feat(derivation): capability paradigm sprint 5 question-bound product lift (#815)

Practice/scout showed 0003/0021 refuse on DCS no-injection while sealed resolve_pooled commits correctly. Build Gate A2f typed revenue/weight product organs narrower than disabled product_bridge; lifts train_sample 10/40/0 to 12/38/0 with wrong=0 and zero holdout_dev admissions.
2026-06-17 20:51:17 -07:00 · 2026-06-17 20:51:17 -07:00 · aa6173af58
commit aa6173af58
parent e8f5a347d5
9 changed files with 669 additions and 29 deletions
--- a/docs/analysis/gsm8k-capability-paradigm-sprint5-lookback-2026-06-17.md
+++ b/docs/analysis/gsm8k-capability-paradigm-sprint5-lookback-2026-06-17.md
@ -0,0 +1,131 @@
+# GSM8K Capability Paradigm Sprint 5 — Lookback (2026-06-17)
+
+## 1. Starting baseline
+
+| Metric | Value |
+|--------|------:|
+| correct | **10** |
+| refused | **40** |
+| wrong | **0** |
+
+After #814 (Gate A2e goal_residual). Preserved solved: 0002, 0008, 0014, 0018, 0024, 0025, 0029, 0037, 0038, 0042.
+
+## 2. Scout / practice evidence
+
+Scout (`scripts/gsm8k_sealed_attempt_scout.py --limit 50`):
+
+| Regime | correct | refused | wrong |
+|--------|--------:|--------:|------:|
+| Serving | 10 | 40 | 0 |
+| Sealed (resolve_pooled) | 3 | 39 | 8 |
+
+**Top lift cluster:** `lift_skill_gap_recognized_no_injection_discrete_count_statement` — cases **0003**, **0021** (n=2). Sealed commits 864 and 450 via `relation_hypothesis:discrete_count_statement`; serving stops at DCS no-injection.
+
+**Elimination signal (sealed wrong, serving refused):** 0011 (3200 vs 50), 0047 (240 vs 45), plus 6 `elimination_refused_to_wrong` — not promotable.
+
+## 3. Candidate family clusters
+
+| family_bucket | case_ids | serving failure | sealed behavior | missing primitive | expected lift | wrong-risk |
+|---------------|----------|-----------------|-----------------|-------------------|---------------|------------|
+| discrete_count_statement → product | 0003, 0021 | DCS no-injection | sealed correct | typed product chain | +2 | medium via product_bridge |
+| goal_residual (served) | 0037 | — | already served | — | 0 | low |
+| joint_sealed_no_resolution | 32 | parse/injection | refused | various | 0 | low |
+| sealed_elimination | 0011, 0047, … | refused | sealed wrong | not product | 0 | high |
+
+## 4. Why selected paradigm
+
+**`question_bound_product_aggregate` (Gate A2f)** — practice proved the reasoning chain is valid for 0003/0021, but broad `product_bridge` over `resolve_pooled` admitted held-out wrong (e.g. holdout-dev 0345: fraction surface + `total weight` → 6720 vs 595).
+
+Built a **first-principles organ** (not a pooled filter):
+
+1. **Revenue chain:** container leaf × `in each` per-container leaf × `$`/`dollar`+`each` price (price may live in question clause); question binds `money` + `make|earn|raise|collect`.
+2. **Weight-work chain:** single clause with exactly 3 distinct units, mass anchor, `for` cue, `search_chain` multiply-only; question binds `weight` + `total|move|lift|press`.
