core-labs/core
1
0
Fork 0
Code Issues Pull requests 1 Projects Releases Packages Wiki Activity Actions
core/docs/decisions/ADR-0052-teaching-grounded-surface.md

13 KiB
Raw Permalink Blame History

ADR-0052 — Teaching-Grounded Surface for Cold-Start CAUSE / VERIFICATION

Status: Accepted Date: 2026-05-18 Author: Shay

Context

ADR-0048 added a pack-grounded surface for cold-start DEFINITION / RECALL, and ADR-0050 extended that to COMPARISON. Both consult the ratified en_core_cognition_v1 pack as a second source of grounding alongside the session vault and brought the cognition eval to surface_groundedness=69.2% / term_capture_rate=58.3%.

The remaining unlift cases on the 13-case public split were:

Case Prompt Intent
cause_light_007 "Why does light exist?" CAUSE
cause_knowledge_032 "Why does knowledge require evidence?" CAUSE
verification_memory_037 "Does memory require recall?" VERIFICATION
correction_specific_015 "No, correction means reviewed repair" CORRECTION

correction_specific_015 is out of scope — it requires prior-turn context, not cold-start teaching-store inference.

The three cold-start cases share a structural property the pack alone cannot satisfy: they ask about a relation between two pack-known lemmas (light ↔ truth, knowledge ↔ evidence, memory ↔ recall). Pack semantic_domains describe a single subject, not a relation. Fabricating the relation would violate the no-LLM-fallback doctrine.

However, the system already has a documented place for reviewed relational memory: per CLAUDE.md's Teaching Safety section and ADR-0018, reviewed memory may contribute grounding evidence as long as it goes through a teaching-store path and stays auditable. The doctrinally clean fix is a third grounding source — a small, ratified corpus of cognition chains, treated as immutable reviewed memory at runtime, whose every emitted atom is verifiably pack-sourced.

Decision

Add a deterministic teaching-grounded surface as a third branch of _maybe_pack_grounded_surface, identical guardrails to the DEFINITION/RECALL and COMPARISON branches.

Corpus

A new file teaching/cognition_chains/cognition_chains_v1.jsonl holds the reviewed chain corpus. Each line:

{
  "chain_id": "cause_knowledge_requires_evidence",
  "subject": "knowledge",
  "intent": "cause",
  "connective": "requires",
  "object": "evidence",
  "domains_subject_k": 2,
  "domains_object_k": 1,
  "provenance": "adr-0052:reviewed:2026-05-18"
}

Constraints enforced at load time (chat/teaching_grounding.py:_corpus_index):

  • intent ∈ {cause, verification} (any other value is dropped).
  • subject and object MUST both be present in the ratified en_core_cognition_v1 pack with non-empty semantic_domains.
  • connective MUST be a key recognised by generate.semantic_templates.humanize_predicate (e.g. requires, reveals, evidences, is_grounded_in).

Entries that fail any constraint are silently dropped — the runtime cannot leak a non-pack atom into a surface even if the corpus is edited improperly.

Three chains ship in v1:

chain_id subject intent connective object
cause_light_reveals_truth light cause reveals truth
cause_knowledge_requires_evidence knowledge cause requires evidence
verification_memory_requires_recall memory verification requires recall

Each connective is corroborated by the subject's own pack semantic_domains (light → meaning.revelation; memory → recall.surface; knowledge → epistemic.ground).

Surface format

{subject} — teaching-grounded ({corpus_id}): {ds1}; {ds2}. {subject} {connective} {object} ({do1}). No session evidence yet.

Every visible non-template token is either one of the two lemmas, a verbatim semantic_domains string from the pack, or the connective predicate already humanised by humanize_predicate. No rewording, no LLM, no synthesis.

Engagement conditions

teaching_grounded_surface(subject_lemma, intent_tag) returns a non-None surface only when all hold:

  • subject_lemma is a non-empty string,
  • intent_tag is IntentTag.CAUSE or IntentTag.VERIFICATION,
  • the (subject, intent) pair has a chain in the corpus.

