The digit-free grammar is a character whitelist¶
Decided in 0.3.0 (slots.PROSE_CHARS). Recorded in 0.12.0.
Context and problem statement¶
The digit-free grammar (ADR-0002) reaches the model as a
JSON-schema pattern that the model server compiles into a sampling grammar. The obvious pattern,
^[^0-9]*$, was probed on llama.cpp's json-schema-to-grammar converter and failed: the pattern is
compiled outside JSON string escaping, so the negated class admitted raw control bytes, and the
model emitted replies that json.loads could not parse.
Decision drivers¶
- Nothing the grammar admits may need JSON escaping.
- The pattern must compile on the converter's regex subset.
- A length cap must hold under the grammar, not only in validation afterwards.
Considered options¶
- A negated class:
[^0-9]. - A negated class that also excludes control characters and quotes.
- A whitelist of the characters prose needs: letters, space and
,.;:'()/&?!-.
Decision outcome¶
Chosen option 3, PROSE_CHARS = [A-Za-z ,.;:'()/&?!-]. Constraints of the converter shaped the
rest:
maxLengthis not enforced alongsidepattern, so a capped run carries its cap inside the pattern as{0,N}. A slot counts as one repetition, so the grammar's cap is approximate; the pydantic field keeps the exactmax_length.- The regex subset has no
(?:, so groups are plain(...). - The complement used by
to_proseis derived fromPROSE_CHARS, so the two cannot drift.
Consequences¶
- Good: a grammar-admitted reply always parses as JSON.
- Bad: prose loses characters such as
",%,$and non-ASCII letters; figures and symbols come from rendered slots. - Bad: a stricter or richer alphabet is a change to every consumer's grammar, so
PROSE_CHARSis part of the public contract.