core/docs/briefs/parallel-2026-05-23/L12-ADR-0131-G4-multi-clause.md

L12 brief — ADR-0131.G.4 — Capability axis: multi-clause composition (conjunctions, distributive subjects, embedded quantifier phrases)

Worktree setup (do this first, non-negotiable):

git worktree add ../core-adr-0131-g4-multi-clause -b feat/adr-0131-g4-multi-clause origin/main
cd ../core-adr-0131-g4-multi-clause

Scope. Capability-axis iteration: extend the candidate parser + binding graph to handle within-sentence composition that the per-statement parser currently refuses on. Baseline clusters: Aaron and his brother Carson each saved up $40 (conjoined subject + distributive each), Francine has five full boxes of crayons and 5 loose crayons (conjoined object NPs sharing a verb), Ella has 4 bags with 20 apples in each bag and six bags with 25 apples in each bag (embedded quantifier with N <unit> in each <bag> + conjunction).

Target shapes (closed set):

• Conjoined subjects with each: <A> and <B> each <verb> <N> <unit>. → emits two InitialPossession candidates (one per actor), same (N, unit).
• Conjoined object NPs sharing a verb: <Entity> has <N1> <unit1> and <N2> <unit2>. → emits two InitialPossession candidates for the same entity.
• Embedded quantifier phrases: <Entity> has <N> <container> with <M> <unit> in each <container>. → emits a derived InitialPossession with value = N*M, unit = <unit> only if the round-trip filter admits the product. The multiplication is a candidate, not a guarantee; the binding graph picks admissible compositions.
• Conjoined embedded quantifiers: <Entity> has <N1> <container> with <M1> <unit> in each and <N2> <container> with <M2> <unit> in each. → emits two derived candidates and a sum candidate.

This is the highest-risk axis of the four — multi-clause composition is where confabulation risk is highest. Refusal-first stays paramount; admission gains must be small and load-bearing, not maximum-rate-chasing.

Reference docs (read these, only these):

1. docs/decisions/ADR-0131.G-gsm8k-coverage-probe.md — iteration discipline. The "smell test" (admission moves on GSM8K but new axis cases don't all pass → reject) bites hardest here.
2. generate/math_candidate_graph.py — the candidate-graph topology. New candidates must compose through the same graph; nothing in the binding/admissibility layer changes.

What to ship:

• Parser extension in generate/math_candidate_parser.py: three new extractors (_conj_subject_each_candidates, _conj_object_candidates, _embedded_quantifier_candidates) emitting multiple candidates per match. Source-span provenance covers the full sentence for each candidate.
• Optional graph-side note in generate/math_candidate_graph.py (read-only audit; only edit if a composed candidate is unreachable through existing edges — if so, the edge addition is a one-line widening, not a new admissibility rule). Decision to edit or not is documented in the ADR.
• Curated coverage cases at evals/math_capability_axes/G4_multi_clause/v1/cases.jsonl (~30 cases): ≥6 per shape + ≥6 refusal probes for shapes that look multi-clause but are not in the closed set (e.g. cross-sentence coreference Aaron has 5. He gives 2 to Bob., ambiguous each scope, three-way conjunctions). Refusal probes are load-bearing — they pin the scope boundary the architecture refuses to cross.
• Runner + report at evals/math_capability_axes/G4_multi_clause/v1/.
• Tests at tests/test_adr_0131_G4_multi_clause.py (~15): per-shape at-least-one passing, refusal probes refuse typed, wrong == 0 (especially load-bearing here), replay byte-equality, GSM8K probe re-run with admission strictly increases OR multi-clause refusals strictly decrease (declare in ADR, gate on it), B3 + G.1/G.2/G.3 lanes unchanged.
• ADR docs/decisions/ADR-0131.G.4-multi-clause.md. Cite ADR-0131.G parent and ADR-0126 (candidate graph). Pin the closed shape set; document the each-scope policy (always distributive, never collective — refuse collective readings); document why cross-sentence coreference stays deferred.
• Refresh evals/gsm8k_math/train_sample/v1/train_sample_coverage_report.json.

Hard constraints:

• wrong == 0 is non-negotiable. Multi-candidate emission means the round-trip filter does more work — if any composed candidate slips a wrong answer, remove the shape, do not weaken the filter.
• Closed shape set. Every recognized multi-clause structure matches exactly one of the listed extractors. No paraphrase tolerance.
• No cross-sentence state. This axis is strictly within-sentence. Pronoun/coreference across sentences stays refused.
• Distributive each only. Collective readings (Aaron and Carson saved $40 together) must refuse — explicit adversarial probe required.
• No solver changes. If a multi-clause case parses but does not solve, file as follow-up ADR. Do not stub.
• No new modules under algebra/, chat/, core/.
• Determinism. Multi-candidate ordering pinned; report byte-equal across runs.

Out of scope: verb classes (L9/G.1), comparatives (L10/G.2), numeric literals (L11/G.3), cross-sentence coreference, ellipsis (Aaron has 5, Carson 3), three-way+ conjunctions, collective readings.

Target branch. PR against main. Title: feat(ADR-0131.G.4): multi-clause composition — admission N/50 (Δ+N). Body: per-shape case counts, refusal-set documentation, admission or refusal-family delta, explicit acknowledgement that this is the highest-risk axis and the wrong == 0 evidence to back it.

Exit criterion. CI green; multi-clause runner exits 0 with wrong == 0; chosen GSM8K-probe gate satisfied; B3 + L9/G.1 + L10/G.2 + L11/G.3 lanes unchanged; refreshed coverage report committed.

Do not stack on another agent's branch. Target main directly.