core/teaching/discovery.py
Shay b5ba9b6d6f feat(adr-0064): cross-pack teaching chains + relations_chains_v1 seed (Phase 1.3+1.4)
ADR-0064 is the corpus-layer sibling of ADR-0063.  The teaching-grounded
surface composer was hardcoded to cognition_chains_v1, so kinship CAUSE/
VERIFICATION prompts fell through to the universal disclosure even though
en_core_relations_v1 was mounted on the live runtime (ADR-0063).

Architectural change in chat/teaching_grounding.py:

  - New TeachingCorpusSpec dataclass (corpus_id, path, pack_id).
  - TEACHING_CORPORA tuple registers every active corpus.  Each
    corpus is 1:1-bound to one lexicon pack — cross-domain triples
    deferred per docs/teaching_order.md §5.
  - _load_corpus(spec) loads one corpus with pack-residency scoped
    to its declared pack.
  - _all_chains_index() aggregates across all registered corpora
    (first-match-wins; cognition first preserves byte-identity).
  - _pack_for_corpus(corpus_id) → bound pack lexicon.
  - clear_teaching_caches() atomic cache invalidation.
  - TeachingChain gains corpus_id field → surface tag follows resolving corpus.

Wiring updates:

  - teaching_grounded_surface + teaching_grounded_surface_composed
    consult _all_chains_index; surface tag follows chain.corpus_id.
  - teaching/discovery.py gate uses chat.pack_resolver.is_resolvable
    (any mounted pack) + _all_chains_index (any registered corpus).
  - teaching/replay.py _swap_corpus_path rewrites the registry path
    + clears all teaching caches during the gate's transient phase.
    Active corpus bytes unchanged (replay invariant preserved).
  - evals/learning_loop/run_demo.py scene-5 swap mirrors the new
    pattern so the demo still grounds against transient corpora.

Back-compat preserved: _corpus_index, _CORPUS_PATH, TEACHING_CORPUS_ID
remain cognition-corpus-specific for audit/replay consumers.

Phase 1.4 — relations_chains_v1 seeded with 7 reviewed kinship chains:
  cause_parent_precedes_child
  cause_child_follows_parent
  cause_ancestor_precedes_descendant
  cause_descendant_follows_ancestor
  cause_family_grounds_parent
  verification_child_requires_parent
  verification_descendant_requires_ancestor

5 of 8 relations lemmas covered.  All connectives already humanised.
Strict pack-internal to en_core_relations_v1 (no cross-domain in v1).
Seed pattern matches cognition_chains_v1's original pre-ADR-0055 seed.

Live verification:
  > Why does parent exist?
  parent — teaching-grounded (relations_chains_v1):
  kinship.ascendant.direct; kinship.parent.
  parent precedes child (kinship.descendant.direct).
  grounding_source = teaching

Cognition eval byte-identical to pre-ADR baseline:
  public:  intent 100% / surface 100% / term 91.7% / closure 100%
  holdout: intent 100% / surface 100% / term 83.3% / closure 100%

Lanes green: smoke 67 / cognition 121 / teaching 17 / packs 6 /
runtime 19 / algebra 132 / full 1933 passed.
2026-05-18 16:04:20 -07:00

