Gemini trial · September 4, 2026

Which model hears the Hebrew?

The same three warm WhatsApp templates were evaluated independently by Gemini 3.8 Flash and Gemini 3.5 Flash Lite. Placeholder names were explicitly described as hints: each model could put any text in a slot, but could not change the fixed copy surrounding it.

3 templatesOne template per request to avoid cross-template confusion.

6 API callsTemperature 0 with the same JSON schema and scoring rubric.

3.8 is stricterLite is faster, but awards perfect scores much more readily.

What the models were told

Variable names such as {{time}}, {{place}}, and {{invitee}} are suggestions rather than semantic limits. The model may place any content in any variable, including moving information between them, when that improves the rendered message. Fixed text cannot change, so prefixes and suffixes constrain which inserted values sound natural. For generic use, a score of 8 or higher requires the exact recommended variable value and the complete rendered message.