core/generate/derivation/accumulate.py
Shay 9dcb00248c feat(derivation): route accumulation through ADR-0184 S2 semantic ledger
ADR-0184 S1 (state/bind.py, state/change.py) landed the reusable referent and
change-cue helpers. S2 adds the first explicit state-transition substrate and
routes both accumulation readings through it:

- state/model.py   frozen SET/GAIN/LOSS transition model over one entity/unit
                   StateKey, with structural validation (closed op set; unit is a
                   str with "" = unitless, mirroring the extractor's Quantity.unit
                   contract; entity optional, mirroring leading_subject_token).
- state/ledger.py  build_accumulation_ledger() — the proven single-referent
                   gain/loss reading, expressed as a SemanticLedger.
- state/replay.py  replay_accumulation_ledger() — the only bridge to
                   GroundedDerivation; refuses any non-accumulation ledger shape
                   (non-SET start, cross-key mutation).

accumulate.py's _build_accumulation / _build_accumulation_anchor_skip now build a
ledger and replay it instead of constructing the arithmetic chain inline; the
public compose_accumulation()/accumulation_candidates() surfaces are unchanged.

Behavior-equivalent: a byte-for-byte differential over 937 real GSM8K problems
(accumulation_candidates, compose_accumulation, pooled_candidates, resolve_pooled)
shows 0 differences vs main. Semantic worlds never commit directly — replay emits
GroundedDerivation and verify.py / pool.py remain the sole authority. Sealed: no
serving/runtime/chat import; no CLAIMS or eval-lane-SHA movement.