326 lines
12 KiB
Python

"""ADR-0055 Phase B — DiscoveryCandidate emission from the turn loop.
A ``DiscoveryCandidate`` is **structured evidence** — never a
mutation. When a turn's audit trail satisfies a deterministic
predicate, an entry is emitted to the discovery candidate stream.
Candidates **never** load into the active teaching corpus; the only
path to corpus extension is the review-gated
``TeachingChainProposal`` (Phase C, not yet built).
Trigger set (Phase B lands the first; the others are stubbed in the
``Literal`` so the structure is stable when later phases add them):
- ``would_have_grounded`` — the turn fell through to the universal
"insufficient grounding" disclosure, the classified intent was
``CAUSE`` or ``VERIFICATION``, the subject lemma is in the
ratified cognition pack, and no active chain matched
``(subject, intent)``. A reviewed chain of that subject/intent
would have grounded the turn.
- ``successful_comparison`` — open question §5 in ADR-0055; not
fired in Phase B.
- ``hedge_acknowledged`` — open question §5 in ADR-0055; not
fired in Phase B.
- ``oov_resolved_via_decomp`` — not fired in Phase B.
Determinism contract:
- ``extract_discovery_candidates`` is a pure function of its
inputs.
- ``candidate_id`` is a SHA-256 hash of a canonical JSON encoding
of the candidate's load-bearing fields; identical inputs always
produce the identical id.
- No LLM, no stochastic sampling, no clock-time read.
Trust boundary:
- This module reads pack + corpus indices and a ``TurnEvent``.
It never writes to the corpus, the pack, or runtime state.
- The ``source_turn_trace`` is the upstream ``TurnEvent.trace_hash``
when present; absent that, the empty string. Tying every
candidate to a replayable turn is the load-bearing audit
property.
"""
from __future__ import annotations
import hashlib
import json
from dataclasses import dataclass
from typing import Any, Literal
from generate.intent import IntentTag
# ``chat.pack_grounding`` and ``chat.teaching_grounding`` are
# imported lazily inside ``extract_discovery_candidates`` to break a
# circular import chain when an entry-point (e.g. the CLI) imports
# ``teaching.proposals`` → ``teaching.discovery`` before ``chat``
# has been fully initialized.
DiscoveryTrigger = Literal[
"would_have_grounded",
"successful_comparison",
"hedge_acknowledged",
"oov_resolved_via_decomp",
]
# ADR-0056 Phase C1: typed claim domain for the contemplation loop.
ClaimDomain = Literal["factual", "relational", "evaluative"]
@dataclass(frozen=True, slots=True)
class EvidencePointer:
"""One unit of admissible evidence used by the contemplation loop.
Only three source families admit a pointer: reviewed teaching
corpus chains, ratified pack atoms, and vault entries stamped
``EpistemicStatus.COHERENT``. SPECULATIVE / CONTESTED / FALSIFIED
vault entries contest but do not contribute as evidence.
"""
source: Literal["corpus", "pack", "vault_coherent"]
ref: str
polarity: Literal["affirms", "falsifies"]
epistemic_status: str
def as_dict(self) -> dict[str, Any]:
return {
"source": self.source,
"ref": self.ref,
"polarity": self.polarity,
"epistemic_status": self.epistemic_status,
}
@dataclass(frozen=True, slots=True)
class SubQuestion:
"""One decomposed sub-question + its outcome (ADR-0056 §SubQuestion).
``outcome="gap_recorded"`` is the load-bearing case from Call 1
in ADR-0056: the sub-question could not be decomposed further so
the system records the gap and stops.
"""
sub_id: str
proposed_subject: str
proposed_intent: str
outcome: Literal["grounded", "gap_recorded", "depth_failsafe"]
evidence: tuple[EvidencePointer, ...] = ()
def as_dict(self) -> dict[str, Any]:
return {
"sub_id": self.sub_id,
"proposed_subject": self.proposed_subject,
"proposed_intent": self.proposed_intent,
"outcome": self.outcome,
"evidence": [e.as_dict() for e in self.evidence],
}
@dataclass(frozen=True, slots=True)
class DiscoveryCandidate:
"""Structured evidence that a reviewed chain would have helped.
Phase B emits the Phase-B fields only. ADR-0056 Phase C1 adds
typed contemplation fields (``polarity``, ``claim_domain``,
``evidence``, ``sub_questions``, ``contemplation_depth``,
``recursion_overflow``). Defaults make a freshly-emitted Phase B
candidate a trivially-valid un-contemplated C1 candidate.
"""
candidate_id: str
proposed_chain: dict[str, Any]
trigger: DiscoveryTrigger
source_turn_trace: str
pack_consistent: bool
boundary_clean: bool
review_state: Literal["unreviewed"] = "unreviewed"
# Phase C1 fields. Defaults preserve byte-equality with Phase B
# ``as_dict`` output when the candidate has not been contemplated.
polarity: Literal["affirms", "falsifies", "undetermined"] = "undetermined"
claim_domain: ClaimDomain = "factual"
evidence: tuple[EvidencePointer, ...] = ()
sub_questions: tuple[SubQuestion, ...] = ()
contemplation_depth: int = 0
recursion_overflow: bool = False
def as_dict(self) -> dict[str, Any]:
out: dict[str, Any] = {
"candidate_id": self.candidate_id,
"proposed_chain": self.proposed_chain,
"trigger": self.trigger,
"source_turn_trace": self.source_turn_trace,
"pack_consistent": self.pack_consistent,
"boundary_clean": self.boundary_clean,
"review_state": self.review_state,
}
