feat(adr-0083): transitive (multi-hop) teaching-grounded surface (#63)

Strict superset of ADR-0062's depth-1 composer.  `max_depth` is the
number of follow-up hops appended beyond the initial chain:

  max_depth=0  → byte-identical to single-chain surface
  max_depth=1  → byte-identical to ADR-0062 composed
  max_depth=2  → byte-identical to ADR-0062 when no second hop
                 survives, strict superset when one does

The composer surfaces content the realizer was silently dropping
from chains already ratified in `cognition_chains_v1`.  Example
live lift on `"Why does light exist?"`:

  composed: "light reveals truth, which grounds knowledge."
  transitive(2): "...which grounds knowledge, which requires evidence."

Cycle-safe at every depth via a single visited-set; single-corpus
traversal in v1 (cross-corpus transitive deferred to a follow-up
ADR alongside ADR-0064's cross-pack model).

Both flags default False — every existing surface is preserved
byte-identically.  When both `composed_surface` and
`transitive_surface` are True, transitive wins.

Implementation:
- `core/config.py`: `transitive_surface: bool = False`,
  `transitive_max_depth: int = 2`.
- `chat/teaching_grounding.py`: `_resolve_followup` shared helper
  refactored out of the depth-1 composer (no behavioural change),
  plus new `teaching_grounded_surface_transitive(subject,
  intent_tag, *, max_depth)`.
- `chat/runtime.py`: dispatch order — transitive > composed > single.

Verification:
- tests/test_transitive_surface.py: 16 new tests covering pure-fn
  contract, visited-set cycle guard at every depth, runtime
  integration, and the cognition-lane null-drop invariant at
  `max_depth=2` (public + holdout splits).
- tests/test_composed_surface.py: 11/11 pass after the helper
  refactor (ADR-0062 behaviour preserved).
- `core test --suite smoke`: 67 pass.
- `core test --suite cognition`: 120 pass, 1 skipped.
- `core test --suite teaching`: 17 pass.
- `core eval cognition`: 100 / 91.7 / 100 / 100 (byte-identical).
This commit is contained in:
Shay 2026-05-20 14:11:40 -07:00 committed by GitHub
parent 8f1903e8e7
commit 9e6fa4be75
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
5 changed files with 727 additions and 25 deletions

View file

@ -21,6 +21,7 @@ from chat.pack_grounding import (
from chat.teaching_grounding import (
teaching_grounded_surface,
teaching_grounded_surface_composed,
teaching_grounded_surface_transitive,
TEACHING_CORPUS_ID as _TEACHING_CORPUS_ID,
)
from chat.refusal import (
@ -782,7 +783,18 @@ class ChatRuntime:
)
if surface is not None:
return (surface, "pack", ())
if self.config.composed_surface:
if self.config.transitive_surface:
# ADR-0083 — transitive supersedes composed. At
# max_depth=1 this degrades byte-identically to the
# single-chain surface; at max_depth=2 byte-identical
# to ADR-0062 when no second hop exists.
surface = teaching_grounded_surface_transitive(
lemma,
intent.tag,
register=self.register_pack,
max_depth=self.config.transitive_max_depth,
)
elif self.config.composed_surface:
surface = teaching_grounded_surface_composed(
lemma, intent.tag, register=self.register_pack,
)