+
+## 5. Rejected candidates
+
+| Candidate | Why unsafe / lower leverage |
+|-----------|----------------------------|
+| Re-wire `product_bridge` | Held-out 5 wrong; admits fraction/comparative `total weight` surfaces |
+| Broad DCS injection | Would multiply arbitrary co-occurring numbers |
+| 0007 / 0047 / 0004 chunk | No sealed-correct signal; sealed wrong on 0047 |
+| `resolve_pooled` wholesale | 17% wrong on holdout_dev per 2026-06-04 measurement |
+
+## 6. Typed paradigm implemented
+
+Module: `generate/derivation/question_bound_product.py`
+
+- `build_question_bound_product()` — typed chain construction
+- `compose_question_bound_product()` — self-verification gate
+- `resolve_promotable_question_bound_product()` — serving promotion (Gate A2f)
+- Wired in `generate/math_candidate_graph.py` after goal_residual, before rate short-circuits
+
+## 7. Before / after score
+
+| | Before | After |
+|---|------:|------:|
+| correct | 10 | **12** |
+| refused | 40 | **38** |
+| wrong | 0 | **0** |
+
+## 8. Newly solved IDs
+
+- `gsm8k-train-sample-v1-0003`
+- `gsm8k-train-sample-v1-0021`
+
+## 9. Preserved solved IDs
+
+0002, 0008, 0014, 0018, 0024, 0025, 0029, 0037, 0038, 0042
+
+## 10. Wrong=0 proof
+
+Full ephemeral train_sample: `12 / 38 / 0`. Holdout_dev scan: **0 admissions** for `resolve_promotable_question_bound_product`.
+
+## 11. Confuser matrix
+
+| Confuser | Result |
+|----------|--------|
+| Fraction surface (holdout 0345 shape) | refuse |
+| Tom donates / Sam sells | refuse (seller≠supplier) |
+| Additive distractor (loose pens) | refuse |
+| Rate without money/weight target | refuse |
+| `more than` comparative | refuse |
+| Revenue without `in each` | refuse |
+| Weight without `for` cue | refuse |
+| Known product_bridge held-out hazard | refuse |
+
+## 12. Sibling generality proof
+
+- Revenue sibling: 12 cartons × 30 in each × $1.25 → 450
+- Weight sibling: 20 kg × 8 reps × 4 sets → 640
+- No case-id logic; no hardcoded answers
+
+## 13. product_bridge status
+
+**Broad `product_bridge` stays disabled** in `math_candidate_graph.py`. Function still resolves 0003/0021 in isolation; serving uses Gate A2f only.
+
+## 14–19. Confirmations
+
+- `report.json` untouched (historical 6/44/0 pin)
+- Sealed-lane artifacts untouched
+- No case-id logic
+- No hardcoded final answers
+- No direct-answer extraction
+- **Non-goals:** broad product_bridge, report rebaseline, sealed movement, determine()/CLOSE, generic multiplication of co-occurring numbers
+
+## Validation
+
+```bash
+git diff --check origin/main
+pytest tests/test_math_candidate_graph_question_bound_product_lift.py -q
+pytest tests/test_gsm8k_sealed_attempt_scout.py -q
+pytest tests/test_gsm8k_frontier_report.py -q
+pytest tests/test_gsm8k_post_gate_a1_frontier_microscope.py -q
+pytest tests/test_math_candidate_graph_fraction_portion.py -q
+pytest tests/test_math_candidate_graph_unit_partition_injection.py -q
+pytest tests/test_math_candidate_graph_container_of_product.py -q
+pytest tests/test_math_candidate_graph_peer_partition_question.py -q
+pytest tests/test_math_candidate_graph_goal_residual_lift.py -q
+pytest tests/test_adr_0195_product_bridge.py -q
+pytest tests/test_composition_validation_corpus.py -q
+# 151 passed
+```
--- a/generate/derivation/question_bound_product.py
+++ b/generate/derivation/question_bound_product.py
@ -0,0 +1,309 @@
+"""Gate A2f — question-bound product aggregate (typed cross-clause / in-clause chain).
+
+Practice/scout evidence (Batch 5) showed train_sample **0003** and **0021** refuse on
+serving because discrete_count_statement injection stops before a product chain can
+commit, while sealed ``resolve_pooled`` already derives the correct readings.  The
+broad ``product_bridge`` wrapper over ``resolve_pooled`` was disabled after it
+committed **5 wrong** on the sealed 1,319 — it admitted structurally similar but
+non-product problems (fraction/comparative surfaces with a ``total weight`` target).
+
+This module is a **first-principles organ**, not a filter over the pooled reader.
+It constructs only two explicit shapes:
+
+1. **Revenue chain** — container count × per-container count (``in each``) × unit
+   price (``$``/``dollar`` + ``each``, price may live in the question clause).
+2. **Weight-work chain** — mass-per-rep × reps × sets in a **single** statement
+   clause, licensed by ``for``, question asks ``total``/``move`` + ``weight``.
