core/docs/decisions/ADR-0136.S.4-novel-initial-form.md

ADR-0136.S.4 — Novel Initial-Form Subject-Slot Widenings

Status: Accepted
Parent: ADR-0136 (Statement Layer Corridor)
Date: 2026-05-23

Context

The v3 refusal taxonomy (post-S.3 rescan) identified 5 cases with novel_initial_form as their primary barrier. These cases have initial-state sentences that the existing _INITIAL_HAS_RE and _INITIAL_THERE_ARE_RE cannot parse because the subject slot uses a form neither pattern admits.

The 5 cases and their S.4 disposition:

Case	Primary Barrier	S.4 Target Shape	Secondary Barriers	S.4 Result
0028	novel_initial_form	—	rate_price + temporal_frequency	No change (multi-clause)
0030	novel_initial_form	—	compound_comparative	No change (impersonal "It" subject)
0038	novel_initial_form	Shape B (prep-prefix existential)	compound_comparative on sentence 2	Barrier shift
0045	novel_initial_form	—	rate_price	No change (rate-price shape)
0046	novel_initial_form	Shape A (indefinite-article subject)	fraction_operand on sentence 2	Barrier shift

Only 2 of the 5 cases are single-shape targets addressable by S.4. The remaining 3 have load-bearing secondary barriers that S.4 cannot resolve:

• 0028: "It cost $100,000 to open a bakery" — impersonal "It" subject + rate_price + temporal_frequency. All three barriers must fall together.
• 0030: "It takes 2 times as many" — impersonal "It" + compound_comparative.
• 0045: Sentence 1 is actually a rate-price shape ("The price of 1 orange is $0.50"), misclassified as novel_initial_form in v2; S.4 does not target it.

Decision

Two sibling regexes, not a widened _ENTITY

_ENTITY = r"(?:[A-Z]\w+|[Tt]he\s+\w+)" is used by many patterns across the file. Widening it to include indefinite-article forms would cascade through all regex sites, breaking the closed-grammar discipline of every shape that depends on it.

Instead S.4 adds two sibling patterns (_INITIAL_HAS_INDEF_RE and _INITIAL_THERE_ARE_PREFIX_RE) with their own entity slots, applied in dedicated extractor functions that are wired at the end of extract_initial_candidates. All other regex sites are unchanged.

Shape A — indefinite-article subject

^[Aa]\s+<noun>\s+has\s+<N>(?:\s+<adjective>)?(?:\s+<unit>)?(?:\s+of\s+<substance>)?\.$

Restrictions:

• [Aa]\s+ — capital or lowercase "A" followed by whitespace only. "An" is deliberately excluded: [Aa] does not match the "n" that follows in "An", so the \s+ fails. This avoids collision with money-amount or other shapes that may start with "an".
• Anchor: has only (singular third-person). "A school have" is ungrammatical; widening to "have/had" is deferred until a real GSM8K case requires it.
• Entity stored as the bare noun, lowercased (e.g., "school", "factory").
• All existing value-slot widenings (_VALUE) and substance qualifiers apply.

gsm8k-0046 sentence 1: "A school has 100 students." → InitialPossession(entity="school", quantity=Quantity(100, "students"))

Shape B — prepositional-prefix existential

^In\s+[Aa]\s+<place>,?\s+there\s+(?:are|were|is|was)\s+(?:a\s+)?<N>\s+<unit>
(?:\s+on\s+the\s+\w+(?:-floor)?)?(?:\s+\w+ing)?\.$

Restrictions:

• Prefix: In\s+[Aa]\s+ — "In a" or "In A" only. Other prepositions ("At a", "On a") are deferred.
• (?:a\s+)? before the value slot: handles the "a hundred" construction where "a" is an article for the numeral phrase. The article is consumed without being captured; the value group sees "hundred" which resolves via WORD_NUMBERS.
• Ordinal-floor qualifier (on the first-floor) and participial phrase (studying, shelving) are optional and discarded.
• Entity stored as the bare place noun, lowercased (e.g., "building").
• Participial pattern \w+ing matches present participles only; past participles ("shelved") are not matched — consistent with the only GSM8K target sentence (gsm8k-0038 uses "studying").

gsm8k-0038 sentence 1: "In a building, there are a hundred ladies on the first-floor studying." → InitialPossession(entity="building", quantity=Quantity(100, "ladies"))

No short-circuit path

Both extractors produce CandidateInitial objects that flow through the existing graph machinery. No bypass route is added in parse_and_solve.

Per-Case Impact

Case	Pre-S.4 Barrier	S.4 Shape	S.4 Effect	Post-S.4 Barrier
0028	novel_initial_form	—	No change	novel_initial_form + rate_price + temporal_frequency
0030	novel_initial_form	—	No change	novel_initial_form (impersonal "It")
0038	novel_initial_form	Shape B	Shifted	compound_comparative (sentence 2: "three times that many")
0045	novel_initial_form	—	No change	rate_price (misclassification; sentence 1 is a rate-price shape)
0046	novel_initial_form	Shape A	Shifted	fraction_operand (sentence 2: "Half of the students are girls")

Barrier shifts: +2 (0038 → compound_comparative; 0046 → fraction_operand).

GSM8K Admission Count

Metric	Pre-S.4	Post-S.4
Admitted (correct)	3	3
Wrong	0	0
Refused	47	47
Barrier shifted	—	2 (0038, 0046)

Direct admission delta: 0. Both 0038 and 0046 shift their barrier from sentence 1 to sentence 2, which introduces a barrier that S.4 does not address.

Axis Lane

20 curated cases at evals/math_capability_axes/S4_novel_initial_form/v1/:

Category	Count	Gate
indefinite_article_canonical	4	answer == expected
indefinite_article_substance	3	answer == expected
prep_prefix_canonical	4	answer == expected
prep_prefix_ordinal_participial	3	answer == expected
refusal_missing_unit	2	answer is None
refusal_indefinite_quantity	2	answer is None
refusal_definite_article_regression	2	answer == expected (regression gate)

wrong == 0 across all 20 cases.

The refusal_definite_article_regression category verifies that definite-article subjects ("The school has 100 students.") still resolve through the existing _INITIAL_HAS_RE path and are not blocked by the new sibling regex.

Deferred

• Impersonal "It" subjects — "It cost $100,000 to open." The subject slot "It" is not a named entity; resolving it requires coreference or a special implicit-entity rule. Cases 0028 and 0030 have this form.
• "An" article prefix — "An apple has 10 seeds." Excluded from [Aa]\s+ to avoid collision with money-amount or other shapes; no GSM8K target requires it.
• Global _ENTITY widening — cascade-unsafe; every shape depending on _ENTITY would need independent testing.
• Multi-clause cases — 0028 has rate_price + temporal_frequency as co-barriers; addressing them requires S.1-level rate-parsing and multi-clause composition, both out of S.4 scope.
• Other preposition prefixes — "At a school, there are..." / "On a floor, there are..." — no GSM8K targets require them in the current 50-case sample.

PR Checklist Answers

• Capability added: Two closed subject-slot widenings for initial-state sentences; 20-case axis lane with wrong == 0.
• Invariant: versor_condition(F) < 1e-6 is field-level; this change is pure parser — no algebra path touched.
• CLI lane: S.4 axis runner + S.3 + S.1 + rescan-v3 gates all pass.
• No hidden normalization: No drift repair, no approximate recall, no stochastic fallback introduced.
• Trust boundary: User-supplied text enters via parse_and_solve (existing boundary). Both new patterns are closed-grammar (no wildcards); unmatched sentences refuse, not wrong.