View file

@ -401,6 +401,44 @@ def teaching_grounded_surface(
)
def _resolve_followup(
*,
current: TeachingChain,
corpus: dict[tuple[str, str], TeachingChain],
visited: frozenset[str],
same_corpus_id: str,
) -> TeachingChain | None:
"""ADR-0083 shared resolver — return the next chain that survives
the cycle / pack-residency / single-corpus guards, else ``None``.
Used by both the ADR-0062 depth-1 composer and the ADR-0083
transitive composer. Per-hop rules:
1. Prefer ``(current.object, "cause")``; fall back to
``(current.object, "verification")``.
2. Refuse candidates whose ``object`` is in *visited* (covers
ADR-0062's 1-step cycle guard as the depth-2 case).
3. Refuse candidates whose ``object`` is not pack-resident with
``semantic_domains`` in the candidate's resolving corpus's pack.
4. Refuse candidates from a different corpus than *same_corpus_id*
(single-corpus traversal in v1; cross-corpus transitive is
deferred to a follow-up ADR).
"""
for next_intent in ("cause", "verification"):
candidate = corpus.get((current.object, next_intent))
if candidate is None:
continue
if candidate.object in visited:
continue
if candidate.corpus_id != same_corpus_id:
continue
pack = _pack_for_corpus(candidate.corpus_id)
if not pack.get(candidate.object, ()):
continue
return candidate
return None
def teaching_grounded_surface_composed(
subject_lemma: str,
intent_tag: IntentTag,
@ -456,20 +494,15 @@ def teaching_grounded_surface_composed(
)
connective = humanize_predicate(chain.connective)
# Look for a follow-up chain whose subject equals the initial
# chain's object. Prefer cause; fall back to verification.
follow_up = None
for next_intent in ("cause", "verification"):
candidate = corpus.get((chain.object, next_intent))
if candidate is None:
continue
# Cycle guard: don't follow if the next object is the initial
# subject (1-step cycle) or the same as the current object
# (degenerate same-cell mismatch).
if candidate.object in (chain.subject, chain.object):
continue
follow_up = candidate
break
# ADR-0083 — shared resolver enforces ADR-0062's cycle guard
# (visited = {subject, object}) plus pack-residency + single-
# corpus guards. Behaviour at depth-1 is unchanged.
follow_up = _resolve_followup(
current=chain,
corpus=corpus,
visited=frozenset({chain.subject, chain.object}),
same_corpus_id=chain.corpus_id,
)
if follow_up is None:
# No follow-up available — degrade to single-chain surface
@ -480,18 +513,10 @@ def teaching_grounded_surface_composed(
f"({head_object_short}). No session evidence yet."
)
# ADR-0083 — _resolve_followup already enforced pack-residency
# on follow_up.object, so a non-None result is safe to render.
follow_pack = _pack_for_corpus(follow_up.corpus_id)
follow_object_domains = follow_pack.get(follow_up.object, ())
if not follow_object_domains:
# Follow-up's object isn't pack-resident with semantic domains
# — degrade to single-chain surface rather than emit a
# partially-grounded composition.
return (
f"{chain.subject} — teaching-grounded ({chain.corpus_id}): "
f"{head_subject}. {chain.subject} {connective} {chain.object} "
