core/docs/adr/ADR-0131.G.1-verb-classes-initial-state.md

ADR-0131.G.1 — Capability axis: state-introducing verb classes

Status: Accepted Date: 2026-05-23 Author: CORE agents Parent: ADR-0131.G

---

Context

ADR-0131.G introduced the GSM8K coverage probe to measure the capability of the bounded grammar layers while maintaining the safety rail (wrong == 0). This decision record details the first capability-axis iteration on top of the coverage probe (G.1), which extends the grammar parser to support a closed set of acquisition / action verbs that introduce quantity.

Decision

We recognize that sentences of the form <Entity> <verb> <N> <unit> can introduce a quantity without an explicit "has/have" possession verb. The verbs fall into two classes:

Class A — Pure-possession anchors (initial-possession slot)

Kept in _INITIAL_HAS_RE. These verbs have no semantic overlap with operation verbs and produce no candidate ambiguity:

• had (past possession)
• started / started with (opening state)

Class B — Acquisition/action verbs (operation slot, add-kind)

Handled exclusively as add operations in ADD_VERBS / SUBTRACT_VERBS. The solver defaults the actor's pre-operation state to 0 when no initial possession exists, so a single-statement sentence like "Sam buys 5 apples." resolves correctly as 0 + 5 = 5.

These verbs were not added to _INITIAL_HAS_RE because they also appear in the operation verb registry (math_roundtrip.KIND_TO_VERBS). Adding them to both registries causes branch-disagreement refusals: when a canonical "has" initial for entity E is followed by an acquisition sentence for the same E, the candidate-graph emitter produces two branches—one treating the acquisition as a second initial (wrong answer) and one treating it as an add operation (correct answer)—and the decision rule refuses on disagreement.

Verb	Operation kind	Already in verb registry
buys / bought	add	ADD_VERBS
collected	add	ADD_VERBS
saved / saved up	add	ADD_VERBS
makes / made	add	ADD_VERBS
sells / sold	subtract	SUBTRACT_VERBS

Code Changes

1. _INITIAL_HAS_RE narrowed to pure possession anchors only: (?P<anchor>has|have|had|started)(?:\s+(?:up|with))?

2. CandidateInitial.__post_init__ validation updated to match the narrower set.

3. Optional verb particle added to _op_pattern between verb and value: (?:\s+(?:up|down|out|back|off|in|away))? This allows the operation regex to match "saved up N", "picked up N", etc. without listing particle-bearing forms as initial anchors.

4. ADD_VERBS / SUBTRACT_VERBS in math_roundtrip.py already include all Class B verbs—no changes required there.

Out of Scope

The following capabilities are explicitly deferred to sibling axes:

• Rate-introducing verbs: Multipliers and rates (e.g. "makes $18 an hour") continue to refuse on this axis.
• Comparatives: Multiplicative/additive comparison structures (e.g., "twice as many", "3 more than").
• Acquisition-with-cost: Transactional semantics (buying items at a given price).
• Multi-statement coreference.

Invariants

• wrong == 0: Every evaluation run over both the G1 curated axis and the GSM8K probe must yield zero wrong answers.
• Closed Set: No synonymous expansion or paraphrase tolerance beyond the enumerated verbs.
• Determinism: Evaluator outputs must be byte-equal across consecutive runs.
• No initial/operation overlap: Verbs that appear in ADD_VERBS or SUBTRACT_VERBS must not also appear in _INITIAL_HAS_RE.