feat(adr-0182): anchor-skip + intra-clause accumulation — distractor 0016 refuses, twin 0017 solves (confuser wrong 1→0) (#481)

* feat(adr-0182): anchor-skip + intra-clause accumulation — distractor 0016 refuses, twin 0017 solves (confuser wrong 1->0)

The last confuser wrong. 0016 ("A train travels 60 miles per hour for 2 hours. Tom
has 8 tickets and buys 4 more tickets. How many tickets?") committed the blunt
product 60x2x8x4=3840 because it was the unique complete reading: the train sentence
is an all-foreign anchor-position block (2 quantities, can't seed an anchor) and the
Tom sentence packs state+change in ONE sentence ("has 8 ... and buys 4 more"), which
the sentence-level reader couldn't decompose -> accumulation_candidates was empty ->
no rival reading -> product committed.

The microscope confirmed intra-clause state+change is a REAL GSM8K pattern (train-0010
"Yun had 20 paperclips initially, but then lost 12"; practice-0121 "Sam has 30 apples
and gives 10 to Anna") -- so this is genuine comprehension, not a 0016-only patch.

Mechanism (added to accumulation_candidates ONLY -> feeds only the pool; train_sample
serving + practice use compose_accumulation, unchanged -> 3/47/0 byte-identical by
construction):
- _sub_clauses: sentence clauses further split on coordinating conjunctions
  (and / then / and then). LOCAL to the ungated candidate generator -- the global
  segmenter (GB-1/GB-2/serving) is untouched.
- _build_accumulation_anchor_skip: anchor = first single-quantity sub-clause (leading
  non-anchorable all-foreign blocks are skipped); chain the conjunction-mate change
  ("buys 4 more" -> +4). Referent guard + polarity-cue requirement still gate it
  (a no-cue sub-clause -> refuse), so GB-2 same-unit lists ("6 apples and 4 apples")
  produce no spurious candidate.

Result (sealed lane):
- confuser wrong 1 -> 0. distractor-quantity 0 wrong / 2 refused: 0016 refuses
  (product 3840 [complete] vs additive 12 [exempt] disagree). BONUS: the clean twin
  0017 ("Tom has 8 tickets and buys 4 more tickets", no distractor) now *solves* 12
  -- genuine comprehension of a real positive, the comprehension-vs-surface-match
  discrimination the corpus exists to measure. genuine positives 7 -> 8 solved.
- the only non-clean verdict left is 1 spurious (0010 multi-referent "altogether" --
  a separate H1 graduation question, not this lever).
- train_sample 3/47/0 and practice 3/47/0 byte-identical; 243 derivation/pool tests
  + 40 architectural invariants green.

Tests:
- test_adr_0182_pool.py TestAnchorSkipIntraClause: intra-clause twin resolves to 12;
  the 0016 anchor-skip candidate classifies `exempt`; 0016 refuses via disagreement;
  no spurious extra candidate without a conjunction.
- test_adr_0163_f2_confusers.py: baseline wrong 1->0, positives_solved 7->8; renamed
  test_distractor_quantity_refuses asserts BOTH 0014 and 0016; new
  test_intra_clause_twin_0017_solves.

Stacked on #480 (prior-state guard); merge #480 first. Confuser arc wrong 7->5->2->1->0.

* test(adr-0182): close anchor-skip refuse-branch coverage gap (lookback finding)

The four-PR lookback review (EX-6/pooling/prior-state/anchor-skip) found the
anchor-skip refuse branches in _build_accumulation_anchor_skip untested: the
referent guard (new named actor -> refuse) and the polarity-cue requirement were
asserted but not proven (no test would fail if removed). Per the schema-obligation
discipline, add NON-VACUOUS failing-under-violation tests:

- test_anchor_skip_referent_guard_discriminates_actor: identical structure, change
  sub-clause subject is a pronoun ('he', same referent -> 12 IS produced) vs a new
  name ('Sara' -> guard suppresses it). Removing _same_referent makes the new-actor
  case also produce 12 -> the second assertion fails. (Non-vacuous: the positive
  control proves the path is reachable, so the negative isn't trivially-empty.)
- test_anchor_skip_requires_a_polarity_cue: 'gets 4 more' (cue -> 12) vs 'owns 4'
  (no cue -> not guessed). Removing the polarity gate makes the no-cue case produce
  a reading -> fails.
- test_anchor_skip_refuses_without_single_quantity_anchor: no single-quantity
  sub-clause -> () (does not force a multi-quantity clause to anchor).

