core/docs/handoff/VERB-EXPANSION-DISPATCH-PACK.md

Verb-Coverage Expansion — Dispatch Pack

Goal: Widen the narrowness boundary that gates discrete_count_statement admission so the held-hypothesis substrate landed in ADR-0174 Phase 3a has cases to act on. Concretely: review the 11 verbs surfacing in evals/gsm8k_math/train_sample/v1 empty-anchor refusals and decide, per verb, whether to add to _ACQUISITION_VERBS / _POSSESSION_VERBS, add to a new depletion-verb class, or refuse and route to a different recognizer category.

Parent ADRs: ADR-0163 (path to GSM8K mastery), ADR-0150/0152/0155/0161 (contemplation → proposal → HITL → ratification corridor), ADR-0167 (audit-as-teaching-evidence), ADR-0174 (held-hypothesis comprehension)

Parent briefs:

• docs/handoff/PHASE-3.1-FOLLOWUP-RECOGNIZER-EXPANSION.md — analysis identifying the verb-coverage bottleneck.

Type: Implementation dispatch + operator decision pack; not a doctrine ADR. Touches a hardcoded narrowness boundary so every acceptance is operator-reviewed and individually justified against the wrong = 0 invariant.

---

Why this brief exists

ADR-0174 Phase 3a (PR #423) shipped the lookback substrate (reevaluate, PronounResolution, held-hypothesis emission). The substrate is correct and tested but does not fire on any train_sample case because all 21 empty-anchor discrete_count_statement refusals fail at the verb-whitelist layer before reaching the pronoun-resolution check.

The whitelist lives in generate/recognizer_match.py:

_POSSESSION_VERBS: Final[frozenset[str]] = frozenset({
    "has", "have", "had",
})

_ACQUISITION_VERBS: Final[frozenset[str]] = frozenset({
    "collected", "collects", "collect",
    "received", "receives", "receive",
    "bought", "buys", "buy",
    "got", "gets", "get",
})

This is a deliberate narrowness boundary, not registry-derived data. Widening it is a code change with operator review (not a pack mutation through the HITL corridor). The brief below produces the evidence the operator needs to decide each verb.

---

Bundling rule

Multiple PRs, one per verb class (not one per verb). Each PR ships:

1. The narrowness expansion in generate/recognizer_match.py
2. A test in tests/test_adr_0174_phase3_lookback.py (or sibling) asserting the widened class admits cases the held-hypothesis substrate now handles
3. A wrong = 0 proof against train_sample/v1 AND case 0050 (the canary).

Three suggested PRs, each independently mergeable:

• VE-A — Acquisition widening (feat/verb-expansion-acquisition) Adds reviewed acquisition verbs to _ACQUISITION_VERBS.
• VE-B — Depletion class introduction (feat/verb-expansion-depletion) Adds a new _DEPLETION_VERBS: Final[frozenset[str]] and routes it to CandidateOperation(kind="subtract"). Includes its own anchor_kind="depletion" branch in inject_discrete_count_statement.
• VE-C — Refusal-typed evidence (feat/verb-expansion-non-arithmetic) For verbs that should NOT widen any whitelist (the verb does not grammatically denote possession / acquisition / depletion), emit a structured refusal-evidence record so the contemplation lane can decide whether a new recognizer category is needed (e.g. capacity_statement, descriptive_action_frame). No production whitelist changes; just better trace.

VE-A and VE-B are independent. VE-C requires nothing in the others. All three can be in flight in parallel.

---

Operator profile

• Opus for VE-A (load-bearing wrong=0 surface — widening the acquisition class touches every train_sample case via the recognizer pipeline). Case 0050 hazard pin is mandatory.
• Opus for VE-B (adds an entirely new operation-kind path; same wrong=0 hazard surface). Includes the depletion-class case-0050 hazard check (a depletion-verb addition must not flip case 0050 from refused to wrong).
• Sonnet for VE-C (instrumentation-only; no production behavior change). Tight scope, no architectural risk.