Refs ADR-0184 section 7 (S2).
2026-06-10 12:00:23 -07:00

156 lines
7.9 KiB
Python

"""ADR-0178 GB-3b.1 — single-referent accumulation chaining.
The first cross-clause *comprehension* reading: one actor's quantity changes over
successive clauses (``Sam has 14 apples. He buys 9 more.`` -> ``14 + 9``). It is
the safe specialisation of the cross-clause sum that GB-3a's referent guard
correctly refuses wholesale (the ``Alice has 6 … Tom has 2 …`` hazard): we chain
across clauses **only** when (a) the later clause stays on the **same referent**
and (b) it carries a **licensed change cue** whose polarity is unambiguous.
Otherwise we refuse — the guard is generalised, never weakened.
Reading:
1. **Anchor** — clause 1 must establish exactly one quantity ``(actor, N, unit)``.
2. **Change steps** — each later quantity-bearing clause applies ``+ M`` (gain) or
``- M`` (loss) to the running total, where ``M`` is the clause's single grounded
quantity, taken **in the anchor's unit** (``9 more`` = 9 more *apples*; the unit
is inherited from the running total, which is what accumulation means).
3. **Gate** — the constructed chain runs through the unchanged self-verification
gate (grounding ∧ unit ∧ completeness ∧ uniqueness). The gate keeps
wrong=0; this only proposes a structurally-licensed candidate.
ADR-0184 S1 extracted the reusable referent and change-cue helpers into
``generate.derivation.state``. S2 routes both accumulation readings through a
minimal semantic ledger (SET/GAIN/LOSS over one entity/unit key) and replays it to
``GroundedDerivation`` — the transition construction now lives in ``state.ledger``
and ``state.replay``, not as inline arithmetic here. This module remains the public
accumulation composer surface; behavior is intentionally unchanged, and the replayed
``GroundedDerivation`` still faces the unchanged verifier/pool.
Sealed (no ``chat/`` import); deterministic; refuse-preferring.
"""
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
from generate.derivation.state.ledger import build_accumulation_ledger
from generate.derivation.state.replay import replay_accumulation_ledger
from generate.derivation.verify import Resolution, select_self_verified
def _quantity_clauses(problem_text: str) -> list[str]:
"""Sentence-level clauses that carry extracted quantities (the ledger anchor +
change-clause sequence)."""
return [c for c in segment_clauses(problem_text) if extract_quantities(c)]
def _build_accumulation(
problem_text: str, *, drop_isolated_foreign: bool
) -> GroundedDerivation | None:
"""Construct the single-referent accumulation chain (ungated).
``drop_isolated_foreign`` (ADR-0182): when a change clause carries more than
one quantity, drop those with a **non-empty unit foreign to the anchor's unit**
(a candidate distractor — ``studies for 3 hours`` among ``pencils``) and proceed
if exactly one same-unit/unitless change remains. With the flag off this is the
strict GB-3b.1 reading (a multi-quantity change clause refuses), so
:func:`compose_accumulation` is byte-identical to its pre-ADR-0182 behavior.
The distractor-skip reading is **never committed alone** — it only ever enters
the pool to force a disagreement refusal (see :mod:`generate.derivation.pool`).
ADR-0184 S2: the SET/GAIN/LOSS reading is built as a semantic ledger and replayed
to ``GroundedDerivation``; the construction is unchanged in behavior.
"""
ledger = build_accumulation_ledger(
_quantity_clauses(problem_text), drop_isolated_foreign=drop_isolated_foreign
)
if ledger is None:
return None
return replay_accumulation_ledger(ledger)
# ADR-0182 anchor-skip: sub-clause split on conjunctions. A single sentence can pack
# a state and its change ("Tom has 8 tickets and buys 4 more tickets") — the
# sentence-level segmenter (used everywhere; not changed) keeps them together. This
# finer split is *local* to the ungated candidate generator, so it cannot move
# GB-1/GB-2/serving/practice (which never call it). Lexeme-level (ADR-0165): it names
# coordinating conjunctions, it does not parse grammar.
_CONJUNCTION_SPLIT: Final[re.Pattern[str]] = re.compile(r",?\s+(?:and then|and|then)\s+")
def _sub_clauses(problem_text: str) -> list[str]:
"""Sentence clauses, each further split on coordinating conjunctions."""
parts: list[str] = []
for clause in segment_clauses(problem_text):
parts.extend(p.strip() for p in _CONJUNCTION_SPLIT.split(clause) if p.strip())
return parts
def _build_accumulation_anchor_skip(problem_text: str) -> GroundedDerivation | None:
"""ADR-0182 — accumulation over sub-clauses, skipping a leading all-foreign block.
Reads ``A train travels 60 mph for 2 hours. Tom has 8 tickets and buys 4 more
tickets.`` by skipping the (anchor-position) train block — its quantities cannot
seed an anchor (≠1 quantity) — and anchoring on the first single-quantity
sub-clause (``Tom has 8 tickets``), then chaining its conjunction-mate change
(``buys 4 more`` → +4). The skipped block's quantities go unused; the pool's
isolated-foreign exemption then classifies the reading ``exempt`` (commit-
ineligible), so it can only force a disagreement refusal, never commit. Ungated.
"""
sub_clauses = [(s, extract_quantities(s)) for s in _sub_clauses(problem_text)]
quantity_subs = [(s, qs) for s, qs in sub_clauses if qs]
if len(quantity_subs) < 2:
return None
# Anchor = first single-quantity sub-clause; leading non-anchorable (≠1
# quantity) sub-clauses are skipped (candidate distractor blocks). The anchor and
# its trailing change sub-clauses are then read by the same semantic ledger as the
# sentence-level path (ADR-0184 S2): same referent guard, same one-change-per-clause
# rule, same polarity gate. drop_isolated_foreign stays off here — the anchor-skip
# reading drops a whole leading block, not an in-clause foreign quantity.
anchor_idx = next((i for i, (_, qs) in enumerate(quantity_subs) if len(qs) == 1), None)
if anchor_idx is None:
return None
selected = [sub for sub, _ in quantity_subs[anchor_idx:]]
ledger = build_accumulation_ledger(selected, drop_isolated_foreign=False)
if ledger is None:
return None
return replay_accumulation_ledger(ledger)
def compose_accumulation(problem_text: str) -> Resolution | None:
"""GB-3b.1 composer — single-referent gain/loss accumulation. Refuse-preferring.
The strict (commit) reading: it gates the no-distractor-skip derivation through
the unchanged self-verification gate. Behavior is byte-identical to pre-ADR-0182.
"""
derivation = _build_accumulation(problem_text, drop_isolated_foreign=False)
if derivation is None:
return None
return select_self_verified([derivation], problem_text, target_units=())
def accumulation_candidates(problem_text: str) -> tuple[GroundedDerivation, ...]:
"""ADR-0182 — the ungated accumulation readings for cross-composer pooling.
Three readings: the strict GB-3b.1 reading, the distractor-skip reading (drops an
isolated-foreign quantity in a multi-quantity change clause — 0014), and the
anchor-skip reading (skips a leading all-foreign block + reads a conjunction-mate
intra-sentence change — 0016). Ungated: the pool classifies each (``complete``
commits, ``exempt`` refuses-only) and the disagreement rule does the wrong=0 work.
Deterministic; de-dup is the pool's job.
"""
candidates: list[GroundedDerivation] = []
for drop in (False, True):
derivation = _build_accumulation(problem_text, drop_isolated_foreign=drop)
if derivation is not None:
candidates.append(derivation)
anchor_skip = _build_accumulation_anchor_skip(problem_text)
if anchor_skip is not None:
candidates.append(anchor_skip)
return tuple(candidates)