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.
8.2 KiB
ADR-0064 — Cross-pack teaching chains
Status: Accepted Date: 2026-05-18 Author: Shay Supersedes: none (extends ADR-0052 / ADR-0062 / ADR-0063) Phase: Plan Phase 1 (corpus flywheel)
Context
ADR-0052 introduced reviewed teaching chains as a third grounding
source alongside vault and pack-grounded surfaces. The corpus
(teaching/cognition_chains/cognition_chains_v1.jsonl) is reviewed,
immutable, append-only memory. ADR-0063 brought
en_core_relations_v1 (kinship pack) onto the live runtime, but the
teaching-grounded surface composer was still hardcoded to the
cognition corpus:
# chat/teaching_grounding.py — pre-ADR-0064
TEACHING_CORPUS_ID: str = "cognition_chains_v1"
_CORPUS_PATH = .../cognition_chains/cognition_chains_v1.jsonl
Every cold-start CAUSE/VERIFICATION prompt on a kinship lemma fell through to the universal disclosure even though the relations pack was mounted, because no kinship chain corpus existed and no path existed to register one.
ADR-0064 closes that gap architecturally. ADR-0063 was the resolver unlock at the pack layer; ADR-0064 is the same unlock at the teaching corpus layer.
Decision
1. Teaching corpus registry
A new dataclass + constant in chat/teaching_grounding.py:
@dataclass(frozen=True, slots=True)
class TeachingCorpusSpec:
corpus_id: str
path: Path
pack_id: str
TEACHING_CORPORA: tuple[TeachingCorpusSpec, ...] = (
TeachingCorpusSpec("cognition_chains_v1", .../cognition_chains_v1.jsonl, "en_core_cognition_v1"),
TeachingCorpusSpec("relations_chains_v1", .../relations_chains_v1.jsonl, "en_core_relations_v1"),
)
Each corpus is 1:1-bound to exactly one lexicon pack. The 1:1
binding is the structural invariant that prevents cross-domain
leakage during cold-start surface composition: a relations chain
cannot accidentally surface a cognition-pack atom (or vice versa)
because pack-residency at load time is scoped to the corpus's
declared pack. Cross-domain chain shapes are out of scope for v1
per docs/teaching_order.md §5.
2. Aggregated chain index
_all_chains_index() — lru_cached aggregator that loads every
registered corpus via _load_corpus(spec) and unions them into a
single {(subject, intent): TeachingChain} view. Registration order
is the resolution order; cognition is registered first so cognition-
lane byte-identity is preserved on any future cross-corpus
collision.
TeachingChain gains a corpus_id field so the surface tag and
audit trail are unambiguous when multiple corpora are active.
3. Surface composers consult the aggregated view
teaching_grounded_surface and teaching_grounded_surface_composed
now call _all_chains_index(). The surface tag follows the chain's
resolving corpus_id:
parent → teaching-grounded (relations_chains_v1)
light → teaching-grounded (cognition_chains_v1)
Cognition-lane surfaces remain byte-identical; relations-pack lemmas now ground on the live path.
4. Discovery gate updated for cross-corpus residency
teaching/discovery.py previously gated on
(lemma in cognition_pack) AND ((lemma, intent) not in cognition_corpus).
Updated to:
from chat.pack_resolver import is_resolvable
from chat.teaching_grounding import _all_chains_index
if not is_resolvable(lemma): # any mounted pack
return ()
if (lemma, intent_name) in _all_chains_index(): # any registered corpus
return ()
A kinship CAUSE prompt that lacks a relations chain is now correctly flagged as a discovery candidate, instead of being suppressed because the cognition pack doesn't carry the lemma.
5. Replay-equivalence gate registers the swap
teaching/replay.py's _swap_corpus_path was extended to also
rewrite the registry entry's path for the swapped corpus AND
invalidate _all_chains_index cache, so surface composers re-read
the swapped corpus during the gate's transient phase. The active
corpus on disk remains byte-identical to its pre-swap state — the
replay invariant is preserved.
