Benchmarks

Two independent evaluations prove GCF's superiority: a comprehension accuracy test (do LLMs understand it?) and a token efficiency test (how many tokens does it cost?).

LLM Comprehension Accuracy

Setup: 500 symbols, 200 edges. Same payload encoded in GCF, TOON, and JSON. Six structured extraction questions sent to an LLM. Deterministic ground truth, no LLM judge.

Format	Accuracy	Tokens	vs JSON
GCF	100% (6/6)	11,090	79% fewer
TOON	100% (6/6)	16,378	69% fewer
JSON	66.7% (4/6)	53,341	baseline

What JSON got wrong

Question	Expected	JSON answered
"How many symbols?"	500	320
"How many targets (distance 0)?"	166	240

At 500 records, JSON's field-name repetition creates enough noise that the model loses count. It's not hallucinating; it's drowning in structural tokens that carry no semantic content.

What GCF got right

All six questions answered correctly:

1. Symbol count: 500 ✓
2. Edge count: 200 ✓
3. Highest-scored symbol name ✓
4. Kind of highest-scored symbol ✓
5. Target count (distance 0): 166 ✓
6. All unique edge types (alphabetical) ✓

GCF achieves this at 32% fewer tokens than TOON, which also scored 100%.

Reproduce

git clone https://github.com/blackwell-systems/gcf-go
cd gcf-go/eval && GOWORK=off go test -run TestComprehension -v -timeout 15m

Eval source: gcf-go/eval

---

Token Efficiency (TOON's Own Benchmark)

We inserted GCF into TOON's token efficiency benchmark. Their datasets, their tokenizer (gpt-tokenizer, o200k_base), their methodology. The only change: one additional formatter.

Mixed-Structure Track

Datasets with nested or semi-uniform structures. This is where most real-world data lives.

Semi-uniform event logs (2000 records):
  TOON   ████████████████████████████████████████████████████  154,032
  GCF    ████████████████████████████████████░░░░░░░░░░░░░░░  107,269  ◀ 44% smaller

E-commerce orders (500 orders, nested items):
  TOON   ████████████████████████████████████████████████████   73,246
  GCF    ████████████████████████████████████████████░░░░░░░   61,592  ◀ 19% smaller

Mixed-structure total:
  TOON   ████████████████████████████████████████████████████  227,896
  GCF    █████████████████████████████████████░░░░░░░░░░░░░░  169,554  ◀ 34% smaller

Flat-Only Track

Pure tabular data. TOON's claimed sweet spot. GCF still wins.

Employee records (2000 rows):
  TOON   ████████████████████████████████████████████████████   49,966
  GCF    ██████████████████████████████████████████████████░░   49,054  ◀ 2% smaller

Analytics time-series (365 days):
  TOON   ████████████████████████████████████████████████████    9,127
  GCF    ████████████████████████████████████████████████░░░░    8,397  ◀ 8% smaller

Flat-only total:
  TOON   ████████████████████████████████████████████████████   67,837
  GCF    ██████████████████████████████████████████████████░░   66,026  ◀ 3% smaller

Per-Dataset Breakdown

Dataset	Structure	GCF	TOON	CSV	JSON
Event logs	Semi-uniform	107,269	154,032	n/a	181,141
E-commerce	Nested	61,592	73,246	n/a	109,574
Nested config	Deep	693	618	n/a	905
Employees	Flat	49,054	49,966	47,137	127,050
Analytics	Flat	8,397	9,127	8,395	22,257
GitHub repos	Flat	8,575	8,744	8,512	15,144

TOON's only win: deeply nested configuration. A 75-token difference on a 618-token payload.

Why GCF wins on semi-uniform data

TOON's tabular format requires all rows to have identical fields. When data is semi-uniform (event logs where 50% have nested error objects), TOON falls back to its less efficient nested encoding for the entire array.

GCF handles semi-uniformity natively: primitive fields encode positionally, nested fields attach inline only when present. No format-level "mode switch" is required.

Reproduce

git clone https://github.com/blackwell-systems/toon.git
cd toon && git checkout gcf-comparison
cd benchmarks && pnpm install && pnpm benchmark:tokens

Fork: blackwell-systems/toon@gcf-comparison

---

Summary

Metric	GCF	TOON	JSON
Comprehension accuracy (500 sym)	100%	100%	66.7%
Input tokens (500 symbols)	11,090	16,378	53,341
Output tokens (100 symbols)	5,619	11,650	22,180
Generation validity	5/5	5/5	N/A
vs JSON input savings	79%	69%	baseline
vs TOON input savings	32%	baseline	n/a
vs JSON output savings	75%	40%	baseline
vs TOON output savings	52%	baseline	n/a
Mixed-structure efficiency	best	34% larger	72% larger
Flat-data efficiency	best	3% larger	149% larger
Session dedup (5th call)	92.7%	unavailable	unavailable
Delta encoding	81.2%	unavailable	unavailable

GCF wins on input efficiency, output efficiency, and offers session/delta features no competitor has. See the generation eval for output token methodology.