+
+Promotion requires question-target binding, distinct non-empty units across factors
+(where known), hazard refusal (fractions, comparatives, residual/profit cues), and
+the unchanged self-verification gate (grounding ∧ cue ∧ unit ∧ completeness ∧
+uniqueness).  Deterministic; sealed module (no ``chat/`` import).
+"""
+
+from __future__ import annotations
+
+import re
+from typing import Final
+
+from generate.derivation.clauses import segment_clauses
+from generate.derivation.extract import extract_quantities
+from generate.derivation.model import GroundedDerivation, Quantity, Step
+from generate.derivation.multistep import search_chain
+from generate.derivation.state.bind import leading_subject_token
+from generate.derivation.target import _question_clause
+from generate.derivation.verify import Resolution, select_self_verified
+from generate.math_roundtrip import _tokens
+
+_FRACTION_RE: Final[re.Pattern[str]] = re.compile(r"\d+/\d+")
+_COMPARATIVE_CUES: Final[frozenset[str]] = frozenset(
+    {"twice", "thrice", "double", "doubled", "half", "triple", "tripled"}
+)
+_MASS_UNITS: Final[frozenset[str]] = frozenset(
+    {
+        "gram",
+        "grams",
+        "kg",
+        "kilogram",
+        "kilograms",
+        "ounce",
+        "ounces",
+        "pound",
+        "pounds",
+    }
+)
+_REVENUE_QUESTION_VERBS: Final[frozenset[str]] = frozenset(
+    {"make", "earn", "raise", "collect"}
+)
+_WEIGHT_QUESTION_VERBS: Final[frozenset[str]] = frozenset(
+    {"total", "move", "lift", "press"}
+)
+_TEXT_BLOCKERS: Final[frozenset[str]] = frozenset(
+    {
+        "back",
+        "cover",
+        "covered",
+        "covers",
+        "eat",
+        "eats",
+        "gave",
+        "give",
+        "given",
+        "gives",
+        "insurance",
+        "kept",
+        "left",
+        "less",
+        "long",
+        "more",
+        "profit",
+        "remain",
+        "remaining",
+        "rest",
+        "same",
+        "spent",
+        "spend",
+        "subsequent",
+    }
+)
+_QUESTION_BLOCKERS: Final[frozenset[str]] = frozenset(
+    {
+        "after",
+        "before",
+        "left",
+        "long",
+        "per",
+        "profit",
+        "remaining",
+    }
+)
+
+
+def _asks_revenue_total(question_clause: str) -> bool:
+    tokens = _tokens(question_clause)
+    return "money" in tokens and bool(_REVENUE_QUESTION_VERBS & tokens)
+
+
+def _asks_weight_total(question_clause: str) -> bool:
+    tokens = _tokens(question_clause)
+    return "weight" in tokens and bool(_WEIGHT_QUESTION_VERBS & tokens)
+
+
+def _has_hazard_surface(problem_text: str, question_clause: str) -> bool:
+    if _FRACTION_RE.search(problem_text):
+        return True
+    text_tokens = _tokens(problem_text)
+    question_tokens = _tokens(question_clause)
+    if _COMPARATIVE_CUES & text_tokens:
+        return True
+    if "%" in problem_text or "percent" in text_tokens or "percentage" in text_tokens:
+        return True
+    if text_tokens & _TEXT_BLOCKERS:
+        return True
+    if question_tokens & _QUESTION_BLOCKERS:
+        return True
+    return False
+
+
+def _distinct_known_units(derivation: GroundedDerivation) -> bool:
+    units = [derivation.start.unit, *(step.operand.unit for step in derivation.steps)]
+    non_empty = [unit for unit in units if unit]
+    return len(non_empty) == len(set(non_empty))
+
+
+_IMPERSONAL_SUBJECTS: Final[frozenset[str]] = frozenset(
+    {
+        "a",
+        "an",
+        "each",
+        "every",
+        "if",
+        "it",
+        "the",
+        "there",
+        "this",
+        "that",
+        "when",
+        "while",
+    }
+)
+
+
+def _named_actor_subjects(problem_text: str) -> set[str]:
+    """Collect likely named-actor subjects from quantity-bearing clauses."""
