core/docs/decisions/ADR-0178-GB3b-referent-accumulation-scope.md

ADR-0178 GB-3b — referent-aware accumulation chaining (scope)

Status: Proposed (scope only — no code). Sub-phase of ADR-0178 Gap B. Builds on the GB-3a referent guard (PR #456) and the CP-2a measurement (PR #461).

One-line thesis. The first cross-clause reading that produces flips in a chunk and is genuinely "comprehension": a single actor's quantity changes over successive clauses (X has N. He buys M more. → N + M). The win is gated by a tight referent + change-cue model so the Alice/Bob hazard still refuses.

---

1. The measured opportunity (why this is the next move, not a guess)

Run on the 200 sealed cases (50 train_sample + 150 ADR-0163-F additive practice), deterministic:

lane	single product/sum == gold	needs multi-step
train_sample	2 / 50	48
practice-additive	46 / 150 (sum-of-all == gold)	—

All 46 additive cases are the same shape and all currently refuse (practice lane solves 0/150). Inspected, every one is single-referent accumulation:

Sam has 14 apples. He buys 9 more. How many apples does Sam have now?   -> 14 + 9 = 23
Lisa has 30 coins. She earns 15 more by doing chores. How many...?      -> 30 + 15 = 45
Kate has 18 pencils. Her teacher gives her 5 more. How many...?         -> 18 + 5 = 23

The CP-2a measurement (PR #461) showed no cue is reliable because the blunt search's readings are crude. GB-3b is the better candidate generation that measurement said was the real lever — and the 46 cases are the concrete, gold- verified chunk it should flip. The subtract analog (has N, gives/loses/uses M) is the same shape and adds a comparable second chunk.

2. Why they refuse now — and why that is correct

GB-3a made compose_sequential refuse all multi-clause same-unit aggregation, to kill the referent hazards H1/H2/H3 (Alice has 6 apples. Tom has 2 apples. must not sum to Alice's total). These 46 cases are multi-clause, so GB-3a refuses them — correctly, given what it can currently see.

The difference between a safe accumulation and the hazard is exactly two readable things:

	clause 2	safe?
accumulation	He buys 9 more — same actor (pronoun), change verb	✅ +9
hazard (H1)	Tom has 2 apples — new actor, no change verb	❌ refuse

GB-3b adds precisely this discrimination: chain across clauses iff (same referent) ∧ (a licensed change cue); otherwise refuse. It generalises GB-3a (refuse-all-multi-clause → chain-when-licensed-else-refuse), it does not weaken it.

3. The reading GB-3b adds

1. Anchor. The running total is clause 1's resolved GB-1 leaf: (actor, N, unit).
2. Change steps. Each later clause that (a) is same-referent and (b) carries a change cue applies ± M to the running total, where M is that clause's grounded quantity. Op + polarity come from the cue lexeme; the operand from the clause's text quantity; the order from clause order.
3. Gate. The constructed chain runs through the unchanged self-verification gate (grounding ∧ cue ∧ unit ∧ completeness ∧ uniqueness). The gate already keeps wrong=0; GB-3b only proposes a structurally licensed candidate for it to verify.

Closed change-cue lexeme sets (ADR-0165-safe — lexemes, not grammar templates; refined by the CP ledger, not asserted complete):

• GAIN (+): buys, bought, gets, gments? → gets, got, receives, received, finds, found, earns, earned, adds, added, gains, gained, makes, made, picks, more
• LOSS (−): gives, gave, loses, lost, spends, spent, uses, used, eats, ate, sells, sold, drops, dropped, removes, removed, fewer, less

(Final sets curated against the corpus; an unknown change verb → refuse, never guess.)

4. Referent model (wrong=0-critical) — cross-reference + decision

Prior art (must not reinvent): ADR-0164.2 (pronoun-entity-resolution), ADR-0164.3 (cross-sentence-state), and ADR-0174 (PronounResolution in generate/math_candidate_graph.py, with the multi-actor pronoun hazard — gender-blind most-recent-antecedent — recorded in memory project-adr-0174-multi-actor-pronoun-hazard). That comprehension subsystem is heavyweight, did not move GSM8K scoring, and was partly retired (generate/ comprehension/lifecycle.py is dual-use/inert in scoring per project-lifecycle- reader-dual-use).

