Skip to content

Spec v3.2.0: Nested Object Flattening

Choose a tag to compare

View all tags
@blackwell-systems blackwell-systems released this 23 Jun 04:13
v3.2.0
92a8072

Nested Object Flattening (Section 7.4.6)

Encoders may now flatten fixed-shape nested objects into > path column names, reducing tokens by 20-48% on deeply nested API data while maintaining 100% comprehension on every frontier model.

What changed

  • "customer>name" in a tabular header means the name field inside the customer object
  • Multiple levels chain: "billing>address>city" produces 3-level nesting
  • Decoder reconstructs nested objects from > path columns
  • Alternative to inline schema (Section 7.4.5.1); both remain valid
  • Round-trip guarantee preserved: decode(encode(value)) == value

Validation

  • 40+ comprehension eval runs across 20 models, 8 providers (Anthropic, OpenAI, Google, xAI, Moonshot, DeepSeek, Mistral, Meta, Amazon)
  • 100% comprehension on every frontier model (Claude, GPT-5.5, Gemini, Grok)
  • Generation validity: 0% (current GCF) to 96-100% (flat GCF) on Mistral Medium
  • Budget experiment: GCF fits 1.9x more data than JSON in the same token budget
  • 50M+ fuzz round-trips across JSON, YAML, MessagePack, CSV
  • 17-format mega-gauntlet (JSON, XML, MessagePack, YAML, BSON, TOML, CBOR, Protobuf, CSV, JSON5, Avro, Arrow, Parquet, Pickle, INI, NDJSON, Plist)

Conformance

  • 15 new fixtures in tests/conformance/flatten/
  • Updated existing fixtures for flattening-default encoder output

SDK implementations (on flatten-v3.2 branches, publishing shortly)

  • Go (gcf-go)
  • Rust (gcf)
  • TypeScript (@blackwell-systems/gcf)
  • Python (gcf-python)
  • Swift (gcf-swift)
  • Kotlin (gcf-kotlin)
Assets 2
Loading