No code change; behavior unchanged. Confuser wrong stays 0; 281 derivation tests +
40 invariants green.
This commit is contained in:
Shay 2026-05-29 15:05:31 -07:00 committed by GitHub
parent d8b37069c0
commit f008550d79
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 178 additions and 11 deletions

View file

@ -178,6 +178,67 @@ def _build_accumulation(
return GroundedDerivation(start=start, steps=tuple(steps))
# ADR-0182 anchor-skip: sub-clause split on conjunctions. A single sentence can pack
# a state and its change ("Tom has 8 tickets and buys 4 more tickets") — the
# sentence-level segmenter (used everywhere; not changed) keeps them together. This
# finer split is *local* to the ungated candidate generator, so it cannot move
# GB-1/GB-2/serving/practice (which never call it). Lexeme-level (ADR-0165): it names
# coordinating conjunctions, it does not parse grammar.
_CONJUNCTION_SPLIT: Final[re.Pattern[str]] = re.compile(r",?\s+(?:and then|and|then)\s+")
def _sub_clauses(problem_text: str) -> list[str]:
"""Sentence clauses, each further split on coordinating conjunctions."""
parts: list[str] = []
for clause in segment_clauses(problem_text):
parts.extend(p.strip() for p in _CONJUNCTION_SPLIT.split(clause) if p.strip())
return parts
def _build_accumulation_anchor_skip(problem_text: str) -> GroundedDerivation | None:
"""ADR-0182 — accumulation over sub-clauses, skipping a leading all-foreign block.
Reads ``A train travels 60 mph for 2 hours. Tom has 8 tickets and buys 4 more
tickets.`` by skipping the (anchor-position) train block its quantities cannot
seed an anchor (1 quantity) and anchoring on the first single-quantity
sub-clause (``Tom has 8 tickets``), then chaining its conjunction-mate change
(``buys 4 more`` +4). The skipped block's quantities go unused; the pool's
isolated-foreign exemption then classifies the reading ``exempt`` (commit-
ineligible), so it can only force a disagreement refusal, never commit. Ungated.
"""
sub_clauses = [(s, extract_quantities(s)) for s in _sub_clauses(problem_text)]
quantity_subs = [(s, qs) for s, qs in sub_clauses if qs]
if len(quantity_subs) < 2:
return None
# Anchor = first single-quantity sub-clause; leading non-anchorable (≠1
# quantity) sub-clauses are skipped (candidate distractor blocks).
anchor_idx = next((i for i, (_, qs) in enumerate(quantity_subs) if len(qs) == 1), None)
if anchor_idx is None:
return None
anchor_sub, anchor_qs = quantity_subs[anchor_idx]
start = anchor_qs[0]
anchor_subject = _subject_token(anchor_sub)
steps: list[Step] = []
for sub, qs in quantity_subs[anchor_idx + 1:]:
if not _same_referent(sub, anchor_subject):
return None # new named actor -> referent hazard -> refuse
if len(qs) != 1:
return None # one change per sub-clause (multi-change is GB-3b.2)
polarity = _polarity(sub)
if polarity is None:
return None # no unambiguous licensed change cue -> refuse
change = qs[0]
operand = Quantity(value=change.value, unit=start.unit, source_token=change.source_token)
op = "add" if polarity > 0 else "subtract"
steps.append(Step(op=op, operand=operand, cue=_cue(sub, polarity)))
if not steps:
return None
return GroundedDerivation(start=start, steps=tuple(steps))
def compose_accumulation(problem_text: str) -> Resolution | None:
"""GB-3b.1 composer — single-referent gain/loss accumulation. Refuse-preferring.
@ -193,13 +254,19 @@ def compose_accumulation(problem_text: str) -> Resolution | None:
def accumulation_candidates(problem_text: str) -> tuple[GroundedDerivation, ...]:
"""ADR-0182 — the ungated accumulation readings for cross-composer pooling.
The strict reading and (when it differs) the distractor-skip reading. Ungated:
the pool classifies each (``complete`` commits, ``exempt`` refuses-only) and the
disagreement rule does the wrong=0 work. Deterministic; de-dup is the pool's job.
Three readings: the strict GB-3b.1 reading, the distractor-skip reading (drops an
isolated-foreign quantity in a multi-quantity change clause 0014), and the
anchor-skip reading (skips a leading all-foreign block + reads a conjunction-mate
intra-sentence change 0016). Ungated: the pool classifies each (``complete``
commits, ``exempt`` refuses-only) and the disagreement rule does the wrong=0 work.
Deterministic; de-dup is the pool's job.
"""
candidates: list[GroundedDerivation] = []
for drop in (False, True):
derivation = _build_accumulation(problem_text, drop_isolated_foreign=drop)
if derivation is not None:
candidates.append(derivation)
anchor_skip = _build_accumulation_anchor_skip(problem_text)
if anchor_skip is not None:
candidates.append(anchor_skip)
return tuple(candidates)