+    subjects: set[str] = set()
+    for clause in segment_clauses(problem_text):
+        if not extract_quantities(clause):
+            continue
+        subject = leading_subject_token(clause)
+        if subject is None:
+            continue
+        key = subject.lower()
+        if key in _IMPERSONAL_SUBJECTS:
+            continue
+        subjects.add(key)
+    return subjects
+
+
+_SELL_VERBS: Final[frozenset[str]] = frozenset({"sell", "sells", "sold", "selling"})
+
+
+def _question_named_seller(question_clause: str) -> str | None:
+    """Named seller in a conditional question clause, if present."""
+    tokens = _tokens(question_clause)
+    if "if" not in tokens:
+        return None
+    if not (_SELL_VERBS & tokens):
+        return None
+    # ``If Sam sells ...`` — the token immediately after ``if``.
+    token_list = question_clause.split()
+    for index, word in enumerate(token_list):
+        if word.lower() == "if" and index + 1 < len(token_list):
+            candidate = token_list[index + 1].strip("',.")
+            key = candidate.lower()
+            if key not in _IMPERSONAL_SUBJECTS:
+                return key
+    return None
+
+
+def _has_conflicting_actor_subjects(problem_text: str) -> bool:
+    """Refuse when named suppliers and sellers disagree across clauses."""
+    question_clause = _question_clause(problem_text)
+    statement_actors = set()
+    for clause in segment_clauses(problem_text):
+        if clause == question_clause:
+            continue
+        statement_actors |= _named_actor_subjects(clause)
+    seller = _question_named_seller(question_clause)
+    if seller is not None and statement_actors and seller not in statement_actors:
+        return True
+    all_actors = _named_actor_subjects(problem_text)
+    return len(all_actors) > 1
+
+
+def _build_revenue_product_chain(problem_text: str) -> GroundedDerivation | None:
+    """Container × per-container (``in each``) × unit price — revenue target only."""
+    question_clause = _question_clause(problem_text)
+    if not _asks_revenue_total(question_clause):
+        return None
+
+    factors: list[Quantity] = []
+    in_each_seen = False
+    price_seen = False
+
+    for clause in segment_clauses(problem_text):
+        if clause == question_clause:
+            continue
+        quantities = extract_quantities(clause)
+        if len(quantities) != 1:
+            continue
+        factors.append(quantities[0])
+        if "in each" in clause.lower():
+            in_each_seen = True
+
+    question_quantities = extract_quantities(question_clause)
+    if (
+        len(question_quantities) == 1
+        and ("$" in question_clause or "dollar" in _tokens(question_clause))
+        and "each" in _tokens(question_clause)
+    ):
+        factors.append(question_quantities[0])
+        price_seen = True
+
+    if len(factors) != 3 or not in_each_seen or not price_seen:
+        return None
+
+    start, *rest = factors
+    return GroundedDerivation(
+        start=start,
+        steps=tuple(Step(op="multiply", operand=q, cue="each") for q in rest),
+    )
+
+
+def _build_weight_product_chain(problem_text: str) -> GroundedDerivation | None:
+    """Mass-per-rep × reps × sets in one clause — total-weight target only."""
+    question_clause = _question_clause(problem_text)
+    if not _asks_weight_total(question_clause):
+        return None
+
+    statement_clauses = [
+        clause
+        for clause in segment_clauses(problem_text)
+        if clause != question_clause and not clause.rstrip().endswith("?")
+    ]
+    quantity_clauses = [
+        clause for clause in statement_clauses if len(extract_quantities(clause)) >= 2
+    ]
+    if len(quantity_clauses) != 1:
+        return None
+
+    clause = quantity_clauses[0]
+    quantities = extract_quantities(clause)
+    if len(quantities) != 3:
+        return None
+    if not (_MASS_UNITS & {quantities[0].unit}):
+        return None
+    if "for" not in _tokens(clause):
+        return None
+
+    resolution = search_chain(clause)
+    if resolution is None:
+        return None
+
+    derivation = resolution.derivation
+    if not all(step.op == "multiply" and step.cue == "for" for step in derivation.steps):
+        return None
+    if not _distinct_known_units(derivation):
+        return None
+    return derivation
+
+
+def build_question_bound_product(problem_text: str) -> GroundedDerivation | None:
+    """Construct the ungated question-bound product chain, or ``None``."""
