#title: ASON description: The native LLM-first format: canonical syntax, columnar records, dictionaries, and retrievable references.

#ASON

ASON is the LLM-first structured format designed and implemented natively by ash. It reduces repeated structure in model context while preserving strict schemas, canonical encoding, framed boundaries, and recoverable evidence.

#Example

s:0
a:search
d{p}:
0:"src/runtime.rs"
r{p,l,c}:
0,42,governor
0,88,stable_merge
z{shown,total}:
2,17
q:ref-7

• s:0 is a scalar field.
• d{p} is a path dictionary column.
• r{p,l,c} contains homogeneous search rows with one column header.
• z{shown,total} carries projection metadata.
• q is a complete-evidence reference available for slice or search.

#Canonical rules

1. Message schemas fix field and column order.
2. Safe text uses bare atoms; UTF-8 strings use the shortest required escapes.
3. Duplicate fields, duplicate columns, unknown required structures, and noncanonical numbers fail.
4. The parser enforces document, collection, string, and depth limits before full model construction.
5. Decode followed by encode must yield one unique byte sequence.

Bare atoms additionally reserve # ? | >; together with / - they form ASH's single-byte formula operators. Other punctuation remains quoted.

#Token evidence

The fixed format corpus produces the same ratio under pinned cl100k_base and o200k_base tokenizers:

This is a format-corpus regression datum, not a full agent-task quality or runtime performance claim. See benchmark evidence to reproduce it.

#Why not ordinary JSON

JSON repeats keys for every homogeneous record and tends to inline long paths and large values. ASON moves structural repetition into schemas and column headers, then leaves large values in bounded storage while returning references and useful projections. It is not a compatibility layer for another format.