---

The 11 verbs — recommended classification

Each row is the operator's input for the dispatch. Final classification is the operator's call after reviewing the contemplation evidence; this column is a recommendation only.

Verb (lemma)	Surface forms in evidence	Recommended	Why
gain	gained, gains, gain	Acquisition (VE-A)	Grammatically gains quantity to actor: "Orlando gained 5 pounds" → +5. Hazard: gained can be delta-of-attribute (weight, age) — the existing _ACQUISITION_VERBS comment explicitly excludes it for this reason. Operator must confirm.
earn	earned, earns, earn	Acquisition (VE-A)	Same shape as gained; safer because monetary/discrete-count semantics are more uniform.
save	saved, saves, save	Acquisition (VE-A)	Ambiguous per existing comment ("saved time" vs "saved money"). Refusal-preferring approach: require the noun class (time vs money) to gate admission.
accumulate	accumulated, accumulates, accumulate	Acquisition (VE-A)	Unambiguous.
acquire	acquired, acquires, acquire	Acquisition (VE-A)	Unambiguous.
donate	donated, donates, donate	Depletion (VE-B)	Grammatically removes quantity from actor: "The bookstore donated 48 boxes" → -48. Subject is the giver.
give	gave, gives, give	Depletion (VE-B)	Same shape as donate. Existing TRANSFER_VERBS handles give to X — depletion is the no-target form.
lose	lost, loses, lose	Depletion (VE-B)	"She lost 3 marbles" → -3. Hazard: lost can semantically invert direction in comparisons ("Alice lost 3 more than Bob"). Existing comparison code excludes lost for this reason. Depletion-class addition must preserve that exclusion in comparison contexts.
spend	spent, spends, spend	Depletion (VE-B)	Money-class depletion; safer than lose because monetary context constrains semantics.
eat	eats, ate, eat	Depletion (VE-B)	"The guests eat 1 pan" → -1. Constrained to food-noun classes for safety.
bench presses / splits / runs / bakes / invests / is / wants	—	Non-arithmetic (VE-C)	These verbs do NOT carry possession/acquisition/depletion semantics. Routing them through the discrete_count_statement recognizer is the wrong intent. They want a separate recognizer category (e.g. rate_statement, capacity_statement, descriptive_frame) that is out of scope for this brief.

Hazard pinning — non-negotiable

Every VE-A / VE-B PR MUST include two safety tests.

Hazard test #1 — case 0050 canary

def test_case_0050_remains_refused_after_verb_widening(self) -> None:
    """gsm8k-train-sample-v1-0050 is the wrong=0 canary. The verb
    widening must NOT flip it from refused to wrong."""
    from generate.math_candidate_graph import parse_and_solve
    text = (
        "Mark does a gig every other day for 2 weeks. "
        "He gets paid $50 per gig. He then gets a 50% raise. "
        "How much money does he make per week?"
    )
    r = parse_and_solve(text)
    assert r.answer is None, (
        f"case 0050 wrong=0 hazard violated: verb widening admitted "
        f"answer {r.answer!r}; must remain refused"
    )

Reference: memory feedback-wrong-zero-hazard-case-0050.

Hazard test #2 — multi-actor pronoun safety (ADR-0174 Phase 3a)

VE-A / VE-B widens the verb whitelist, which means more discrete_count_statement cases will reach the ADR-0174 Phase 3a lookback wiring. When the widened verb appears in a multi-actor problem, the Phase 3a defense (no_antecedent_ambiguous refusal on multiple distinct prior subjects) MUST fire — otherwise the pronoun could be mis-resolved and produce a wrong answer with no downstream safety net catching it.