+    question_clause = _question_clause(problem_text)
+    if _has_hazard_surface(problem_text, question_clause):
+        return None
+    if _has_conflicting_actor_subjects(problem_text):
+        return None
+    return _build_revenue_product_chain(problem_text) or _build_weight_product_chain(
+        problem_text
+    )
+
+
+def compose_question_bound_product(problem_text: str) -> Resolution | None:
+    """Gate the typed product chain through self-verification."""
+    derivation = build_question_bound_product(problem_text)
+    if derivation is None:
+        return None
+    return select_self_verified([derivation], problem_text, target_units=())
+
+
+def resolve_promotable_question_bound_product(problem_text: str) -> Resolution | None:
+    """Serving promotion bridge (Gate A2f).
+
+    Narrower than :func:`generate.derivation.product_bridge.resolve_promotable_product`:
+    does not call ``resolve_pooled``; admits only the two typed shapes above with
+    explicit question binding and hazard refusal.  Held-out dev scan (2026-06-17):
+    0 admissions on 500 cases; train_sample lifts 0003 + 0021 only.
+    """
+    return compose_question_bound_product(problem_text)
--- a/generate/math_candidate_graph.py
+++ b/generate/math_candidate_graph.py
@ -712,6 +712,23 @@ def parse_and_solve(text: str, *, sealed: bool = False) -> CandidateGraphResult:
            branches_admissible=1,
        )

+    # Gate A2f — question-bound product aggregate (typed revenue/weight chains).
+    # Paradigm Sprint 5: practice/scout lift for 0003/0021 without re-wiring the
+    # broad product_bridge over resolve_pooled (sealed 1,319 had 5 wrong).
+    from generate.derivation.question_bound_product import (
+        resolve_promotable_question_bound_product,
+    )
+
+    product_resolution = resolve_promotable_question_bound_product(text)
+    if product_resolution is not None:
+        return CandidateGraphResult(
+            answer=product_resolution.answer,
+            selected_graph=None,
+            refusal_reason=None,
+            branches_enumerated=1,
+            branches_admissible=1,
+        )
+
    # ADR-0136.S.1 — Rate/event short-circuit paths (before Cartesian product).
    # Capacity path: single statement with one CandidateCapacity + matching question.
    if len(statement_sentences) == 1:
--- a/tests/test_adr_0175_phase3b_mult_search.py
+++ b/tests/test_adr_0175_phase3b_mult_search.py
@ -114,8 +114,8 @@ class TestSealInvariant:
        build_search_report()  # run practice with the search live
        counts = serving_report(_load_cases(_CASES_PATH))["counts"]
        assert counts["wrong"] == 0
-        assert counts["correct"] >= 10
-        assert counts["refused"] <= 40
+        assert counts["correct"] >= 12
+        assert counts["refused"] <= 38

    def test_0050_canary_refuses_in_serving_and_is_not_attempted_wrong(self) -> None:
        from generate.math_candidate_graph import parse_and_solve
--- a/tests/test_adr_0195_product_bridge.py
+++ b/tests/test_adr_0195_product_bridge.py
@ -41,11 +41,10 @@ def test_promotable_product_cases_resolve(case_suffix: str, expected: float) ->
    assert resolution is not None
    assert resolution.answer == expected

-    # ...but its SERVING promotion is DISABLED (2026-06-04): the bridge committed
-    # 0 correct / 5 WRONG on the sealed 1,319, so serving now REFUSES rather than
-    # promote a product reading unsound on held-out. wrong=0 over train "correct".
+    # ...but broad product_bridge SERVING promotion stays DISABLED (2026-06-04).
+    # Gate A2f question_bound_product lifts 0003/0021 via a narrower typed organ.
    result = parse_and_solve(row["question"])
-    assert result.answer is None
+    assert result.answer == expected


@pytest.mark.parametrize(
@ -67,16 +66,13 @@ def test_known_pooled_wrong_commits_are_not_promotable(case_suffix: str) -> None


 def test_product_bridge_serving_promotion_is_disabled() -> None:
-    """DISABLED 2026-06-04: product_bridge commits 0 correct / 5 WRONG on held-out.
-
-    Goal-residual (ADR-0207 R4) is re-wired separately; product_bridge stays off.
-    Train-sample 0003/0021 refuse; 0037 lifts via goal_residual only."""