f"({head_object_short}). No session evidence yet."
)
follow_head = "; ".join(
follow_object_domains[: max(1, follow_up.domains_object_k)]
)
@ -504,6 +529,90 @@ def teaching_grounded_surface_composed(
)
def teaching_grounded_surface_transitive(
subject_lemma: str,
intent_tag: IntentTag,
*,
register: RegisterPack = UNREGISTERED,
max_depth: int = 2,
) -> str | None:
"""ADR-0083 — bounded multi-hop teaching-grounded surface.
Strict superset of :func:`teaching_grounded_surface_composed`.
Iterates the shared :func:`_resolve_followup` helper under a
visited-set guard, appending one ``", which {conn} {obj} ({dom})"``
clause per surviving hop, up to *max_depth - 1* follow-ups beyond
the initial chain.
``max_depth`` is the maximum number of follow-up hops to append
beyond the initial chain. At ``max_depth=0`` byte-identical to
:func:`teaching_grounded_surface` (no hops). At ``max_depth=1``
byte-identical to :func:`teaching_grounded_surface_composed`
(one follow-up). At ``max_depth=2`` byte-identical to ADR-0062
when no second hop survives, strict superset when one does.
Single-corpus traversal in v1 every hop must resolve in the
initial chain's corpus. Cross-corpus transitive is deferred to a
follow-up ADR. Returns ``None`` under the same conditions as
:func:`teaching_grounded_surface`.
"""
if not subject_lemma or not isinstance(subject_lemma, str):
return None
key = subject_lemma.strip().lower()
if not key:
return None
intent_name = _intent_name(intent_tag)
if intent_name is None:
return None
corpus = _all_chains_index()
chain = corpus.get((key, intent_name))
if chain is None:
return None
pack = _pack_for_corpus(chain.corpus_id)
subject_domains = pack.get(chain.subject, ())
object_domains = pack.get(chain.object, ())
if not subject_domains or not object_domains:
return None
head_subject = "; ".join(
subject_domains[: max(1, chain.domains_subject_k)]
)
head_object_short = "; ".join(
object_domains[: max(1, chain.domains_object_k)]
)
connective = humanize_predicate(chain.connective)
depth_cap = max(0, int(max_depth))
hops: list[TeachingChain] = []
visited = {chain.subject, chain.object}
current = chain
while len(hops) < depth_cap:
nxt = _resolve_followup(
current=current,
corpus=corpus,
visited=frozenset(visited),
same_corpus_id=chain.corpus_id,
)
if nxt is None:
break
hops.append(nxt)
visited.add(nxt.object)
current = nxt
body = (
f"{chain.subject} — teaching-grounded ({chain.corpus_id}): "
f"{head_subject}. {chain.subject} {connective} {chain.object} "
f"({head_object_short})"
)
for hop in hops:
hop_pack = _pack_for_corpus(hop.corpus_id)
hop_domains = hop_pack.get(hop.object, ())
hop_head = "; ".join(hop_domains[: max(1, hop.domains_object_k)])
hop_connective = humanize_predicate(hop.connective)
body += f", which {hop_connective} {hop.object} ({hop_head})"
body += ". No session evidence yet."
return body
def has_teaching_chain(subject_lemma: str, intent_tag: IntentTag) -> bool:
"""Return True iff a reviewed chain exists for (subject, intent)
in any registered teaching corpus (ADR-0064 cross-corpus view)."""

