One paragraph
Mom made cookies. I took them all.
FORMAT ATLAS / EARLY FIELD GUIDE
The same meaning.
A different way in.
Good developer guidance starts with exact inputs. A recipe earns its place through testing, not through a claim that one format always works.
Three layers that are easy to confuse
- Transport. JSON carrying a request to the server. Escaping a newline for transport does not necessarily change the text the model receives.
- Model-visible content. A paragraph, labeled turns, or a serialized role array. These produce different tokens.
- Native chat template. The model’s expected role markers and special tokens, applied by the runtime.
Recipe used in this published pilot
Native JEV decision calls · tree scoring · cache enabled · JSON role-array content · three topic-local fields: reading, strength and clarity. Each model’s information panel preserves its measured profile.
Inspect model profiles →One question, several controlled presentations
The examples below illustrate layouts. They are not new measured results.
Two user paragraphs
Mom made cookies. I took them all.
Role-array content
[{"role":"user","content":"Mom made cookies."},
{"role":"user","content":"I took them all."}]What a reliable recipe needs
Model and weight revision. Precision. Runtime and template. Exact content. Task and answer schema. Cache conditions. A development set for choosing a format—and separate examples for evaluating that choice.
Our earlier format experiments motivate this atlas. Their full public comparison is still being prepared. The present site publishes the arc pilot’s actual calling configuration, not a universal best-format recommendation.