Decision: GB-3b builds a minimal, refuse-preferring referent guard in the derivation lane — it does not resurrect the retired reader and does not reuse the gender-blind antecedent resolver (itself a hazard). Rationale: the clean derivation lane is what has been making progress; the old resolver's failure mode is exactly the multi-actor confusion we must refuse on.

The guard (lexeme/orthographic-anchored):

• Clause 1's subject = its leading proper noun (a capitalised token that is not sentence-initial-only) or pronoun.
• A later clause is same-referent iff it introduces no new proper-noun actor — i.e. it continues via a pronoun (He/She/They/Her/His) or states no new named subject.
• If a later clause introduces a different proper-noun actor → multi-actor → refuse (defensive; this is the ADR-0174 hazard's required fix, finally built).

This is deliberately conservative: it refuses on any referent ambiguity rather than resolving it. Pronoun gender/number matching is not attempted (that was the old resolver's trap); "a new name appeared" is the only signal, and it triggers refusal, not resolution.

5. wrong=0 obligations (must be proven, not asserted)

A test must fail if any of these is violated (CLAUDE.md proof-obligation rule):

1. Grounded + cue-licensed + unit-consistent change steps (the existing gate).
2. Multi-actor → refuse. H1 (Alice … Tom …) stays None. A new proper-noun subject in any clause refuses the whole chain.
3. No change cue → refuse. Two clauses with quantities but no gain/loss lexeme do not get summed (no implicit aggregation).
4. Ambiguous polarity → refuse. A clause carrying both a gain and a loss lexeme, or an unknown verb where a change is implied, refuses.
5. Comparative bound elsewhere (H2) → refuse; question-temporal-scope (H3, "how many before giving any away") → refuse (handling "before/after" question scoping is deferred to a later increment; until then it must refuse, never apply the change blindly).
6. Determinism, seal (no chat/ import), serving 3/47/0 byte-identical (sealed substrate until a Phase-5 ratification, as with GB-3a/CP).

6. Cue-precision coupling (this is what closes the CP-2a loop)

CP-2a measured every existing cue at ~0.0 reliability. The change cues GB-3b introduces are structurally different: buys always adds, gives always subtracts — they should record high reliability in the CP ledger once GB-3b's candidates are trained through CP-2a. GB-3b is the phase that finally produces a reliable cue pattern, which is the prerequisite CP-2b (trust) was blocked on. The two tracks compound here: structure produces clean readings → cue-precision finds signal → trust becomes possible.

7. Increments (one PR each, sealed, measured gold-checked)

• GB-3b.1 — single-referent gain/loss accumulation. The anchor + one-or-more change steps + the minimal referent guard (incl. pronoun-continuation, since the 46 use He/She). Flips the 46 additive + cases and the subtract analog. This is the chunk.
• GB-3b.2 — multi-step / mixed accumulation. 3+ changes, gain and loss in one problem; running total across several clauses.
• GB-3b.3 — question-temporal scope. "before/after he gave any away" — refuse- aware reading of the question's time reference (closes H3 properly).

8. Honest flip expectation (so we are not surprised)

• Big chunk in practice-additive: GB-3b.1 should take the additive lane from 0 correct toward ~46 (the + cases) and, with the subtract analog, a comparable additional set — the first time the comprehension reading visibly works at scale. This is the progress signal, measured and gold-checked.
• Modest in train_sample: real GSM8K skews to multi-step mixed-operation problems (48/50 need >1 op); single-referent accumulation will flip only the few that are pure accumulation. GB-3b is the foundational primitive the multi-step work composes on, not a one-shot GSM8K solve.
• Serving stays 3/47/0 (sealed). Progress shows as practice correct rising with wrong held at 0 (or eliminations only) — the two-regime contract.

9. Acceptance (Proposed → Accepted)

1. GB-3b.1 lands; the 46 + cases (and the subtract analog) flip to correct in the practice lane under self-verification; wrong does not rise; serving frozen.
2. The §5 obligations each have a test that fails under the violation (multi-actor admitted, unlicensed/absent change summed, wrong polarity, H1/H2/H3).
3. Change cues record measurable reliability in the CP ledger (CP-2a report), demonstrating the structure→precision coupling of §6.
4. Determinism/seal invariants hold; the GB-3a hazards H1/H2/H3 remain refused.