View file

@ -34,8 +34,17 @@ from evals.gsm8k_math.confusers.v1.runner import (
# *before* a stated change ("...before lunch?") is refused — the forward reader
# computes the net, the wrong temporal point. 0020 refuses; its 'left' twin 0021
# still solves. wrong 2->1 (only 0016 remains). pair_tells 1->0.
_BASELINE_WRONG = 1
# 2026-05-29 (ADR-0182 anchor-skip + intra-clause): a sentence packing state+change
# ("Tom has 8 tickets and buys 4 more tickets") is read via conjunction split, and
# a leading all-foreign block ("A train travels 60 miles ... for 2 hours") is
# skipped from anchor selection. 0016 refuses (product 3840 vs additive 12). BONUS:
# the clean twin 0017 ("Tom has 8 ... and buys 4 more", no distractor) now *solves*
# 12 — genuine comprehension, not just refusal. wrong 1->0; genuine positives 7->8
# solved. The 1 remaining spurious is 0010 (multi-referent "altogether", a separate
# H1 graduation question).
_BASELINE_WRONG = 0
_BASELINE_PAIR_TELLS = 0
_BASELINE_POSITIVES_SOLVED = 8
class TestSchema:
@ -76,7 +85,10 @@ class TestProbeBaseline:
results = run_probe()
positives = [r for r in results if r.expected == "solve"]
solved = sum(1 for r in positives if r.verdict == "solved")
assert solved >= 7, f"genuine positives solving dropped to {solved}"
assert solved >= _BASELINE_POSITIVES_SOLVED, (
f"genuine positives solving dropped to {solved} (baseline "
f"{_BASELINE_POSITIVES_SOLVED})"
)
# and a positive must never be answered WRONG (that would be a real defect).
assert all(r.verdict != "wrong" for r in positives)
@ -97,14 +109,20 @@ class TestProbeBaseline:
"divisor is no longer reaching the completeness/polarity gate."
)
def test_distractor_0014_refuses_via_pooling(self) -> None:
# ADR-0182: 0014's blunt product (300) and its competing additive reading
# (25) disagree under cross-composer pooling -> refuse. Fails loudly if
# pooling regresses (the unique product would commit 300 again). 0016 is
# NOT asserted here — its distractor sits in the anchor clause and needs
# anchor-skip, a separate step (see ADR-0182 §3b).
def test_distractor_quantity_refuses(self) -> None:
# ADR-0182: both distractor confusers refuse. 0014 via the change-clause
# drop (product 300 vs additive 25); 0016 via anchor-skip + intra-clause
# (product 3840 vs additive 12). Fails loudly if either regresses.
by_id = {r.case_id: r for r in run_probe()}
assert by_id["confuser-v1-0014"].verdict == "refused"
assert by_id["confuser-v1-0016"].verdict == "refused"
def test_intra_clause_twin_0017_solves(self) -> None:
# the discrimination: 0017 ("Tom has 8 tickets and buys 4 more tickets", the
# clean twin of 0016) is genuinely *read* to 12 — comprehension, not refusal.
by_id = {r.case_id: r for r in run_probe()}
assert by_id["confuser-v1-0017"].verdict == "solved"
assert by_id["confuser-v1-0017"].answered == 12.0
def test_disguised_polarity_does_not_misfire(self) -> None:
# ADR-0182 bonus: the spurious "buys X for N <unit>" product disagrees with

