When the models split
A VibeScore reading comes from several models. Here is how to tell a real split from ordinary noise, using the figures from our October 2026 tests.
A VibeScore reading is not one model's opinion. Several models read the same text, and their readings are combined into one. In the October 2026 tests, that meant four production models, averaged into a consensus reading.
Most of the time the models broadly agree. Sometimes they do not. This tutorial shows you how to read those moments without over-reading them.
Start with what agreement means
In October, the models agreed with each other closely. Their average correlation on the 0 to 1000 score was 0.90 across 500 comments labeled by people, and 0.96 on 500 phrases we wrote.
They agreed with people far less. The consensus reading's strongest emotion matched the person's label 42% of the time. Chance would be 12%.
Hold on to both numbers. Agreement between models shows consistency, not truth. A split does not make a reading wrong, and agreement does not make it right. A split tells you where the reading is less settled.
1. Know which emotions split more often
Agreement was not the same for every emotion. In October it ran from 0.76 for trust, the lowest, to 0.90 for anger, the highest.
So a split on trust is less surprising than a split on anger.
2. Separate a split from ordinary noise
A model does not give exactly the same number twice. When the same model read the same text a second time, the two readings landed this far apart, on average, out of 1000:
- Claude Haiku 5.5: 12.0 points
- GPT-6 Luna: 16.6 points
- Grok 4.7: 18.3 points
- Gemini 3.5 Flash-Lite: 26.5 points
These were measured on phrases we wrote, not yet on the comments labeled by people. A difference smaller than these numbers is within normal variation. Two readings that differ by less than that are not telling you much.
3. Check whether one model is pulling away
If your reading shows each model's numbers, look at who disagrees. In October, Gemini 3.5 Flash-Lite repeated itself least and sat furthest from the other three. A split that comes from that one model, and stays inside its own run-to-run spread, is a weaker signal than a split between models that usually agree.
4. Go back to the words
A split is a reason to look, not a verdict. Go back to the text and read the line again. Decide for yourself what it carries. The reading describes the words. It does not tell you how a specific person feels.
What these figures cover
The figures describe the engine as tested in October 2026, on short text, with Claude Haiku 5.5, GPT-6 Luna, Gemini 3.5 Flash-Lite and Grok 4.7. A reading made with a different set of models is not covered by them. They are preliminary, and the research page lists what has not been measured yet.