ADR-0131.G.3 — Capability Axis: Numeric Literals (Money, Fractions, Compound Numbers)

Status

Accepted

Date

2026-05-23

Context

As part of the ADR-0131 re-benchmarking roadmap, the GSM8K coverage probe identifies capability axes along which the NL-to-typed-graph parser must be extended. This decision specifies the capability axis G.3: support for money, fractions, and compound numeric literals.

Historically, the parser was restricted to simple integer digits and single-word numbers (1 to 12). Real-world math problems introduce currency notation, fractional quantities (both digit slash forms and spelled equivalents), adjectives qualifying units, and hyphenated numeric nouns.

Decision

Extend the candidate-graph parser layer to consume the existing semantic packs en_units_v1 (ADR-0127) and en_numerics_v1 (ADR-0128) to recognize these literal shapes without hardcoding inline regex alternations.

1. Closed Literal Classes

We implement recognition for four distinct literal classes:

Money: Recognizes currency symbols and currency word forms. Leading symbols map to canonical plural units via en_units_v1.
Fractions: Recognizes digit slash forms (e.g., 3/4) and spelled equivalents in the pack (e.g., one-half, three-quarters). Fractions are converted to floats in Quantity structures. Prepositional of <substance> tails (e.g. 3/4 of a cake) are parsed, using the substance as the unit where explicit unit tokens are absent.
Word-number compositions: Recognizes combinations of spelling numbers, adjectives, and head nouns (e.g., five full boxes of crayons). The adjective is correctly retained as part of the unit (full boxes of crayons ≢ boxes).
Hyphenated compound numerics: Recognizes compound adjectives modifying a unit head noun (e.g., 10 one-hour videos).

2. Currency-Symbol → Unit Mapping

Leading currency symbols resolve deterministically to plural units in the unit pack:

$ → "dollars"
¢ → "cents"
€ → "euros"
£ → "pounds"
¥ → "yens"
₱ → "pesos"

Currency symbols compose with rate denominators (e.g., $ and hour → "dollars per hour"; $0.75 each → "dollars per item").

3. Scope Pinning and Refusal Probes

To preserve the wrong == 0 firewall, inputs that exceed well-defined boundaries are cleanly refused:

Decimals > 2 in currency: $18.0000 is refused at the parser layer.
Division by zero: Fractions like 1/0 fail solver/graph construction and refuse.
Ambiguous hyphenations: Expressions like one-hour-old refuse.
Out of scope notation: Percentages (10%) and scientific notation (1e3) are excluded and refuse.

4. Deferred Shapes

The following shapes remain out of scope for G.3 and will be addressed in future capability iterations:

Distributive operators: Distributive scopes (like "each saved up $40" or "each weighing 5 ounces") are deferred to G.4.
Unit inheritance: Re-using the last referenced unit (e.g., "gives 1/4 of that") requires state threading and is deferred to P3 grammar integration.
Multi-clause sentence syntax: Compound sentences linked by conjunctions (like but then lost 12 or and her friend has) are refused.

Verification Plan

Curated Benchmarks: Author 25 isolated capability cases under evals/math_capability_axes/G3_numerics/v1/cases.jsonl (5 per literal class, 5 refusal probes).
Runner Verification: The runner at evals/math_capability_axes/G3_numerics/v1/runner.py must score correct_rate == 100% against expectations and wrong == 0.
Determinism: The runner output report.json must be byte-equal across consecutive runs.
Regression Testing: All test suites must remain green (pytest tests/test_adr_0131_G3_numerics.py and tests/test_adr_0131_*.py).

Consequences

The G3 capability axis is fully verified and landed under the safety rail.
Parsing of money, fractions, and compound noun phrases is now fully pack-driven and cached at start-up.
The baseline GSM8K train-sample coverage probe shows that targeted first-sentence refusals (like Tina makes $18.00 an hour) now successfully parse, moving the refusal frontier further down the problem structure.

4.2 KiB Raw Blame History