View file

@ -167,3 +167,85 @@ class TestPriorStateQuestionGuard:
# discrimination: the twin asking 'left' commits the forward net (30).
resolution = resolve_pooled(_LEFT_TWIN)
assert resolution is not None and resolution.answer == 30.0
_ANCHOR_SKIP_0016 = (
"A train travels at 60 miles per hour for 2 hours. Tom has 8 tickets and "
"buys 4 more tickets. How many tickets does Tom have?"
)
_INTRACLAUSE_TWIN = "Tom has 8 tickets and buys 4 more tickets. How many tickets does Tom have?"
class TestAnchorSkipIntraClause:
"""ADR-0182 anchor-skip: a sentence packing state+change ("Tom has 8 tickets and
buys 4 more tickets") is read by splitting on the conjunction, and a leading
all-foreign block ("A train travels 60 miles ... for 2 hours") is skipped from
anchor selection. Its quantities go unused -> the pool's isolated-foreign
exemption makes the reading commit-ineligible -> it forces a disagreement refusal
on the distractor case, while the clean twin commits."""
def test_intra_clause_state_and_change_resolves(self) -> None:
# the clean twin: "has 8 ... and buys 4 more" -> 12, committed.
resolution = resolve_pooled(_INTRACLAUSE_TWIN)
assert resolution is not None and resolution.answer == 12.0
def test_anchor_skip_candidate_is_exempt(self) -> None:
# the 0016 reading skips the train block; 8+4=12 leaves 60/2 unused-foreign.
cands = accumulation_candidates(_ANCHOR_SKIP_0016)
assert cands, "expected an anchor-skip accumulation candidate"
twelve = [d for d in cands if d.answer == 12.0]
assert twelve, "expected the 8+4=12 reading"
assert classify_derivation(twelve[0], _ANCHOR_SKIP_0016) == "exempt"
def test_distractor_0016_refuses_via_disagreement(self) -> None:
# product 3840 (complete) vs additive 12 (exempt) disagree -> refuse.
assert resolve_pooled(_ANCHOR_SKIP_0016) is None
def test_no_anchor_skip_candidate_without_conjunction(self) -> None:
# a plain single-quantity sentence yields no spurious extra reading.
cands = accumulation_candidates("Sam has 14 apples. He buys 9 more apples.")
assert all(d.answer == 23.0 for d in cands)
# --- ADR-0182 lookback: the anchor-skip refuse branches (failing-under-violation) ---
def test_anchor_skip_referent_guard_discriminates_actor(self) -> None:
# NON-VACUOUS failing-under-violation: the ONLY difference between these two
# is the change sub-clause's subject (pronoun 'he' vs new name 'Sara').
# Same referent -> the 8+4=12 anchor-skip reading IS produced; new actor ->
# the referent guard suppresses it. If `_same_referent` were removed, the
# new-actor case would also produce 12 and the second assertion fails.
same_referent = (
"A train travels at 60 miles per hour for 2 hours. Tom has 8 tickets and "
"he buys 4 more tickets. How many tickets does Tom have?"
)
new_actor = (
"A train travels at 60 miles per hour for 2 hours. Tom has 8 tickets and "
"Sara buys 4 more tickets. How many tickets does Tom have?"
)
assert any(d.answer == 12.0 for d in accumulation_candidates(same_referent)), (
"same-referent anchor-skip should produce the 8+4=12 reading"
)
assert all(d.answer != 12.0 for d in accumulation_candidates(new_actor)), (
"anchor-skip chained across a new actor (Sara onto Tom)"
)
def test_anchor_skip_requires_a_polarity_cue(self) -> None:
# NON-VACUOUS: same structure, change sub-clause with vs without a licensed
# cue. "gets 4 more" -> +4 -> 12 produced; "owns 4" (no gain/loss/more cue)
# -> polarity None -> not guessed -> no candidate.
with_cue = (
"A train travels at 60 miles per hour for 2 hours. Tom has 8 tickets and "
"gets 4 more tickets. How many tickets does Tom have?"
)
no_cue = (
"A train travels at 60 miles per hour for 2 hours. Tom has 8 tickets and "
"owns 4 tickets. How many tickets does Tom have?"
)
assert any(d.answer == 12.0 for d in accumulation_candidates(with_cue))
assert all(d.answer != 12.0 for d in accumulation_candidates(no_cue))
def test_anchor_skip_refuses_without_single_quantity_anchor(self) -> None:
# No sub-clause yields a single-quantity anchor -> the anchor-skip path
# produces nothing (it does not force a multi-quantity clause to anchor).
no_anchor = "A train travels at 60 miles per hour for 2 hours and covers 5 towns."
assert accumulation_candidates(no_anchor) == ()