# Phase C1 fields are emitted only when contemplation has
# produced non-default content. This keeps a freshly-emitted
# Phase B candidate's JSONL line byte-identical to the pre-C1
# encoding.
if (
self.polarity != "undetermined"
or self.claim_domain != "factual"
or self.evidence
or self.sub_questions
or self.contemplation_depth != 0
or self.recursion_overflow
):
out["polarity"] = self.polarity
out["claim_domain"] = self.claim_domain
out["evidence"] = [e.as_dict() for e in self.evidence]
out["sub_questions"] = [s.as_dict() for s in self.sub_questions]
out["contemplation_depth"] = self.contemplation_depth
out["recursion_overflow"] = self.recursion_overflow
return out
_TEACHING_INTENT_NAME: dict[IntentTag, str] = {
IntentTag.CAUSE: "cause",
IntentTag.VERIFICATION: "verification",
}
def _hash_candidate_id(payload: dict[str, Any]) -> str:
"""Deterministic SHA-256 over a canonical JSON encoding.
Sorted keys + tight separators keep the hash stable across
Python runtimes and dict-insertion order. This is the
``candidate_id`` — used both as the on-disk JSONL line key and
by Phase C to look up the originating candidate.
"""
blob = json.dumps(payload, sort_keys=True, separators=(",", ":"))
return hashlib.sha256(blob.encode("utf-8")).hexdigest()
def _boundary_clean(turn_event: Any) -> bool:
"""Return True iff the source turn produced no safety/ethics
refusal and no hedge injection.
Tolerates events that pre-date the bundled-verdicts era (ADR-0039
onward) by reading the canonical fields directly.
"""
refusal_emitted = bool(getattr(turn_event, "refusal_emitted", False) or False)
hedge_injected = bool(getattr(turn_event, "hedge_injected", False) or False)
if refusal_emitted or hedge_injected:
return False
verdicts = getattr(turn_event, "verdicts", None)
if verdicts is not None:
if bool(getattr(verdicts, "refusal_emitted", False) or False):
return False
if bool(getattr(verdicts, "hedge_injected", False) or False):
return False
return True
def _trace_hash(turn_event: Any) -> str:
value = getattr(turn_event, "trace_hash", "") or ""
return str(value)
def extract_discovery_candidates(
turn_event: Any,
intent_tag: IntentTag | None,
intent_subject_lemma: str | None,
*,
grounding_source: str | None = None,
) -> tuple[DiscoveryCandidate, ...]:
"""Return zero or more DiscoveryCandidates for a single turn.
Phase B only fires the ``would_have_grounded`` trigger. All
other triggers in the ``DiscoveryTrigger`` Literal are reserved
for later phases.
Fires when **every** condition holds (deterministic predicate):
1. ``grounding_source`` is ``"none"`` or absent — the turn
fell through to the universal disclosure.
2. ``intent_tag`` is ``CAUSE`` or ``VERIFICATION`` — the
intent set the teaching-grounded surface answers.
3. ``intent_subject_lemma`` is a non-empty pack lemma in the
ratified cognition pack.
4. ``(subject_lemma, intent_name)`` is **not** in the active
corpus — a chain of that shape would have grounded the
turn but does not exist.
Order of conditions matters for tests: short-circuit on the
cheapest predicate first.
"""
source = (grounding_source or getattr(turn_event, "grounding_source", "none") or "none").lower()
if source != "none":
return ()
if intent_tag is None or intent_tag not in _TEACHING_INTENT_NAME:
return ()
if not intent_subject_lemma or not isinstance(intent_subject_lemma, str):
return ()
lemma = intent_subject_lemma.strip().lower()
if not lemma:
return ()
from chat.pack_resolver import is_resolvable
from chat.teaching_grounding import _all_chains_index
# ADR-0064 — discovery gate uses cross-pack residency (any mounted
# lexicon pack) AND cross-corpus chain lookup (any registered
# teaching corpus). A kinship CAUSE prompt whose subject is in
# the relations pack but has no relations-chain in the active
# corpus is now also a discovery signal.
if not is_resolvable(lemma):
return ()
intent_name = _TEACHING_INTENT_NAME[intent_tag]
if (lemma, intent_name) in _all_chains_index():
return ()
# The candidate's proposed_chain is intentionally partial: Phase B
# can only assert that a chain of this (subject, intent) would
# have helped. Connective and object remain null; Phase C is
# where a complete proposed entry is constructed and review-gated.
proposed_chain = {
"subject": lemma,
"intent": intent_name,
"connective": None,
"object": None,
}
trace_hash = _trace_hash(turn_event)
boundary_clean = _boundary_clean(turn_event)
trigger: DiscoveryTrigger = "would_have_grounded"
hash_payload = {
"proposed_chain": proposed_chain,
"trigger": trigger,
"source_turn_trace": trace_hash,
}
candidate_id = _hash_candidate_id(hash_payload)
candidate = DiscoveryCandidate(
candidate_id=candidate_id,
proposed_chain=proposed_chain,
trigger=trigger,
source_turn_trace=trace_hash,
pack_consistent=True, # subject is in pack; object is null pending Phase C
boundary_clean=boundary_clean,
review_state="unreviewed",
)
return (candidate,)
def format_candidate_jsonl(candidate: DiscoveryCandidate) -> str:
"""Serialise to one JSONL line (sorted keys, no trailing newline)."""
return json.dumps(candidate.as_dict(), sort_keys=True, separators=(",", ":"))
__all__ = [
"ClaimDomain",
"DiscoveryCandidate",
"DiscoveryTrigger",
"EvidencePointer",
"SubQuestion",
"extract_discovery_candidates",
"format_candidate_jsonl",
]