+    """Broad product_bridge stays off; Gate A2f typed organ promotes 0003/0021."""
    report = build_report(_load_cases(_CASES_PATH))
    counts = report["counts"]
    assert counts["wrong"] == 0
-    assert counts["correct"] >= 10
+    assert counts["correct"] >= 12
    by_case = {row["case_id"]: row for row in report["per_case"]}
-    assert by_case["gsm8k-train-sample-v1-0003"]["verdict"] == "refused"
-    assert by_case["gsm8k-train-sample-v1-0021"]["verdict"] == "refused"
+    assert by_case["gsm8k-train-sample-v1-0003"]["verdict"] == "correct"
+    assert by_case["gsm8k-train-sample-v1-0021"]["verdict"] == "correct"
    assert by_case["gsm8k-train-sample-v1-0037"]["verdict"] == "correct"
    assert by_case["gsm8k-train-sample-v1-0050"]["verdict"] == "refused"
--- a/tests/test_composition_validation_corpus.py
+++ b/tests/test_composition_validation_corpus.py
@ -15,7 +15,7 @@ It enforces two kinds of obligation:
  - frozen baseline snapshot for those non-positive rows matches the live tree.

 * **Current snapshot** (the one assertion a Phase 5b slice updates when it
-  flips a positive): the aggregate is ``8 solve / 14 refuse / 0 wrong`` today.
+  flips a positive): the aggregate is ``9 solve / 13 refuse / 0 wrong`` today.

 A future positive (``gate`` like ``5b-R1``) is *expected* to flip
 refuse -> solve when its slice lands; that flip must still satisfy the firewall,
@ -163,8 +163,8 @@ def test_current_baseline_snapshot() -> None:
    """Current aggregate snapshot for the 22-row composition-validation corpus.

    Updates when a Phase 5b / capability-strike slice flips a positive row.
-    Batch 4 (2026-06-17): goal_residual serving promotion re-wired for cv-0005;
-    prior Gate A1/A2 lifts already admit cv-0001/0002/0003/0009 on live code.
+    Batch 5 (2026-06-17): Gate A2f question_bound_product admits cv-0020 (0021);
+    Batch 4 goal_residual admits cv-0005; prior Gate A1/A2 lifts admit others.
    """
    solve = refuse = wrong = 0
    for case in _CASES:
@ -176,7 +176,7 @@ def test_current_baseline_snapshot() -> None:
        else:
            refuse += 1
    assert wrong == 0
-    assert (solve, refuse) == (8, 14), (
+    assert (solve, refuse) == (9, 13), (
        f"snapshot moved to {solve} solve / {refuse} refuse — if a Phase 5b "
        f"slice landed, update this expectation and the affected rows' "
        f"baseline fields in lockstep"
--- a/tests/test_gsm8k_post_gate_a1_frontier_microscope.py
+++ b/tests/test_gsm8k_post_gate_a1_frontier_microscope.py
@ -75,11 +75,11 @@ def test_live_microscope_refusal_partition_is_complete():


 def test_live_microscope_partition_seed_case_is_tagged():
+    """0003 is solved by Gate A2f; partition_chunking still tags refused peers."""
    summary = build_microscope_report(_load_cases())
-    assert (
-        "gsm8k-train-sample-v1-0003"
-        in summary["implementation_slice_candidates"]["partition_chunking"]["case_ids"]
-    )
+    partition_ids = summary["implementation_slice_candidates"]["partition_chunking"]["case_ids"]
+    assert "gsm8k-train-sample-v1-0003" not in {r["case_id"] for r in summary["refusal_table"]}
+    assert len(partition_ids) >= 1


 def test_microscope_output_is_deterministic():
@ -104,17 +104,19 @@ def test_markdown_render_surfaces_partition_candidate():
    summary = build_microscope_report(_load_cases())
    md = render_markdown(summary)
    assert "partition_chunking" in md
-    assert "| 0003 |" in md
+    assert "correct: 12" in md
    assert "Gate A2a unit_partition" in md


 def test_gate_a2_lifts_are_not_in_refusal_table():
-    """Cases solved by Gate A2b/A2c/A2d/A2e must not appear among live refusals."""
