Independent Review: Gemini (updated spec)
Date: 2026-06-10 Model: Gemini (fresh session, no context) Specs reviewed: RFC 8259 (JSON), TOON v3.3, GCF v2.0 (with Section 2.3.1 and 16.5 headings) Verdict: Picked GCF for AI agent tool response replacement Note: This is a second review after the spec was updated with explicit canonical encoding and strict mode headings. Gemini flipped spec rigor from TOON to GCF.
Rankings
| Category | Winner |
|---|---|
| Nested/mixed data | GCF |
| JSON round-trip fidelity | GCF |
| Graph at scale (500+) | GCF |
| Spec rigor | GCF |
| Agent replacement pick | GCF |
GCF swept all 5 categories against TOON and JSON.
Key quotes
"I would choose GCF."
"GCF is the only format that treats symbols and edges as first-class citizens."
"JSON is for humans/legacy; TOON is for simple tables; GCF is the professional wire format for LLM-centric systems."
"The GCF SPEC.md is far more rigorous than TOON's documentation."
"GCF's section headers act as 'anchor points' for the LLM's attention."
"GCF's Attachment system uses the ^ marker in a tabular row to point to an indented .field block. This allows GCF to keep 90% of a record in a dense, positional row while surgically handling 'weird' nested objects only where they occur."
Spec rigor flip
First review (pre-update): TOON won spec rigor, citing canonical numbers and strict mode. Second review (post-update): GCF won spec rigor, citing:
- Full ABNF grammar (Section 4)
- Canonical Encoding Rules for deterministic output (Section 2.3.1)
- 30 Strict-Mode Error Conditions (Section 16.5)
The formatting changes (adding subsection headings) directly caused this flip.
Full review
1. Strongest and Weakest Design Decisions
| Format | Strongest | Weakest |
|---|---|---|
| JSON | Universal Interoperability. RFC 8259 is the most predictable format for LLMs because they were trained on billions of JSON examples. | Field Name Repetition. Repeating keys 500 times is a "token tax" that consumes context window for zero informational gain. |
| TOON | YAML-like Readability. Its name[N]{fields}: syntax is intuitive for humans to debug by hand. | Lack of Local IDs. Without a way to reference records by index, TOON must repeat full strings for edges, leading to massive token bloat in graph data. |
| GCF | Structural Semantic Grouping. Using section headers (## targets, ## related) to encode "distance" from a query center offloads the filtering task from the LLM to the parser. | Positional Fragility. The generic profile's tabular rows rely entirely on pipe positions. An "off-by-one" error in the encoder is catastrophic and hard for humans to spot. |
2. Nested and Mixed-Type Data
Winner: GCF. The Attachment system (Section 7.4.4) uses ^ markers to keep 90% of a record in dense positional rows while surgically handling nested objects only where they occur.
3. JSON Round-Trip Fidelity
Winner: GCF. Explicitly defines lossless round-trip invariant (Section 1.1). Distinguishes null (-) from missing (~) (Section 7.4.2).
4. Graph at Scale
Winner: GCF. Local IDs cost ~4 tokens per edge vs 100+ for TOON. Section headers act as attention anchor points. 90.7% accuracy at 500 nodes.
5. Spec Rigor
Winner: GCF. Full ABNF grammar (Section 4). Canonical Encoding Rules (Section 2.3.1). 30 Strict-Mode Error Conditions (Section 16.5). Matches RFC 8259 rigor while covering much wider structural optimizations.
6. Missing or Underspecified
- JSON: missing native link/reference primitives
- TOON: missing session state (retransmits everything every call)
- GCF: missing human-friendly rendering mode
7. The Replacement Pick
GCF. "If you are building an AI agent, token efficiency and comprehension accuracy are your primary metrics. GCF is the only format that treats symbols and edges as first-class citizens. JSON is for humans/legacy; TOON is for simple tables; GCF is the professional wire format for LLM-centric systems."