_maybe_pack_grounded_surface invokes this path only when:

  • gate fired with source="empty_vault" (cold-start session),
  • config.output_language == "en",
  • intent ∈ {CAUSE, VERIFICATION},
  • intent.subject resolves to a chain.

Any other condition returns None and the runtime falls through to the universal disclosure unchanged. Safety / ethics refusal still takes priority above this branch (refusal is a remediation tier, not a grounding source).

Provenance tag

ChatResponse.grounding_source and TurnEvent.grounding_source gain a fourth value, "teaching", sibling to "vault", "pack", "none". Downstream audit consumers can filter teaching-grounded surfaces distinctly from pack-grounded surfaces — important because the corpus is reviewed memory, not pack memory, and may receive updates through a different process than pack ratification.

_maybe_pack_grounded_surface was widened to return tuple[str, str] | None so the provenance tag is propagated alongside the surface; the previous boolean "pack" if pack_surface else "none" collapsed too early.

Why this is doctrine-aligned

CLAUDE.md prohibits opaque LLM fallbacks, stochastic sampling, hidden normalisation, hot-path repair, and approximate recall. This ADR is:

  • Not opaque. Every visible atom is either a lemma supplied by the intent classifier or a verbatim pack semantic_domains string. Corpus id and pack id are both inspectable.
  • Not stochastic. Deterministic JSONL read; identical input produces byte-identical output (test_surface_is_deterministic).
  • Not hidden normalisation. The corpus lookup is a separate source of grounding, not a normalisation step inside an existing operator. No versor, no manifold, no field state touched.
  • Not hot-path repair. UnknownDomainGate semantics are unchanged — the gate still fires; this ADR only broadens what the stub-path emits after the gate fires, in a narrow intent-typed branch.
  • Not approximate recall. Exact dictionary lookup on the corpus by (subject, intent) key, exact dictionary lookup on the pack by lemma. No metric, no neighbourhood, no threshold.
  • Reviewed teaching, not unreviewed mutation. The corpus ships as a checked-in artifact; no runtime path mutates it. The corpus is treated as immutable reviewed memory in the spirit of ADR-0018's teaching-store discipline.

The fundamental architectural move is the same as ADR-0048 / ADR-0050: multiple ratified sources contribute to the runtime's grounding basis, each with its own provenance tag. This ADR adds reviewed teaching chains as a third sibling alongside session vault and ratified pack.

Characterisation — core eval cognition

A/B run on the 13-case public cognition split, identical RuntimeConfig except for the merge of this ADR:

Metric Pre-ADR-0052 Post-ADR-0052 Δ
intent_accuracy 100.0 % 100.0 % 0
surface_groundedness 69.2 % 92.3 % +23.1 pp
term_capture_rate 58.3 % 83.3 % +25.0 pp
versor_closure_rate 100.0 % 100.0 % 0
versor_condition < 1e-6 preserved preserved invariant

The three cases that lift are exactly the ones targeted:

"Why does light exist?"
  -> intent.tag = CAUSE, subject="light"
  -> chain: light -reveals-> truth
  -> "light — teaching-grounded (cognition_chains_v1):
      cognition.illumination; logos.core. light reveals truth
      (cognition.truth). No session evidence yet."
  -> grounding_source = "teaching"

"Why does knowledge require evidence?"
  -> intent.tag = CAUSE, subject="knowledge"
  -> chain: knowledge -requires-> evidence
  -> "knowledge — teaching-grounded (cognition_chains_v1):
      cognition.knowledge; epistemic.ground. knowledge requires
      evidence (cognition.evidence). No session evidence yet."
  -> grounding_source = "teaching"

"Does memory require recall?"
  -> intent.tag = VERIFICATION, subject="memory"
  -> chain: memory -requires-> recall
  -> "memory — teaching-grounded (cognition_chains_v1):
      cognition.memory; memory.semantic. memory requires recall
      (operation.recall). No session evidence yet."
  -> grounding_source = "teaching"

The single remaining unlift case (correction_specific_015) is correctly out of scope — it requires prior-turn context, not cold-start inference.