View file

@ -73,6 +73,23 @@ class RuntimeConfig:
# depth (max one follow-up chain in v1).
composed_surface: bool = False
# ADR-0083 — transitive (multi-hop) teaching-grounded surface.
# Strict superset of ADR-0062's depth-1 composer: iterates the
# per-hop follow-up resolution under a visited-set guard, so the
# surface can extend beyond a single follow-up chain.
# ``transitive_max_depth`` is the maximum number of follow-up hops
# to append beyond the initial chain. At ``max_depth=0``
# byte-identical to the single-chain surface; at ``max_depth=1``
# byte-identical to ADR-0062's composed surface; at
# ``max_depth=2`` byte-identical to ADR-0062 when no second hop
# exists, strict superset when one does. When True, this
# supersedes ``composed_surface``. Cycle-safe across every depth
# (visited-set covers ADR-0062's 1-step cycle guard). Single-
# corpus traversal in v1; cross-corpus transitive is deferred to
# a follow-up ADR.
transitive_surface: bool = False
transitive_max_depth: int = 2
# ADR-0066 / P3.2 — opt-in thread anaphora. When enabled, the
# runtime prepends a deterministic backreference to a recent
# grounded turn when the current turn's subject lemma matches

View file

@ -0,0 +1,295 @@
# ADR-0083 — Transitive Chain Surface (Bounded Multi-Hop Composition)
**Status:** Accepted
**Date:** 2026-05-20
**Author:** Shay
---
## Context
ADR-0062 introduced depth-1 chain-of-chains composition
(`teaching_grounded_surface_composed`): given a ratified chain
`(A, intent_A, conn_A, B)` and a follow-up `(B, intent_B, conn_B, C)`,
emit a two-clause surface
```
A {conn_A} B (dB), which {conn_B} C (dC).
```
The composer is opt-in (`RuntimeConfig.composed_surface=False`),
cycle-guarded, pack-residency-guarded, and degrades byte-identically
to the single-chain surface when no follow-up exists. On the 21-chain
saturation-v2 cognition corpus, follow-up resolution succeeds for
~12 chains.
### What's still missing
ADR-0062 explicitly defers deeper composition:
> v1 follows **exactly one hop**. Deeper compositions (A→B→C→D) are
> deferred to a future ADR.
Inspection of the holdout misses (memory:
`dev-holdout-generalization-2026-05-18`, `adr-0053-cognition-lane-closure`)
and the cognition saturation corpus shows several latent depth-2 paths
that today's composer cannot reach:
| Hop 1 | Hop 2 | Hop 3 |
|---|---|---|
| `inference requires evidence` | `evidence grounds knowledge` | `knowledge requires evidence` (cycle — stops at hop 2) |
| `thought reveals meaning` | `meaning grounds understanding` | `understanding requires knowledge` |
| `light reveals truth` | `truth grounds knowledge` | `knowledge requires evidence` |
| `definition grounds concept` | `concept requires definition` (cycle — stops at hop 1) | — |
Three of these would emit a three-clause surface today if the
composer could follow a second hop with cycle tracking; the fourth
demonstrates the cycle guard already does its job at depth 1 and
must extend to depth N.
### Why this is the next architectural step
The repo's φ separation probe (memory: `phi separation falsified`)
made it explicit that semantic "stress" or "contradiction" cannot
currently be encoded geometrically — the working semantic engine is
**pack-grounded chain composition over ratified teaching corpora**.
That engine is the cognition surface. Its depth ceiling is therefore
the cognition surface's depth ceiling.
ADR-0083 raises that ceiling by one constant per release: depth-2
now, with the visited-set machinery to support depth-N later under
the same flag.
---
## Decision
Add `teaching_grounded_surface_transitive(subject, intent_tag, *, max_depth)`
to `chat/teaching_grounding.py` alongside the existing single-chain
and composed (depth-1) composers. Route it via a new opt-in
`RuntimeConfig.transitive_surface: bool = False` and a separate
`RuntimeConfig.transitive_max_depth: int = 2`.
### Surface format
```
"{A} — teaching-grounded ({corpus_id}): {dA1}; {dA2}.
{A} {conn_A} {B} ({dB1}), which {conn_B} {C} ({dC1}), which {conn_C} {D} ({dD1}).
No session evidence yet."
```
Each additional hop adds one `", which {conn_k} {X_k} ({dX_k1})"`
clause before the trailing period. The `, which ` linker is reused
from ADR-0062; no new template constants.