def test_multi_actor_pronoun_with_widened_verb_refuses_safely(self) -> None:
    """ADR-0174 Phase 3a multi-actor hazard — when more than one
    proper-noun subject appears in prior context, the pronoun
    resolver must refuse rather than guess the most-recent.
    Verb widening must preserve this defense."""
    from generate.math_candidate_graph import parse_and_solve
    import json
    # Replace <widened_verb> with one this PR adds.
    text = (
        "Alice has 5 marbles. "
        "Bob has 3 marbles. "
        "She <widened_verb> 2 marbles. "
        "How many marbles does Bob have?"
    )
    r = parse_and_solve(text)
    # MUST refuse — wrong attribution is the hazard.
    assert r.answer is None
    # Trace MUST include no_antecedent_ambiguous.
    lookback_events = [
        json.loads(ev) for ev in r.reader_trace
        if json.loads(ev).get("layer") == "lookback"
    ]
    assert any(
        ev.get("outcome") == "no_antecedent_ambiguous"
        for ev in lookback_events
    ), f"multi-actor defense did not fire; trace={lookback_events}"

Reference: memory project-adr-0174-multi-actor-pronoun-hazard, CLAUDE.md §Lookback Review Discipline (the discipline that surfaced this hazard at the right time).

---

Per-PR workflow

VE-A — Acquisition widening

git fetch origin main && \
  git worktree add /tmp/wt-ve-a origin/main && \
  cd /tmp/wt-ve-a && \
  git checkout -b feat/verb-expansion-acquisition

Touchpoints:

• generate/recognizer_match.py:835 — add reviewed lemmas to _ACQUISITION_VERBS frozenset. Each addition has an inline comment citing the hazard review (e.g. # Reviewed 2026-MM-DD: 'gained' grounds as monetary/discrete-count delta; refuses on delta-of-attribute via unit-noun class check upstream.).
• generate/math_roundtrip.py:60 — ensure each added lemma is also in ADD_VERBS so roundtrip_admissible.verb_registered passes.
• New tests in tests/test_adr_0174_phase3_lookback.py::TestVerbExpansionAcquisition covering: (1) each added verb admits a synthetic pronoun-subject sentence through the held-hypothesis path, (2) case 0050 remains refused, (3) at least one real train_sample case lifts from refused to correct.

Acceptance:

• ≥ 1 case lift on train_sample/v1 (verify via uv run python -m evals.gsm8k_math.train_sample.v1.runner)
• wrong=0 preserved across train_sample AND case 0050
• smoke 67/67, packs 141+/141+, lanes 8/8
• All Phase 1–3a tests still pass (113 baseline)

VE-B — Depletion class introduction

git fetch origin main && \
  git worktree add /tmp/wt-ve-b origin/main && \
  cd /tmp/wt-ve-b && \
  git checkout -b feat/verb-expansion-depletion

Touchpoints:

• generate/recognizer_match.py — add _DEPLETION_VERBS: Final[frozenset[str]] near _ACQUISITION_VERBS. Extend _try_extract_discrete_count_anchor's verb dispatch to include elif verb in _DEPLETION_VERBS: anchor_kind = "depletion".
• generate/recognizer_anchor_inject.py — add anchor_kind == "depletion" branch in inject_discrete_count_statement; emits CandidateOperation(kind="subtract") via a new _build_operation_from_discrete_count_depletion helper that mirrors the existing acquisition helper but with kind="subtract".
• generate/math_roundtrip.py — ensure each depletion lemma is in SUBTRACT_VERBS.

Acceptance:

• Synthetic depletion test passes (e.g. "Sam has 10 apples. He gave 3 apples to a friend. How many apples does Sam have?")
• ≥ 1 case lift on train_sample/v1 from the depletion class
• wrong=0 preserved across train_sample AND case 0050
• Comparison-context hazard: "Alice lost 3 more than Bob" must not parse as depletion (already excluded from comparison verbs).

VE-C — Non-arithmetic refusal evidence

git fetch origin main && \
  git worktree add /tmp/wt-ve-c origin/main && \
  cd /tmp/wt-ve-c && \
  git checkout -b feat/verb-expansion-non-arithmetic

Touchpoints (no production behavior change):

• generate/recognizer_match.py — in _try_extract_discrete_count_anchor, when the verb is not in any whitelist, augment the existing return None with a marker on the anchor (something like _DISCARDED_VERB_TRACE: dict[str, list[str]] = {} that captures the verb + sentence for downstream inspection). DO NOT emit candidates; the production path remains a clean refusal.
• New CLI surface: core teaching unrecognized-verbs --since <case-id> that surfaces the captured verbs as candidate evidence for a future ADR-0163.D.x recognizer-category proposal.