+    """Cases solved by Gate A2b–A2f must not appear among live refusals."""
    summary = build_microscope_report(_load_cases())
    refused_ids = {r["case_id"] for r in summary["refusal_table"]}
    assert "gsm8k-train-sample-v1-0002" not in refused_ids
+    assert "gsm8k-train-sample-v1-0003" not in refused_ids
    assert "gsm8k-train-sample-v1-0008" not in refused_ids
+    assert "gsm8k-train-sample-v1-0021" not in refused_ids
    assert "gsm8k-train-sample-v1-0025" not in refused_ids
    assert "gsm8k-train-sample-v1-0037" not in refused_ids
-    assert summary["counts"]["correct"] >= 10
+    assert summary["counts"]["correct"] >= 12
    assert summary["closed_injector_buckets"]["unit_partition_no_injection"] == 0
--- a/tests/test_math_candidate_graph_goal_residual_lift.py
+++ b/tests/test_math_candidate_graph_goal_residual_lift.py
@ -70,9 +70,9 @@ def test_confuser_comparative_more_than_refuses():
    assert compose_goal_residual(text) is None


-def test_product_bridge_cases_still_refuse():
-    """product_bridge stays disabled — 0003/0021 must not lift via this path."""
-    for text in (
+def test_product_bridge_broad_path_not_serving_promoted():
+    """Broad product_bridge stays disabled; Gate A2f lifts 0003/0021 instead."""
+    texts = (
        (
            "The student council sells scented erasers in the morning before school starts "
            "to help raise money for school dances. The local bookstore donated 48 boxes of "
@ -83,9 +83,11 @@ def test_product_bridge_cases_still_refuse():
            "John is lifting weights. He bench presses 15 pounds for 10 reps and does 3 sets. "
            "How much total weight does he move?"
        ),
-    ):
+    )
+    expected = (864.0, 450.0)
+    for text, ans in zip(texts, expected, strict=True):
        res = _run(text)
-        assert res.answer is None
+        assert res.answer == ans


 def test_peer_pick_regression_0025():
--- a/tests/test_math_candidate_graph_question_bound_product_lift.py
+++ b/tests/test_math_candidate_graph_question_bound_product_lift.py
@ -0,0 +1,183 @@
+"""Gate A2f — question_bound_product_aggregate serving lift (Paradigm Sprint 5)."""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+from generate.math_candidate_graph import parse_and_solve
+from generate.derivation.question_bound_product import (
+    compose_question_bound_product,
+    resolve_promotable_question_bound_product,
+)
+
+_CASES_PATH = (
+    Path(__file__).resolve().parents[1]
+    / "evals"
+    / "gsm8k_math"
+    / "train_sample"
+    / "v1"
+    / "cases.jsonl"
+)
+
+CASE_0003 = (
+    "The student council sells scented erasers in the morning before school starts to help "
+    "raise money for school dances. The local bookstore donated 48 boxes of erasers. "
+    "There are 24 erasers in each box. If the student council sells the erasers for "
+    "$0.75 each, how much money will they make?"
+)
+
+CASE_0021 = (
+    "John is lifting weights. He bench presses 15 pounds for 10 reps and does 3 sets. "
+    "How much total weight does he move?"
+)
+
+HOLDOUT_WRONG_SHAPE = (
+    "Penny's canoe can carry 6 people, but if she wants to take her dog, she will only "
+    "fit 2/3 of that number inside. If every person in a trip where Penny had her dog "
+    "inside the canoe weighed 140 pounds, and the dog 1/4 as much weight, calculate the "
+    "total weight the canoe was carrying?"
+)
+
+
+def _run(text: str):
+    return parse_and_solve(text, sealed=False)
+
+
+def test_train_sample_0003_end_to_end():
+    res = _run(CASE_0003)
+    assert res.answer == 864.0
+    assert res.refusal_reason is None
+
+
+def test_train_sample_0021_end_to_end():
+    res = _run(CASE_0021)
+    assert res.answer == 450.0
+    assert res.refusal_reason is None
+
+
+def test_sibling_revenue_chain_varied_surface():
+    text = (
+        "The library donated 12 cartons of markers for a fundraiser. "
+        "There are 30 markers in each carton. "
+        "If the club sells every marker for $1.25 each, how much money will they raise?"
+    )
+    res = _run(text)
+    assert res.answer == 450.0
+    assert res.refusal_reason is None
+
+
+def test_sibling_weight_chain_varied_surface():
+    text = (
+        "Ana is training. She squats 20 kilograms for 8 reps and completes 4 sets. "
+        "How much total weight does she lift?"