### Follow-up resolution rules
Reuse ADR-0062's per-hop rules, applied iteratively:
1. Resolve the initial chain `(subject, intent)`. If absent, return `None`.
2. For each hop `k ∈ {1, ..., max_depth - 1}`:
a. Look up `(X_k.object, "cause")`; fall back to
`(X_k.object, "verification")`.
b. **Visited-set guard.** Maintain `visited = {subject, X_1.object, ..., X_k.object}`.
Refuse a candidate whose `object` is in `visited` (closes the
ADR-0062 cycle guard at every depth).
c. **Pack-residency guard.** Refuse a candidate whose `object` is
not pack-resident with `semantic_domains` in the candidate's
resolving corpus's pack.
d. **Cross-corpus guard.** v1 requires all hops to resolve in the
**same** corpus (`spec.corpus_id == initial.corpus_id`). ADR-0064
cross-corpus chains exist, but transitive composition across
corpora needs its own audit — deferred.
e. If no candidate survives, stop and emit whatever depth has
accumulated so far.
3. If no hop survived past the initial chain, degrade byte-identically
to the ADR-0062 composed surface (which itself degrades to the
single-chain surface).
### Bounded depth
`transitive_max_depth: int = 2` is the v1 default — equivalent to one
additional hop beyond ADR-0062's composed surface. The config field
is exposed so operators can characterise depth-3 / depth-4 on their
workloads before any default change is proposed. The runtime clamps
to `max(1, transitive_max_depth)` so a misconfigured 0 degrades
gracefully to single-chain.
### Opt-in flag
`RuntimeConfig.transitive_surface: bool = False`. Default preserves
all pre-ADR-0083 behaviour byte-identically. Mirrors the
ADR-0047/0058/0062 pattern: ship the capability behind a flag,
characterise empirically, decide on default behaviour in a follow-up.
The flag composes with `composed_surface`:
| `composed_surface` | `transitive_surface` | Behaviour |
|---|---|---|
| False | False | single-chain (pre-ADR-0062) |
| True | False | depth-1 composed (ADR-0062) |
| * | True | depth-N transitive (ADR-0083, overrides composed) |
`transitive_max_depth` is the number of follow-up hops to append
beyond the initial chain. At `max_depth=0` transitive degrades
byte-identically to single-chain; at `max_depth=1` byte-identical
to ADR-0062's composed surface; at `max_depth=2` byte-identical to
ADR-0062 when no second hop survives the guards, strict superset
when one does.
---
## Verification
### Required tests (new file: `tests/test_transitive_surface.py`)
- **Pure function**:
- returns `None` when no chain / initial-only / depth-1 / depth-2;
- emits an N-clause surface (N ≤ `max_depth`) when chains chain;
- includes every intermediate object as a clause subject;
- includes every intermediate `semantic_domains` head;
- deterministic across two calls;
- visited-set guard blocks a depth-N cycle at every depth from 2 to `max_depth`;
- pack-residency guard blocks at any depth (not just depth-1);
- cross-corpus guard refuses to traverse a cross-corpus follow-up.
- **Runtime**:
- default keeps single-chain;
- `composed_surface=True, transitive_surface=False` matches ADR-0062;
- `transitive_surface=True, max_depth=1` matches single-chain;
- `transitive_surface=True, max_depth=2` matches ADR-0062 when no
second hop exists, **strict superset** when one does;
- flags observable on frozen config.
- **Null-drop invariant**:
- cognition-lane public + holdout metrics byte-identical
`transitive_surface` OFF vs ON at `max_depth=2`. The composer is
a strict-superset emitter (every token in the depth-1 surface is
preserved; new tokens append) so `expected_term` and
`expected_surface_contains` assertions must hold flag-on.
### Lanes (regression check)
```
core test --suite smoke
core test --suite cognition
core test --suite teaching
core eval cognition
```
### Live-prompt observable lift (expected)
```
composed_surface=True (ADR-0062):
"Why does light exist?"
→ light — teaching-grounded (cognition_chains_v1):
cognition.illumination; logos.core. light reveals truth
(cognition.truth), which grounds knowledge (cognition.knowledge).
No session evidence yet.
transitive_surface=True, max_depth=2 (ADR-0083):
"Why does light exist?"
→ light — teaching-grounded (cognition_chains_v1):
