84db129629
GSM8K labels containers/regions with a trailing single-letter or short-numeric
label ('Jar A has 28 marbles', 'Section G has 10 cars', 'District 2 has 19
voters'); the initial-possession entity slot captured only 'Jar' and the label
broke the match. Adds a separate sibling pattern _INITIAL_HAS_LABELED_RE
(mirroring ADR-0136.S.4 localisation) that REQUIRES the label, so the global
_ENTITY is unchanged and bare subjects yield no duplicate candidate.
- Composes with ADR-0193 aggregate question: 'Jar A has 28 marbles. Jar B has
12 marbles. How many marbles are there in total?' -> 40.0.
- 0 real-corpus metric flip (honest substrate): the one real multi-container
aggregate additionally needs comparative + multiplicative + lowercase-ref.
- wrong=0 HOLDS full corpus (7,473 q); train_sample byte-identical 4/46/0;
synthetic-registry capability-axis gate + G5 lane green; smoke 67 passed.
- Label bounded by the possession verb: multi-word nouns ('Jar Apple') do NOT
match. wrong=0 held downstream by completeness + round-trip + disagreement.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
84 lines
4.1 KiB
Markdown
84 lines
4.1 KiB
Markdown
# ADR-0194 — Labeled-container subject entity shape
|
|
|
|
**Status:** Proposed (implemented in this PR).
|
|
**Extends:** [ADR-0136.S.4](./ADR-0136.S.4-novel-initial-form.md) (sibling-pattern
|
|
localisation), [ADR-0123a](./ADR-0123a-inference-shape-synonym.md) (entity slot).
|
|
**Composes with:** [ADR-0193](./ADR-0193-aggregate-existential-question-frame.md)
|
|
(aggregate question frame).
|
|
**Substrate: 0 real-corpus metric flip by design; the value is the entity-shape
|
|
generalisation + proven composition with the aggregate question.**
|
|
|
|
> **One line.** GSM8K labels containers/regions with a trailing single-letter
|
|
> or short-numeric label ("Jar A has 28 marbles", "Section G has 10 cars",
|
|
> "District 2 has 19 voters"). The initial-possession entity slot
|
|
> (`_ENTITY = (?:[A-Z]\w+|[Tt]he\s+\w+)`) captures only "Jar" and then expects
|
|
> the possession verb, so the label breaks the match and the statement parses
|
|
> to nothing. This adds a separate sibling pattern that REQUIRES a label.
|
|
|
|
---
|
|
|
|
## 1. The gap
|
|
|
|
Both reader paths reject the labeled subject:
|
|
- the candidate parser's `_INITIAL_HAS_RE` (`extract_initial_candidates` → 0
|
|
candidates);
|
|
- the recognizer's discrete_count anchor (proper-noun single-token subject →
|
|
0 anchors).
|
|
|
|
"Jamie has 28 marbles" parses (1 candidate); "Jar A has 28 marbles" does not —
|
|
purely because of the trailing label.
|
|
|
|
## 2. Decision
|
|
|
|
Add `_INITIAL_HAS_LABELED_RE` in `generate/math_candidate_parser.py`, consumed
|
|
by a dedicated `_init_has_labeled_candidates` helper wired into
|
|
`extract_initial_candidates`:
|
|
|
|
```
|
|
^<Noun> <label> (has|have|had|started) <value> [adj] [unit] [of/in/for/with …]
|
|
label = a single uppercase letter OR 1-2 digits
|
|
```
|
|
|
|
- **Sibling-pattern localisation** (mirrors ADR-0136.S.4's
|
|
`_INITIAL_HAS_INDEF_RE`): the global `_ENTITY` is unchanged for every other
|
|
path (operations, comparisons, questions).
|
|
- **The label is required**, so a bare subject ("Jamie has 28 marbles") never
|
|
reaches this pattern and yields no duplicate candidate.
|
|
- **The label is bounded by the following possession verb**, so a multi-word
|
|
noun does NOT match: "Jar Apple has 5 marbles" → no candidate ("Apple" is not
|
|
a single-letter label), "Box Set has 12 items" → no candidate.
|
|
- Same value/unit tail and money normalisation as `_INITIAL_HAS_RE`.
|
|
|
|
### Why this is safe (the firewall is the precondition)
|
|
|
|
The label widening only makes a statement *parse* into an initial possession.
|
|
wrong=0 is held downstream by the completeness guard (ADR-0191) + the
|
|
round-trip filter + branch disagreement — a mis-parse leaves source quantities
|
|
uncovered and refuses. Full-corpus verification: **wrong=0 HOLDS** (7,473 q);
|
|
`train_sample` byte-identical **4/46/0**; the synthetic-registry capability-axis
|
|
`wrong=0` gate and the G5 aggregate lane both stay green.
|
|
|
|
## 3. Evidence
|
|
|
|
- **Composes with ADR-0193:** "Jar A has 28 marbles. Jar B has 12 marbles. How
|
|
many marbles are there in total?" → **40.0**; three-container variant → 10.0.
|
|
- **0 real-corpus flip** (honest framing): of the 3 real container-subject
|
|
problems under an aggregate question, the only multi-container aggregate
|
|
("Jar A has 28 marbles. Jar B has 12 more marbles than jar A. Jar C has twice
|
|
as many as jar B. … altogether?") additionally requires **comparative-additive
|
|
+ comparative-multiplicative + lowercase-reference** resolution. This is the
|
|
composition-wall lesson again: an entity-shape widening is necessary, not
|
|
sufficient.
|
|
- **Tests:** `tests/test_labeled_container_subject.py` (labeled containers
|
|
parse; bare subjects yield no duplicate; multi-word nouns don't match a
|
|
label; composes with the aggregate question).
|
|
|
|
## 4. Consequences & follow-ups
|
|
|
|
- The labeled-container entity shape is banked; it becomes load-bearing the
|
|
moment **comparative reading** (the actual aggregate blocker — ADR-0131.G.2 /
|
|
`core-comparatives`) can compose `compare_additive`/`compare_multiplicative`
|
|
ops into an aggregate answer, and the **lowercase-reference** ("jar A" inside
|
|
a later clause) resolves to the labeled entity.
|
|
- Compound enumeration and the recognizer-path labeled subject remain closed;
|
|
open them only if a serving need is proven under the firewall.
|