-provenance (premise structure_keys + rule + the ENTAILED
verdict). idle_tick gains a consolidation pass gated by the new
config.consolidate_determinations (default OFF); IdleTickResult.facts_consolidated
reports the layer.

Invariants held:
- wrong=0 — every consolidated fact is a sound-rule conclusion confirmed by the
  sound+complete decider; member∘member is structurally unreachable (a member fact
  is only ever extended by a subset edge). _verify_subsumption now refuses a
  mislabeled/wrong-arity path (belt-and-suspenders now that consolidation is a
  second caller), so the fallacy cannot be laundered through a corrupted chain.
- honesty — a fact derived from SPECULATIVE premises stays SPECULATIVE / as-told;
  the soundness of the inference never upgrades the standing of the premises.
  COHERENT is never minted.
- teaching-safety — SESSION memory (immediate), an extension of the realize path;
  NOT corpus mutation and NOT coupled to proposals. The HITL path is untouched.
- determinism/replay — pure function of the realized set; sorted write order;
  derived structure_key identical to a told fact's; provenance round-trips through
  the Shape B+ snapshot (consolidated facts resume the SAME life across reboot).
- no new normalization — writes reuse the INV-21-allowed vault writer;
  algebra/versor.py keeps closure.

Falsification — evals/determination_closure: a frozen replay seeds a deep is-a
chain and runs idle ticks; asserts the closure climbs monotonically to a complete
fixed point (no-op final tick), wrong=0 (member∘member canary never derived, no
fabricated membership), all derived facts SPECULATIVE, and every derived record
re-verifies ENTAILED from its recorded premises.

Verified green: smoke, runtime, cognition, architectural invariants, plus the new
D unit + lane tests and the determine/realize/persistence regression net. Five-lens
adversarial review: 4 lenses held; the 5th (normalization) was a misattribution
(pre-existing vault reproject, triggered identically by the merged realize path,
on sanctioned null-vector storage). Design + findings: docs/analysis/
D-close-consolidation-design-2026-06-06.md.
2026-06-06 12:28:09 -07:00

425 lines
18 KiB
Python

"""REALIZE slice R0 — one told fact becomes a realized vault entry.
The boundary that turns comprehension from an EVAL ARTIFACT into accumulating living
knowledge: a comprehended declarative fact is integrated into the held self as a
SPECULATIVE, provenance-rich vault entry that survives reboot and recalls exactly.
A realized record is NOT a new store — it is a structured vault entry
``(versor, metadata)`` so it inherits exact ``cga_inner`` recall, ``EpistemicStatus``
stamping, and bit-exact Shape B+ persistence for free (see
``docs/analysis/REALIZE-scope-2026-06-06.md``).
Slice R0 was deliberately boring: one told fact, SPECULATIVE only, in-vocab subject
only, single non-negated declarative relation. The in-vocab restriction is now LIFTED
(OOV subjects realize too): OOV grounding is deterministic, reboot-stable, and injective
(#591), and correctness rests on the structural key, not the versor. Slice R1 (this
module + ``recall.py``) adds the relation-space KEY R0 lacked: ordered
``relation_arguments`` + a span-free ``structure_key`` so
distinct facts about one subject stay distinct (they collide on the field versor), and
``recall_realized`` retrieves them by exact structural metadata. Dedup is now span-free
(``structure_key``), guarded by an ambiguous-entity-name refusal (wrong=0). Explicitly
still OUT: COHERENT promotion, teaching-loop proposals, the quantitative
``binding_graph`` path (R1c), trace-folding (see
``docs/analysis/REALIZE-R1-DETERMINE-scope-2026-06-06.md``).
wrong=0 at this layer: a ``Refusal`` (or any ineligible / ungrounded input) realizes
NOTHING — it is returned as ``NotRealized(reason)``, never coerced into a vault entry.
"""
from __future__ import annotations
from dataclasses import dataclass
from typing import Any
import numpy as np
from formation.hashing import sha256_of
from generate.meaning_graph.reader import Comprehension, Refusal
from session.context import SessionContext
from teaching.epistemic import EpistemicStatus
# The known failure modes of the session embedding path (probe_ingest):
# RuntimeError on cross-turn versor-condition violation; Key/Index/ValueError on a
# degenerate grounding. Any of them is a clean no-op, never a crashed turn.
_GROUNDING_FAILURES = (RuntimeError, KeyError, IndexError, ValueError)
#: R0 realizes only the neutral MeaningGraph substrate. The record shape already
#: admits a second substrate (the quantitative binding-graph) via ``structure_kind``.
_STRUCTURE_KIND_MEANING_GRAPH = "meaning_graph"
@dataclass(frozen=True, slots=True)
class Derivation:
"""Provenance of a DERIVED realized fact (Step D consolidation).