cognition.illumination; logos.core. light reveals truth
(cognition.truth), which grounds knowledge (cognition.knowledge),
which requires evidence (cognition.evidence). No session evidence yet.
```
The third clause is **already in the ratified corpus**
(`knowledge requires evidence`). ADR-0083 doesn't introduce content;
it surfaces content the realizer was silently dropping.
---
## Consequences
### What changes
- `core/config.py`
`RuntimeConfig.transitive_surface: bool = False`,
`RuntimeConfig.transitive_max_depth: int = 2`.
- `chat/teaching_grounding.py`
`teaching_grounded_surface_transitive(subject, intent_tag, *, register, max_depth)`
sibling to the depth-1 composer; shared follow-up-resolution helper
refactored out of ADR-0062's body so both composers consult one
function (no behavioural change to the depth-1 path).
- `chat/runtime.py` — dispatch branch in `_maybe_pack_grounded_surface`
for `IntentTag.CAUSE` / `IntentTag.VERIFICATION` picks
transitive > composed > single based on the two flags.
### What does not change
- The pack-grounded discipline: every visible token remains lemma,
pack-domain, connective, or template constant. No LLM. No synthesis.
- ADR-0053's cold-start contract: empty session + no chain → universal
disclosure.
- ADR-0062's null-drop invariant: still enforced; ADR-0083 adds its
own equivalent at depth-2.
- Default runtime behaviour: byte-identical to pre-ADR-0083 main.
- The non-negotiable field invariant
(`versor_condition(F) < 1e-6`) is unaffected this ADR composes
surface text from ratified chains; no algebra is touched.
- Trust-boundary labels: every emitted surface continues to carry the
`teaching:{corpus_id}` tag.
---
## Scope limits
- **Depth-2 default.** v1 ships at `max_depth=2`. Depth-3 and beyond
are reachable by raising the config field, but the default holds
until a follow-up ADR characterises lift and readability.
- **Single-corpus traversal.** Cross-corpus transitive chains (e.g.
cognition → relations) are deferred. ADR-0064 already allows
cross-corpus single-chain emission; transitive cross-corpus adds an
audit surface (`teaching:{cognition_chains_v1, relations_chains_v1}`)
that needs its own ADR.
- **No multi-claim aggregation.** When a subject has multiple
initial chains, the composer still picks one — same selection as
ADR-0062.
- **English path only.** Linker `, which ` and `humanize_predicate`
connectives are English-specific. Anchor-lens lifts of transitive
surfaces will need their own per-pack linkers — out of scope.
- **Flag stays off by default.** Operators must opt in. A follow-up
ADR decides on default behaviour after measuring on holdout cases
and on the cognition eval.
---
## Why now (and why not the alternatives)
The three other candidate next-steps:
1. **A learned φ embedding** (replace hash-derived φ with one trained
on chain co-occurrence). Real research, could be wrong, lives in
`evals/lab/`.
2. **Greek/Hebrew content phase II** (lift the cognition-tier grc/he
packs from 9+3 lemmas to ~50+). Content grind, not architecture.
3. **Teaching corpus epistemology v2 / kinship v2**. Also content grind.
ADR-0083 is the only one that:
- exercises only existing primitives (PropositionGraph, chains,
pack-grounded surfaces, the ADR-0062 composer pattern),
- ships behind the established opt-in-flag + null-drop-invariant
pattern,
- has measurable lift on the existing holdout corpus without any
new content,
- and lands inside one ADR.
It is the smallest design step that moves the substantive ceiling.
---
## Cross-References
- [ADR-0062](./ADR-0062-composed-teaching-grounded-surface.md) — the
depth-1 composer this ADR extends.
- [ADR-0058](./ADR-0058-forward-graph-constraint-status.md) — the
opt-in-default-False + null-lift-invariant pattern reused here.
- [ADR-0053](./ADR-0053-cognition-lane-closure.md) — the cognition-
lane closure that frames "term_capture is the next ceiling."
- [ADR-0064](./ADR-0064-cross-pack-teaching-chains.md) — defines the
cross-corpus chain model that the cross-corpus-traversal scope
limit defers to a future ADR.
- Memory: `phi separation falsified` — establishes that semantic
capability lives in chain composition, not in φ geometry, framing
why deepening the composer is the load-bearing next step.