Acceptance:

• All 7 non-arithmetic verbs surface in the new CLI output when run against train_sample/v1.
• No score change; no admission behavior change.
• Refusal trace events gain a verb_class_unknown outcome with the verb lemma.

---

Truth tests

Before any VE-A/B/C PR merges, verify:

1. wrong=0 invariant. uv run python scripts/verify_lane_shas.py passes. train_sample/v1 report counts have wrong: 0.
2. Case 0050 canary. Explicit test that 0050 remains refused.
3. Lift evidence. For VE-A and VE-B, at least one train_sample case moves from refused to correct. If zero lift, the PR is premature — the verb widening exposed no real case, indicating the hazard isn't justified. Refuse to ship and re-scope.
4. Phase 3a substrate fires. For VE-A and VE-B, the lifted case must show a lookback trace event with outcome: "admitted" — proving the held-hypothesis path was on the critical path, not bypassed by the regex parser.

---

Out of scope

• Comparison-verb additions (compare_additive / compare_multiplicative). ADR-0123 owns that surface.
• Rate-verb additions. ADR-0122 / ADR-0163.D.2.2 own that surface.
• Transfer-verb additions (give to X, send to X). Already covered by TRANSFER_VERBS; widening that is a separate brief.
• Verb-class proposals that would change the structure of _try_extract_discrete_count_anchor's narrowness chain (e.g. registry-loadable verb sets). That is its own ADR; this brief preserves the hardcoded boundary because that boundary is the wrong=0 safety surface.
• ADR-0174 Phase 3b (compound-clause held hypotheses). Wait until VE-A or VE-B lands so the compound-clause logic has cases to admit.

---

Sequencing recommendation

VE-C (instrumentation)
  ↓
VE-A (acquisition widening) ─┐
                              ├─→ measure: does Phase 3a lookback fire?
VE-B (depletion class)       ─┘    if yes: ADR-0174 Phase 3b
                                   if no:  re-scope non-firing cause

VE-C is independent and can land first to give us better evidence on which Phase 3b targets are worth pursuing. VE-A and VE-B can land in parallel; their case-0050 tests are independent.

---

Reads required FIRST (for the operator running each VE-x)

• generate/recognizer_match.py:812-840 — current verb-set definitions and the design comments justifying each exclusion.
• generate/recognizer_match.py:912-1010 — the discrete-count anchor extractor (the narrowness chain).
• generate/math_roundtrip.py:60-150 — KIND_TO_VERBS, ADD_VERBS, SUBTRACT_VERBS (must add lemmas here too for roundtrip_admissible parity).
• generate/recognizer_anchor_inject.py:186-300 — injector dispatch; depletion needs its own builder.
• docs/handoff/PHASE-3.1-FOLLOWUP-RECOGNIZER-EXPANSION.md — empirical analysis this brief operationalises.
• Memory: feedback-wrong-zero-hazard-case-0050 (case 0050 hazard is mandatory pinning for every PR).
• Memory: thesis-decoding-not-generating (each verb decision is "does this teach the engine to find better?" — non-arithmetic verbs route to VE-C precisely because widening for them would be "storing another found thing").

---

Decision needed before dispatch

• Authorise VE-A acquisition widening? Specifically: which of gain, earn, save, accumulate, acquire (and their inflections) are in scope. Each adds a wrong=0 hazard surface.
• Authorise VE-B depletion class introduction? Specifically: which of donate, give, lose, spend, eat. The depletion class is a new operation-kind path; whether to introduce it AT ALL is a structural decision.
• Authorise VE-C non-arithmetic refusal evidence? No production risk; gives downstream contemplation/ADR work better evidence.

No timelines proposed; this is operator decision input plus dispatch templates. The PRs ship when the operator says go.