TRANSMISSION ACTIVE

stele

structured data format optimized for LLM consumption

The Problem

Every API call has a cost. Every token in that call adds to it.

JSON burns tokens on ceremony:

{"users":[{"id":1,"name":"alice","role":"admin"},{"id":2,"name":"bob","role":"user"}]}

Quotes. Braces. Colons. Repeated keys. Every row repeats the same field names.

The Solution

stele encodes once, streams values:

Schema declared once. Data rows are pure values. 30-50% fewer tokens.

Design Philosophy

stele is built on one principle: minimize tokens while maximizing model comprehension.

Goal	How
Token efficiency	Eliminate JSON’s syntactic overhead
Model parseability	Structure that LLMs extract accurately without examples
Schema compression	Declare field names once, reference by position

Human readability is a secondary benefit, useful for debugging. But make no mistake: stele exists because every token costs money, and JSON burns tokens on ceremony.

Quick Comparison

Format	Haiku Accuracy	Tokens (50 records)
JSON	baseline	6,757
TOON	59.8%	8,744 (+29%)
stele	100%	5,918 (-12%)

stele parses at parity with JSON while being smaller. Smaller models handle it cold.

## Read the Full Specification ᚠ ENTER THE TABLETS ᚣ

Implementation

The reference implementation is base-d, a Rust CLI and library.

# JSON to stele
echo '{"users":[{"id":1,"name":"alice"}]}' | base-d stele

# stele to JSON
echo '@users|id^i|name^s*1|alice' | base-d stele -d

Format	Model Reads Structure	Compression	Use Case
stele	Yes	30-50%	Working data
carrier98	No	90-97%	Shuttle data