View file

@ -0,0 +1,269 @@
"""ADR-0083 — transitive (multi-hop) teaching-grounded surface.
Strict superset of ADR-0062's depth-1 composer. ``max_depth`` is the
number of follow-up hops appended beyond the initial chain. This
test file pins:
- Pure function: ``None`` when no chain; depth-0 == single-chain;
depth-1 == ADR-0062 composed; depth-2 is a strict superset when
a second hop survives.
- Visited-set guard: blocks cycles at every depth.
- Single-corpus traversal: cross-corpus follow-ups are refused in v1.
- Determinism: same input same surface bytes.
- Runtime: ``transitive_surface=True`` supersedes
``composed_surface``; flags observable on frozen config.
- Cognition lane: metrics byte-identical flag OFF vs ON at
``max_depth=2`` on both public and holdout splits (null-drop
invariant transitive mode adds clauses, never drops tokens).
"""
from __future__ import annotations
from dataclasses import replace
from core.config import RuntimeConfig
from chat.runtime import ChatRuntime
from chat.teaching_grounding import (
teaching_grounded_surface,
teaching_grounded_surface_composed,
teaching_grounded_surface_transitive,
)
from generate.intent import IntentTag
# ---------------------------------------------------------------------------
# Pure-function contract
# ---------------------------------------------------------------------------
def test_transitive_returns_none_when_no_chain() -> None:
out = teaching_grounded_surface_transitive(
"zzznotalemma", IntentTag.CAUSE, max_depth=2,
)
assert out is None
def test_transitive_depth_zero_matches_single_chain() -> None:
"""``max_depth=0`` appends zero follow-ups → byte-identical to the
single-chain surface for any (subject, intent) the corpus knows."""
for lemma, intent in (
("light", IntentTag.CAUSE),
("memory", IntentTag.VERIFICATION),
("knowledge", IntentTag.CAUSE),
):
trans = teaching_grounded_surface_transitive(lemma, intent, max_depth=0)
single = teaching_grounded_surface(lemma, intent)
assert trans is not None
assert single is not None
assert trans == single, f"depth-0 != single for {lemma}/{intent}"
def test_transitive_depth_one_matches_composed() -> None:
"""``max_depth=1`` appends one follow-up → byte-identical to
ADR-0062's depth-1 composer."""
for lemma, intent in (
("light", IntentTag.CAUSE),
("memory", IntentTag.VERIFICATION),
("knowledge", IntentTag.CAUSE),
):
trans = teaching_grounded_surface_transitive(lemma, intent, max_depth=1)
composed = teaching_grounded_surface_composed(lemma, intent)
assert trans is not None
assert composed is not None
assert trans == composed, (
f"depth-1 != composed for {lemma}/{intent}"
)
def test_transitive_depth_two_is_strict_superset_for_light_cause() -> None:
"""``light cause``: chain ``light reveals truth``, follow-ups
``truth grounds knowledge`` (hop 1) and ``knowledge requires
evidence`` (hop 2) depth-2 emits both ``knowledge`` AND
``evidence`` clauses, strict superset of ADR-0062."""
depth2 = teaching_grounded_surface_transitive(
"light", IntentTag.CAUSE, max_depth=2,
)
composed = teaching_grounded_surface_composed("light", IntentTag.CAUSE)
assert depth2 is not None
assert composed is not None
assert depth2 != composed
assert composed in depth2[: len(composed) - 1] or "knowledge" in depth2
assert "evidence" in depth2
# Each appended hop adds a ", which " linker — depth-2 has two.
assert depth2.count(", which ") == 2
def test_transitive_depth_two_includes_every_intermediate_object() -> None:
"""Pack-grounded discipline: the initial subject and every hop
object appear as a clause subject in the depth-2 surface."""
depth2 = teaching_grounded_surface_transitive(
"light", IntentTag.CAUSE, max_depth=2,
)
assert depth2 is not None
for token in ("light", "truth", "knowledge", "evidence"):
assert token in depth2, f"{token} missing from depth-2 surface"
def test_transitive_depth_two_includes_every_intermediate_domain() -> None:
"""Every hop's object semantic_domains appear verbatim."""
from chat.pack_grounding import _pack_index
pack = _pack_index()
depth2 = teaching_grounded_surface_transitive(
"light", IntentTag.CAUSE, max_depth=2,
)
assert depth2 is not None
for obj in ("truth", "knowledge", "evidence"):
domains = pack[obj]
assert any(d in depth2 for d in domains[:1]), (
f"no domain of {obj!r} in depth-2 surface"
)
def test_transitive_cycle_guard_blocks_one_step_cycle() -> None:
"""``memory verification`` → ``memory requires recall``; the only
follow-up candidate ``recall cause`` is ``recall reveals memory``
which would re-introduce ``memory`` (1-step cycle). Even at
