GCF Specification v3.0.0: Inline Schema Encoding

Inline Schema Encoding

The single largest token efficiency improvement since GCF's initial release. Nested objects with 3+ scalar fields are now encoded positionally with a one-time schema declaration.

Before (v2):

## orders [2]{id,total,customer}
@0 1|99.99|^
  .customer {}
    name=Alice
    email=alice@test.com
    tier=premium

After (v3):

## orders [2]{id,total,customer}
@0 1|99.99|^{name,email,tier}
Alice|alice@test.com|premium
@1 2|49.99|^
Bob|bob@test.com|standard

Breaking Changes

• Attachment lines no longer indented (same depth as parent row)
• Inline object fields use positional encoding without field-name prefix
• Objects with 3+ scalar fields encoded with ^{fields} inline schema
• No backwards compatibility with v2 indented attachments

New Features

• Inline object schema: first row declares field order with ^{fields}, subsequent rows use bare ^
• Shared array schemas: identical nested arrays omit {fields} after first row
• MinInlineFields=3: threshold prevents overhead on small objects
• 15 new conformance fixtures, 156 total

Numbers

Metric	Value
Comprehension (standard workloads)	100% on every frontier model
Comprehension (code graphs)	90.7% (vs TOON 68.5%, JSON 53.6%)
Token efficiency (15 datasets)	-25.5% vs TOON, -53% vs JSON
Fuzz iterations	1.006 billion, zero failures
Scale test (1000 orders)	GCF fits (47K), JSON impossible (161K)
LLM evaluations	1,700+ across 10+ models, 3 providers
Implementations	6 languages, all passing

Migration

encodeGeneric / encode_generic now produces v3 output. Decoders accept v3 only (no v2 indented attachment fallback).

Libraries: Go (v1.1.0), TypeScript (v2.0.0), Python (v2.0.0), Kotlin (v2.0.0), Rust (v2.0.0), Swift (v2.0.0)