6. Back-compat
_corpus_index(), _CORPUS_PATH, TEACHING_CORPUS_ID retain
cognition-corpus-specific semantics for consumers whose scope is
explicitly the cognition corpus (audit, replay's cognition-public
runner, replay state-tracking). They are not removed. The aggregated
_all_chains_index is the new abstraction for surface composers
and the discovery gate.
A new helper clear_teaching_caches() drops every teaching-related
lru_cache atomically — replaces ad-hoc _corpus_index.cache_clear()
calls in replay code paths.
Consequences
Capability unlocked
| Path | Cognition lemmas | Kinship lemmas |
|---|---|---|
teaching_grounded_surface(CAUSE/VERIFICATION) |
byte-identical | now grounds for cells in relations_chains_v1 |
teaching_grounded_surface_composed |
byte-identical | composes within relations corpus when chain-of-chains exists |
| Discovery gate | byte-identical | now emits candidates for kinship cells absent from the relations corpus |
Cognition lane: byte-identical
Cognition lemmas resolve to the cognition corpus first; the orthogonal-pack invariant prevents any (subject, intent) collision between corpora. Public/holdout eval baselines unchanged:
public: intent 100% / surface 100% / term 91.7% / closure 100%
holdout: intent 100% / surface 100% / term 83.3% / closure 100%
Live verification
$ core chat
> Why does parent exist?
parent — teaching-grounded (relations_chains_v1): kinship.ascendant.direct;
kinship.parent. parent precedes child (kinship.descendant.direct).
No session evidence yet.
grounding_source = teaching
Future ADRs unlocked
- Cross-domain triples. Once relations corpus saturates
internally, a follow-up ADR can extend the chain shape to allow
subject and object in different packs (e.g.
family grounds identity— family ∈ relations, identity ∈ cognition). - Relations eval lane. Mirror the cognition lane harness with relations-domain holdout cases. The seed corpus is the ground truth.
- Audit + supersede for the relations corpus.
teaching auditandteaching supersedeare cognition-corpus-only today; the registry layer makes generalizing them mechanical.
Trust boundaries
- Each corpus is 1:1-bound to one lexicon pack. The binding is
declared statically in
TEACHING_CORPORA— runtime cannot introduce a new corpus or rebind a pack. - Chain loading is read-only over immutable, reviewed, append-only files. Same trust boundary as ADR-0052.
- Surface tag tokens emitted are corpus_id strings declared in
TEACHING_CORPORA— never derived from user input. - The replay-gate swap rewrites the registry tuple in-place for the duration of the gate's transient phase and restores it on exit. Side-effect-free from outside the contextmanager.
Files changed
chat/teaching_grounding.py registry layer
teaching/discovery.py cross-corpus gate
teaching/replay.py swap-the-registry
teaching/relations_chains/relations_chains_v1.jsonl NEW (seed corpus, 7 chains)
tests/test_relations_chains_v1.py NEW (17 tests)
docs/decisions/ADR-0064-cross-pack-teaching-chains.md NEW (this file)
docs/decisions/README.md index entry
docs/curriculum/relations_chains_v1.md NEW (curriculum unit doc)
Verification
tests/test_relations_chains_v1.py 17 passed
tests/test_teaching_audit.py 23 passed
tests/test_composed_surface.py 11 passed
tests/test_teaching_grounding.py passed
tests/test_discovery_candidates.py 24 passed
Lanes:
core test --suite smoke 67 passed
core test --suite cognition 121 passed
core test --suite teaching 17 passed
core test --suite packs 6 passed
core test --suite runtime 19 passed
core test --suite algebra 132 passed
The non-negotiable field invariant versor_condition(F) < 1e-6 is
unaffected — this ADR is a routing/dispatch change over immutable
corpus data; no algebra, no field operators, no normalization sites
were touched.