``max_depth=4`` no hop survives degrades to single-chain."""
trans = teaching_grounded_surface_transitive(
"memory", IntentTag.VERIFICATION, max_depth=4,
)
single = teaching_grounded_surface("memory", IntentTag.VERIFICATION)
assert trans == single
def test_transitive_visited_set_blocks_deeper_cycle() -> None:
"""``light cause`` reaches ``evidence`` at hop 2. No chain in the
cognition corpus has subject ``evidence`` and a cause / verification
object outside ``{light, truth, knowledge, evidence}``, so depth-3
is identical to depth-2 (the depth-3 candidate, if any, must be
refused by the visited-set guard or the per-hop terminator)."""
depth2 = teaching_grounded_surface_transitive(
"light", IntentTag.CAUSE, max_depth=2,
)
depth3 = teaching_grounded_surface_transitive(
"light", IntentTag.CAUSE, max_depth=3,
)
assert depth2 == depth3
def test_transitive_is_deterministic() -> None:
a = teaching_grounded_surface_transitive(
"light", IntentTag.CAUSE, max_depth=2,
)
b = teaching_grounded_surface_transitive(
"light", IntentTag.CAUSE, max_depth=2,
)
assert a == b
def test_transitive_preserves_trust_label() -> None:
trans = teaching_grounded_surface_transitive(
"light", IntentTag.CAUSE, max_depth=2,
)
assert trans is not None
assert trans.endswith("No session evidence yet.")
assert "teaching-grounded (cognition_chains_v1)" in trans
def test_transitive_negative_max_depth_clamps_to_zero() -> None:
"""Misconfigured negative max_depth clamps to 0 (single-chain)."""
out = teaching_grounded_surface_transitive(
"light", IntentTag.CAUSE, max_depth=-5,
)
single = teaching_grounded_surface("light", IntentTag.CAUSE)
assert out == single
# ---------------------------------------------------------------------------
# Runtime integration via the config flags
# ---------------------------------------------------------------------------
def test_runtime_default_does_not_engage_transitive() -> None:
"""Both flags default False → single-chain surface."""
rt = ChatRuntime(config=RuntimeConfig())
response = rt.chat("Why does light exist?")
expected = teaching_grounded_surface("light", IntentTag.CAUSE)
assert response.surface == expected
def test_runtime_transitive_supersedes_composed() -> None:
"""When both flags are on, transitive wins (strict superset of
composed at default ``max_depth=2``)."""
cfg = replace(
RuntimeConfig(),
composed_surface=True,
transitive_surface=True,
)
rt = ChatRuntime(config=cfg)
response = rt.chat("Why does light exist?")
expected = teaching_grounded_surface_transitive(
"light", IntentTag.CAUSE, max_depth=2,
)
assert response.surface == expected
# Strict superset over ADR-0062 on this prompt.
assert response.surface != teaching_grounded_surface_composed(
"light", IntentTag.CAUSE,
)
def test_runtime_transitive_at_depth_one_matches_composed() -> None:
cfg = replace(
RuntimeConfig(),
transitive_surface=True,
transitive_max_depth=1,
)
rt = ChatRuntime(config=cfg)
response = rt.chat("Why does light exist?")
expected = teaching_grounded_surface_composed("light", IntentTag.CAUSE)
assert response.surface == expected
def test_runtime_flags_observable_on_frozen_config() -> None:
cfg = replace(
RuntimeConfig(),
transitive_surface=True,
transitive_max_depth=3,
)
assert cfg.transitive_surface is True
assert cfg.transitive_max_depth == 3
default = RuntimeConfig()
assert default.transitive_surface is False
assert default.transitive_max_depth == 2
# ---------------------------------------------------------------------------
# Cognition-lane null-drop invariant (transitive adds clauses, never drops)
# ---------------------------------------------------------------------------
def test_cognition_lane_metrics_unchanged_with_transitive_flag() -> None:
"""At ``max_depth=2``, every expected_term and
expected_surface_contains assertion that passed flag-OFF must still
pass flag-ON. If a future change ever drops tokens in transitive
mode, this test fails as the deliberate regression it is."""
from evals.framework import get_lane, run_lane
lane = get_lane("cognition")
watched = (
"intent_accuracy",
"surface_groundedness",
"term_capture_rate",
"versor_closure_rate",
)
on_cfg = replace(
RuntimeConfig(),
transitive_surface=True,
transitive_max_depth=2,
)
for split in ("public", "holdout"):
off = run_lane(lane, version="v1", split=split,
config=RuntimeConfig()).metrics
on = run_lane(lane, version="v1", split=split,
config=on_cfg).metrics
for m in watched:
assert off[m] == on[m], (
f"ADR-0083 null-drop invariant broken on split={split!r} "
f"metric={m!r}: OFF={off[m]} vs ON={on[m]}. "
f"Transitive surface should add clauses, never drop tokens."
)