Independent Review: GPT
Date: 2026-06-10 Model: GPT (fresh session, no context) Specs reviewed: RFC 8259 (JSON), TOON v3.3 (toonformat.dev + GitHub raw), GCF v2.0 Verdict: Picked GCF for AI agent tool response replacement
Rankings
| Category | Winner |
|---|---|
| Nested/mixed data | GCF ("decisively") |
| JSON round-trip fidelity | GCF |
| Graph at scale (500+) | GCF ("by a mile") |
| Spec rigor | TOON (narrowly) |
| Agent replacement pick | GCF |
Key quotes
"I would pick GCF."
"It is the only one here that materially improves both generic structured payloads and graph-heavy payloads."
"GCF handles nested data decisively."
"GCF, by a mile." (on graph structural cues)
"It states an actual round-trip invariant instead of hand-waving."
"Agent tool responses are often not pleasant. They are sparse, nested, mixed, and relationship-heavy. GCF handles that reality better."
"GCF v2.0 is much better than a typical vanity format spec."
Blunt summary (GPT's own words)
- Best standard: JSON
- Best spec among challengers: TOON
- Best design for actual AI tool payloads: GCF
- Best for graph payloads: GCF
- Best for sparse/nested mixed data: GCF
- Best for boring interoperability today: JSON
Why TOON won spec rigor
GPT specifically called out: formal status lifecycle, broader normative coverage, explicit strict/non-strict behavior, conformance classes, security, i18n, IANA, versioning, cleaner separation between core format and optional transforms. Also noted: "GCF v2.0 is much better than a typical vanity format spec" but "feels newer and more coupled to one product shape."
GCF criticisms
- Doing two jobs at once (generic format + graph transport + session/delta protocol)
- Graph profile is specialized for code graphs, less obviously good for arbitrary graphs
- Some token grammars are cramped (ASCII metadata, no whitespace in graph names)
- Session/delta move beyond pure serialization into transport protocol behavior
- MIME registration only "Suggested," not formal
- Ecosystem maturity is the obvious missing piece
TOON criticisms
- "Falls apart exactly where real tool payloads get ugly"
- Spec openly says not intended for non-uniform or deeply nested structures
- Tabular form requires all objects same keys, all values primitives
- Key folding/path expansion creates divergent implementation risk
- Still a Working Draft
- Non-strict mode leaves room for parser divergence
Full review
1. Strongest and Weakest Design Decisions
| Format | Strongest | Weakest |
|---|---|---|
| JSON | Brutal simplicity. Model is tiny: object, array, string, number, boolean, null. | Structurally dumb for repetitive data. Every record repeats keys. Duplicate-key behavior is a mess. |
| TOON | Excellent compression for uniform arrays. Spec is serious (normative syntax, strict-mode, conformance). | Falls apart where real payloads get ugly. Key folding is clever but dangerous. |
| GCF | Solves sparse and nested tabular data instead of pretending everything is clean. Graph profile gives first-class graph structure. | Doing two jobs at once raises implementation complexity. Some grammars too restrictive. |
2. Nested and Mixed-Type Data
Winner: GCF, decisively. TOON spec limits tabular form to uniform primitive-only objects. GCF has tabular rows + ~ for missing + ^ attachments + expanded form for mixed arrays.
3. JSON Round-Trip Fidelity
Winner: GCF. Explicitly promises lossless round-trip. Explicitly calls out what is not preserved. TOON is less explicit about sparse-object handling and key folding creates shape surprises.
4. Graph at Scale
Winner: GCF, by a mile. Real graph structure vs "a pile of objects and arrays."
5. Spec Rigor
Winner: TOON, narrowly. Standards-shaped end to end. But GCF "is much better than a typical vanity format spec."
6. Missing or Underspecified
- JSON: duplicate keys, numeric precision, no canonical form, no structural help
- TOON: working draft, numeric out-of-range is implementation-defined, non-strict mode divergence, key folding complicates syntax, no answer to sparse nested arrays except fallback
- GCF: graph profile specialized not general, some grammars cramped, session/delta are protocol not serialization, MIME only suggested, ecosystem maturity
7. The Replacement Pick
GCF. "It is the only one here that materially improves both generic structured payloads and graph-heavy payloads. Its generic profile is stronger than TOON on sparse and nested data. Its graph profile gives the model better cues than raw JSON or TOON ever will."