A derived fact is the conclusion of a SOUND is-a rule (``member_subset`` or
``subset_subset``) over premises that were themselves realized (told or previously
consolidated), confirmed by the sound+complete proof_chain ROBDD. ``rule`` names
the inference; ``premise_structure_keys`` are the span-free identities of the
premise records (so a replay re-fetches them and re-verifies the chain);
``verdict`` is the decider's outcome — always ``"entailed"`` (a non-entailed
candidate is never consolidated). This makes the soundness claim MEANINGFULLY FAIL
on replay (Schema-Defined Proof Obligations): re-deriving from the recorded
premises must still entail the fact.
"""
rule: str
premise_structure_keys: tuple[str, ...]
verdict: str
def as_dict(self) -> dict[str, Any]:
return {
"rule": self.rule,
"premise_structure_keys": list(self.premise_structure_keys),
"verdict": self.verdict,
}
@classmethod
def from_dict(cls, payload: dict[str, Any]) -> "Derivation":
return cls(
rule=payload["rule"],
premise_structure_keys=tuple(payload.get("premise_structure_keys", [])),
verdict=payload.get("verdict", ""),
)
@dataclass(frozen=True, slots=True)
class RealizedRecord:
"""The realized-knowledge record (mirrors the stored vault metadata)."""
structure_kind: str
structure_canonical: str
relation_predicate: str
#: Ordered argument identities (subject first) — the relation-space key R0's
#: sorted ``entity_names`` discards. Distinct facts about one subject differ
#: here even though they collide on the field versor.
relation_arguments: tuple[str, ...]
entity_names: tuple[str, ...]
source_span: str
content_hash: str
#: Span-FREE structural identity (predicate + negated + ordered argument
#: ids). Idempotency dedups on this, so the same proposition told from a
#: different source/offset collapses (which the span-inclusive ``content_hash``
#: could not). ``content_hash`` is retained for provenance + ``replay_hash``.
structure_key: str
replay_hash: str
epistemic_status: str
tier: str
#: LIVE deque position at recall/write time — authoritative in the unbounded
#: session tier; provenance-only under bounded-vault eviction.
vault_index: int
#: Step D — a TOLD fact is ``derived=False`` (``derivation=None``). A consolidated
#: fact (the conclusion of a sound is-a rule, proof_chain-verified) is
#: ``derived=True`` with a :class:`Derivation`. Defaults preserve the told-record
#: shape (and bit-exact persistence) for every pre-D caller.
derived: bool = False
derivation: "Derivation | None" = None
@dataclass(frozen=True, slots=True)
class Realized:
"""The fact is realized. ``created`` is False on an idempotent dedup hit
(the fact was already realized this session — no new vault entry written)."""
record: RealizedRecord
created: bool
@dataclass(frozen=True, slots=True)
class NotRealized:
"""Nothing was realized (no vault write). ``reason`` is for audit, not control."""
reason: str
def realize_comprehension(
comprehension: Comprehension | Refusal, ctx: SessionContext
) -> Realized | NotRealized:
"""Realize a comprehended declarative fact into ``ctx``'s vault, or decline.
Eligibility (R0): a ``Comprehension`` (not a ``Refusal``) with NO queries and
EXACTLY ONE non-negated relation whose subject argument grounds in-vocabulary.
Everything else is a typed ``NotRealized`` — no vault write (wrong=0).
"""
if isinstance(comprehension, Refusal):
return NotRealized("refusal")
if not isinstance(comprehension, Comprehension):
return NotRealized("not_a_comprehension")
if comprehension.queries:
return NotRealized("query_bearing") # a question is recall, not intake
graph = comprehension.meaning_graph
if len(graph.relations) != 1:
return NotRealized("not_single_relation") # multi-relation intake is ambiguous in R0
rel = graph.relations[0]
if rel.negated:
return NotRealized("negated_relation")
if not rel.arguments:
return NotRealized("empty_relation") # defensive — arity>=1 by construction
# wrong=0 defense (R1b): the MeaningGraph model permits distinct entities to
# share a name (only ``entity_id`` is enforced unique; today the reader sets
# entity_id == name, so this is latent, not live). A name-keyed structural
# identity would collapse a converse/homonym fact, dropping a genuinely
# distinct proposition — so refuse the ambiguous input now, before a future
# reader makes it load-bearing.
names = [e.name for e in graph.entities]
if len(set(names)) != len(names):
return NotRealized("ambiguous_entity_names")
subject_id = rel.arguments[0]
entity = next((e for e in graph.entities if e.entity_id == subject_id), None)
if entity is None:
return NotRealized("ungrounded_subject") # defensive — graph invariant
subject_token = entity.name
# OOV subjects ARE realizable (R0's in-vocab gate is lifted). OOV grounding is
# deterministic and reboot-stable, and #591 makes it injective (distinct token
# content -> distinct field point, closure-by-construction). Correctness does not
# rest on the versor anyway: distinct facts stay distinct by their structural key
# + structural recall (recall.py), and reboot-stability rests on the Shape B+
# snapshot of the exact stored bytes — so even a colliding placement never
# confuses recall.
#
# probe_ingest is side-effect-free for an IN-VOCAB token, but for an OOV token it
# mutates the shared vocab via insert_transient — a SESSION-SCOPED transient that
# is NOT serialized into the snapshot (session/context.py: vocab is a shared
# ratified surface, not session state), so it is re-derived deterministically on
# reboot and never affects reboot stability. The field point still composes with
# the CURRENT session state (deterministic GIVEN that state, not purely
# subject-determined). Closure stays algebra/versor.py's job — REALIZE adds none.
# A degenerate construction raises and is caught below -> NotRealized.
try:
field_state = ctx.probe_ingest([subject_token])
except _GROUNDING_FAILURES:
return NotRealized("grounding_failed")
versor = np.asarray(field_state.F, dtype=np.float32)
# Ordered argument identities (subject first). entity_id == name in today's
# reader; the map keeps this correct if a future reader separates them.
name_of = {e.entity_id: e.name for e in graph.entities}
relation_arguments = tuple(name_of.get(a, a) for a in rel.arguments)
structure_canonical = graph.to_canonical_string()
source_span = rel.span.to_canonical_string()
content_hash = sha256_of(structure_canonical) # span-INCLUSIVE: provenance
# Span-FREE structural identity: the asserted proposition, free of source
# offsets. Idempotency dedups on this, so the same fact told from a different
# source/offset collapses (the span-inclusive content_hash could not).
structure_key = sha256_of(
{
"structure_kind": _STRUCTURE_KIND_MEANING_GRAPH,
"predicate": rel.predicate,
"negated": rel.negated,
"arguments": list(relation_arguments),
}
)
status = EpistemicStatus.SPECULATIVE # COHERENT is never a default (ADR-0021)
replay_hash = sha256_of(
{
"content_hash": content_hash,
"source_span": source_span,
"epistemic_status": status.value,
}
)
entity_names = tuple(sorted(e.name for e in graph.entities))
return _realize_structured(
ctx,
structure_kind=_STRUCTURE_KIND_MEANING_GRAPH,
structure_canonical=structure_canonical,
relation_predicate=rel.predicate,
relation_arguments=relation_arguments,
entity_names=entity_names,
source_span=source_span,
content_hash=content_hash,
structure_key=structure_key,
replay_hash=replay_hash,
versor=versor,
status=status,
)
def _realize_structured(
ctx: SessionContext,
*,
structure_kind: str,
structure_canonical: str,
relation_predicate: str,
relation_arguments: tuple[str, ...],
entity_names: tuple[str, ...],
source_span: str,
content_hash: str,
structure_key: str,
replay_hash: str,
versor: np.ndarray,
status: EpistemicStatus,
derivation: "Derivation | None" = None,
) -> Realized:
"""Dedup-or-store a realized record — the single wrong=0 write path shared by
every substrate (meaning_graph relations, binding_graph quantities, and Step D
derived facts).
Idempotency dedups exactly on the span-free ``structure_key`` (callers refuse
ambiguous identity before getting here, so a hit is always the genuinely-same
proposition, never a distinct one). The ``vault.store`` call is the only
mutation; ``iter_metadata`` is the public read-only accessor and ``idx`` is the
LIVE deque position. The epistemic status is whatever the caller declared —
SPECULATIVE, never COHERENT by default (ADR-0021). A ``derivation`` marks a
consolidated (derived) fact; ``None`` is a told fact (the historical shape).
"""
for idx, meta in ctx.vault.iter_metadata():
if meta.get("kind") == "realized" and meta.get("structure_key") == structure_key:
return Realized(record=_record_from_metadata(meta, idx), created=False)
metadata: dict[str, Any] = {
"kind": "realized",
"structure_kind": structure_kind,
"structure_canonical": structure_canonical,
"relation_predicate": relation_predicate,
"relation_arguments": list(relation_arguments),
"entity_names": list(entity_names),
"source_span": source_span,
"content_hash": content_hash,
"structure_key": structure_key,
"replay_hash": replay_hash,
"tier": "session",
}
if derivation is not None:
# Derived-fact provenance is additive — a told fact omits both keys, so its
# on-disk metadata stays byte-identical to the pre-D encoding.
metadata["derived"] = True
metadata["derivation"] = derivation.as_dict()
vault_index = ctx.vault.store(versor, metadata, epistemic_status=status)
return Realized(
record=RealizedRecord(
structure_kind=structure_kind,
structure_canonical=structure_canonical,
relation_predicate=relation_predicate,
relation_arguments=relation_arguments,
entity_names=entity_names,
source_span=source_span,
content_hash=content_hash,
structure_key=structure_key,
replay_hash=replay_hash,
epistemic_status=status.value,
tier="session",
vault_index=vault_index,
derived=derivation is not None,
derivation=derivation,
),
created=True,
)
def realize_derived(
ctx: SessionContext,
*,
predicate: str,
subject: str,
obj: str,
rule: str,
premise_structure_keys: tuple[str, ...],
) -> Realized | NotRealized:
"""Consolidate a SOUNDLY-DERIVED fact into the held self (Step D — CLOSE).
The caller (``generate.determine.consolidate``) has already found and
proof_chain-VERIFIED the is-a chain that entails ``predicate(subject, obj)`` — this
only writes the conclusion as a realized record so the next ``determine`` reaches it
directly. The record is:
- SPECULATIVE / ``as_told`` — a sound INFERENCE never upgrades the STANDING of its
(SPECULATIVE) premises; COHERENT is never minted here (ADR-0021 honesty).
- SESSION memory (immediate) — NOT reviewed/corpus learning; nothing is proposed,
the teaching/review HITL path is untouched (no parallel learning path).
- keyed by the SAME span-free ``structure_key`` a told ``predicate(subject, obj)``
would carry, so it dedups against / is found by the told path identically, and
is recalled by ``recall_realized`` like any other realized fact.
- provenance-rich (``Derivation``): the premise ``structure_key``s + rule + the
ENTAILED verdict, so a replay re-derives and re-verifies (the soundness claim
can MEANINGFULLY FAIL).
Idempotent (dedups on ``structure_key``); returns ``NotRealized`` on any ineligible
input or grounding failure — never a partial / coerced write (wrong=0).
"""
subject = (subject or "").strip()
obj = (obj or "").strip()
predicate = (predicate or "").strip()
if not subject or not obj or not predicate:
return NotRealized("incomplete_derived_fact")
# The derived fact embeds its subject identically to a told fact about that subject
# (same probe_ingest placement). The versor is not the recall key (structural
# metadata is), so a degenerate placement is a clean no-op, never a wrong write.
try:
field_state = ctx.probe_ingest([subject])
except _GROUNDING_FAILURES:
return NotRealized("grounding_failed")
versor = np.asarray(field_state.F, dtype=np.float32)
relation_arguments = (subject, obj)
structure_canonical = f"derived:{predicate}({subject},{obj})"
source_span = f"derived:{rule}"
content_hash = sha256_of(structure_canonical) # span-inclusive provenance
structure_key = sha256_of(
{
"structure_kind": _STRUCTURE_KIND_MEANING_GRAPH,
"predicate": predicate,
"negated": False,
"arguments": list(relation_arguments),
}
)
status = EpistemicStatus.SPECULATIVE # derived from SPECULATIVE premises — as-told
replay_hash = sha256_of(
{
"content_hash": content_hash,
"source_span": source_span,
"epistemic_status": status.value,
}
)
entity_names = tuple(sorted({subject, obj}))
derivation = Derivation(
rule=rule,
premise_structure_keys=tuple(premise_structure_keys),
verdict="entailed",
)
return _realize_structured(
ctx,
structure_kind=_STRUCTURE_KIND_MEANING_GRAPH,
structure_canonical=structure_canonical,
relation_predicate=predicate,
relation_arguments=relation_arguments,
entity_names=entity_names,
source_span=source_span,
content_hash=content_hash,
structure_key=structure_key,
replay_hash=replay_hash,
versor=versor,
status=status,
derivation=derivation,
)
def _record_from_metadata(meta: dict, idx: int) -> RealizedRecord:
"""Reconstruct a RealizedRecord from a stored vault metadata dict."""
derivation_payload = meta.get("derivation")
return RealizedRecord(
structure_kind=meta.get("structure_kind", _STRUCTURE_KIND_MEANING_GRAPH),
structure_canonical=meta.get("structure_canonical", ""),
relation_predicate=meta.get("relation_predicate", ""),
relation_arguments=tuple(meta.get("relation_arguments", [])),
entity_names=tuple(meta.get("entity_names", [])),
source_span=meta.get("source_span", ""),
content_hash=meta.get("content_hash", ""),
structure_key=meta.get("structure_key", ""),
replay_hash=meta.get("replay_hash", ""),
epistemic_status=meta.get("epistemic_status", "speculative"),
tier=meta.get("tier", "session"),
vault_index=idx,
derived=bool(meta.get("derived", False)),
derivation=(
Derivation.from_dict(derivation_payload)
if isinstance(derivation_payload, dict)
else None
),
)