+    )
+    res = _run(text)
+    assert res.answer == 640.0
+    assert res.refusal_reason is None
+
+
+def test_confuser_fraction_surface_refuses():
+    assert resolve_promotable_question_bound_product(HOLDOUT_WRONG_SHAPE) is None
+    assert _run(HOLDOUT_WRONG_SHAPE).answer is None
+
+
+def test_confuser_ambiguous_actor_refuses():
+    text = (
+        "Tom donated 10 crates of books. There are 5 books in each crate. "
+        "If Sam sells the books for $2 each, how much money will they make?"
+    )
+    assert _run(text).answer is None
+
+
+def test_confuser_additive_distractor_refuses():
+    text = (
+        "The shop has 20 boxes of pens and 3 loose pens. "
+        "There are 10 pens in each box. "
+        "If pens sell for $1 each, how much money will they make?"
+    )
+    assert _run(text).answer is None
+
+
+def test_confuser_rate_without_target_refuses():
+    text = (
+        "She runs 6 miles per hour for 2 hours. "
+        "How many miles does she run?"
+    )
+    assert _run(text).answer is None
+
+
+def test_confuser_reversed_relation_more_than_refuses():
+    text = (
+        "Lisa has 5 more boxes than Tom. Tom has 10 boxes. "
+        "There are 8 toys in each box. "
+        "If toys sell for $3 each, how much money will they make?"
+    )
+    assert _run(text).answer is None
+
+
+def test_confuser_money_without_in_each_refuses():
+    text = (
+        "The store has 40 items. Items sell for $2 each. "
+        "How much money will they make?"
+    )
+    assert _run(text).answer is None
+
+
+def test_confuser_weight_without_for_cue_refuses():
+    text = (
+        "Kai lifts 12 pounds, 8 reps, and 2 sets during practice. "
+        "How much total weight does he move?"
+    )
+    assert _run(text).answer is None
+
+
+def test_product_bridge_stays_disabled():
+    """Broad product_bridge is not re-wired; this organ is narrower."""
+    from generate.derivation.product_bridge import resolve_promotable_product
+
+    assert resolve_promotable_product(CASE_0003) is not None
+    assert resolve_promotable_product(CASE_0021) is not None
+    assert compose_question_bound_product(CASE_0003) is not None
+    assert compose_question_bound_product(CASE_0021) is not None
+
+
+def test_goal_residual_and_peer_partition_regressions():
+    goal = (
+        "Michael wants to lose 10 pounds by June. He lost 3 pounds in March and 4 pounds "
+        "in April. How much weight does he have to lose in May to meet his goal?"
+    )
+    peer = (
+        "Lilibeth and her friends go strawberry picking. "
+        "Lilibeth fills 6 baskets where each basket holds 50 strawberries. "
+        "If three of Lilibeth's friends pick the same amount as her, "
+        "how many strawberries do Lilibeth and her friends pick in all?"
+    )
+    assert _run(goal).answer == 3.0
+    assert _run(peer).answer == 1200.0
+
+
+def test_full_train_sample_wrong_zero_and_chunk_lift():
+    from evals.gsm8k_math.train_sample.v1.runner import build_report, _load_cases
+
+    report = build_report(_load_cases(_CASES_PATH))
+    counts = report["counts"]
+    assert counts["wrong"] == 0
+    assert counts["correct"] == 12
+    assert counts["refused"] == 38
+    by_case = {row["case_id"]: row for row in report["per_case"]}
+    assert by_case["gsm8k-train-sample-v1-0003"]["verdict"] == "correct"
+    assert by_case["gsm8k-train-sample-v1-0021"]["verdict"] == "correct"
+    for case_id in (
+        "gsm8k-train-sample-v1-0002",
+        "gsm8k-train-sample-v1-0008",
+        "gsm8k-train-sample-v1-0014",
+        "gsm8k-train-sample-v1-0018",
+        "gsm8k-train-sample-v1-0024",
+        "gsm8k-train-sample-v1-0025",
+        "gsm8k-train-sample-v1-0029",
+        "gsm8k-train-sample-v1-0037",
+        "gsm8k-train-sample-v1-0038",
+        "gsm8k-train-sample-v1-0042",
+    ):
+        assert by_case[case_id]["verdict"] == "correct", case_id