Consequences

What changes

  • New file: teaching/cognition_chains/cognition_chains_v1.jsonl — reviewed corpus of 3 cognition chains.
  • New module: chat/teaching_grounding.py — corpus loader, teaching_grounded_surface, has_teaching_chain.
  • chat/runtime.py:_maybe_pack_grounded_surface widened to return (surface, grounding_source_tag) and gains a CAUSE / VERIFICATION branch that routes to teaching.
  • ChatResponse.grounding_source and TurnEvent.grounding_source docstrings updated; "teaching" is now an accepted value.
  • Three new cases lift in the cognition eval.

What does not change

  • _UNKNOWN_DOMAIN_SURFACE constant retained; unmodelled intents, unknown subjects, and non-English paths still return the universal disclosure unchanged.
  • UnknownDomainGate semantics unchanged — the gate still fires on every empty-vault cold-start turn.
  • DEFINITION / RECALL / COMPARISON paths unchanged — they still emit grounding_source="pack" via ADR-0048 / ADR-0050.
  • core/cognition/pipeline.py's gate_fired detection (response.grounding_source != "vault") works without modification — "teaching" != "vault".
  • Safety / ethics refusal still takes priority above teaching grounding.
  • versor_condition(F) < 1e-6 invariant unaffected (no algebra changes).

Scope limits

  • English only (en_core_cognition_v1). Same constraint as ADR-0048 / ADR-0050.
  • CAUSE / VERIFICATION only. Other intents (PROCEDURE, TRANSITIVE_QUERY, FRAME_TRANSFER) are out of scope here; they would need their own reviewed corpora.
  • Single-subject CAUSE / VERIFICATION only. Multi-clause causal structures would need richer chain shape and are deferred.
  • Three chains in v1. The corpus is intentionally compact — the pattern is the load-bearing artifact, not corpus volume. Each new chain must be reviewed and reference only pack-known lemmas.
  • Corpus is immutable at runtime. Future ADRs may add a reviewed proposal-and-ratification path; the current contract is checked-in-only.
  • correction_specific_015 remains unlifted by design — corrections reference prior-turn context which cold-start teaching cannot supply.

Cross-References

  • ADR-0018 — teaching-store infrastructure whose discipline this ADR's corpus extends to the surface layer.
  • ADR-0048 — the DEFINITION / RECALL pack-grounded surface this ADR mirrors for CAUSE / VERIFICATION; same trust-boundary discipline, sibling grounding_source tag.
  • ADR-0049 — head-noun subject extraction this ADR consumes. intent.subject for "Why does knowledge require evidence?" is the clean lemma "knowledge".
  • ADR-0050 — the COMPARISON sibling, sharing the same dispatcher and provenance scheme.
  • ADR-0029 / ADR-0033 — sibling-pack-grounded contributor pattern this ADR continues to extend from verdict surfaces to the answer surface, now from a teaching-store source.

Verification

tests/test_teaching_grounding.py                 — 22 tests, all green
tests/test_pack_grounded_comparison.py           — pre-existing, still green
tests/test_pack_grounding.py                     — pre-existing, still green
tests/test_intent_subject_extraction.py          — pre-existing, still green
tests/test_semantic_realizer_integration.py      — pre-existing, still green

Lanes (all green on this branch):
  core test --suite smoke         64 passed (3 pre-existing failures in
                                             tests/test_architectural_invariants.py
                                             flag stale worktree directories,
                                             unrelated to this ADR)
  core test --suite cognition    121 passed
  core test --suite runtime       19 passed
  core test --suite packs          6 passed
  core test --suite teaching      17 passed
  core test --suite algebra      132 passed

core eval cognition (pre → post):
  intent_accuracy        100.0% → 100.0%   (=)
  surface_groundedness    69.2% →  92.3%   (+23.1 pp)
  term_capture_rate       58.3% →  83.3%   (+25.0 pp)
  versor_closure_rate    100.0% → 100.0%   (=)

The non-negotiable field invariant (versor_condition(F) < 1e-6) is preserved: this ADR adds a surface-construction branch on the existing stub path — no algebra changes, no rotor construction, no field update.